Dimensional modeling in a bi environment

Front cover

Dimensional Modeling:
In a Business
Intelligence Environment
Dimensional modeling for easier data
access and analysis

Maintaining flexibility for
growth and change

Optimizing for query
performance

Chuck Ballard
Daniel M. Farrell
Amit Gupta
Carlos Mazuela
Stanislav Vohnik

ibm.com/redbooks

International Technical Support Organization

Dimensional Modeling: In a Business Intelligence
Environment

March 2006

SG24-7138-00

Note: Before using this information and the product it supports, read the information in
“Notices” on page xi.

First Edition (March 2006)

This edition applies to DB2 Universal Database Enterprise Server Edition, Version 8.2, and the
Rational Data Architect, Version 6.1.

© Copyright International Business Machines Corporation 2006. All rights reserved.
Note to U.S. Government Users Restricted Rights -- Use, duplication or disclosure restricted by GSA ADP
Schedule Contract with IBM Corp.

Contents

Notices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Trademarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
The team that wrote this redbook. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
Become a published author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi
Comments welcome. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi

Chapter 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Scope of this redbook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 What this redbook includes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Data modeling and business intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 SQL, OLTP, and E/R modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Dimensional modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Redbook contents abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Chapter 2. Business Intelligence: The destination . . . . . . . . . . . . . . . . . . 21
2.1 Business intelligence overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Information environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.2 Web services. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.3 Activity examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.4 Drivers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Key business initiatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.1 Business performance management . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.2 Real-time business intelligence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.3 Data mart consolidation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2.4 The impact of dimensional modeling. . . . . . . . . . . . . . . . . . . . . . . . . 44

Chapter 3. Data modeling: The organizing structure. . . . . . . . . . . . . . . . . 47
3.1 The importance of data modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 Data modeling techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2.1 E/R modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2.2 Dimensional modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Data warehouse architecture choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.1 Enterprise data warehouse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.2 Independent data mart architecture . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.3 Dependent data mart architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Data models and data warehousing architectures . . . . . . . . . . . . . . . . . . 61
3.4.1 Enterprise data warehouse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

© Copyright IBM Corp. 2006. All rights reserved. iii

3.4.2 Independent data mart architecture . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4.3 Dependent data mart architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.5 Data modeling life cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5.1 Modeling components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5.2 Data warehousing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.3 Conceptual design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5.4 Logical data modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5.5 Physical data modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

Chapter 4. Data analysis techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.1 Information pyramid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.1.1 The information environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.2 BI reporting tool architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3 Types of BI users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Query and reporting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 Multidimensional analysis techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5.1 Slice and dice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5.2 Pivoting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5.3 Drill-down and drill-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.5.4 Drill-across . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5.5 Roll-down and Roll-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.6 Query and reporting tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.6.1 SQL query language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.6.2 Spreadsheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.6.3 Reporting applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.6.4 Dashboard and scorecard applications . . . . . . . . . . . . . . . . . . . . . . . 98
4.6.5 Data mining applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

Chapter 5. Modeling considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.1 Converting an E/R model to a dimensional model . . . . . . . . . . . . . . . . . 104
5.1.1 Identify the business process from the E/R model . . . . . . . . . . . . . 104
5.1.2 Identify many-to-many tables in E/R model . . . . . . . . . . . . . . . . . . 105
5.1.3 Denormalize remaining tables into flat dimension tables . . . . . . . . 107
5.1.4 Identify date and time dimension from E/R model . . . . . . . . . . . . . 108
5.2 Identifying the grain for the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.2.1 Handling multiple, separate grains for a business process . . . . . . . 119
5.2.2 Importance of detailed atomic grain . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2.3 Designing different grains for different fact table types . . . . . . . . . . 124
5.3 Identifying the model dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3.1 Degenerate dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.3.2 Handling time as a dimension or a fact . . . . . . . . . . . . . . . . . . . . . . 139
5.3.3 Handling date and time across international time zones. . . . . . . . . 142
5.3.4 Handling dimension hierarchies . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

iv Dimensional Modeling: In a Business Intelligence Environment

5.3.5 Slowly changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.3.6 Handling fast changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . 163
5.3.7 Identifying dimensions that need to be snowflaked. . . . . . . . . . . . . 171
5.3.8 Identifying garbage dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.3.9 Role-playing dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.3.10 Multi-valued dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5.3.11 Use of bridge tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.3.12 Heterogeneous products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.3.13 Hot swappable dimensions or profile tables . . . . . . . . . . . . . . . . . 188
5.4 Facts and fact tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.4.1 Non-additive facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.4.2 Semi-additive facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.4.3 Composite key design for fact table . . . . . . . . . . . . . . . . . . . . . . . . 202
5.4.4 Handling event-based fact tables . . . . . . . . . . . . . . . . . . . . . . . . . . 205
5.5 Physical design considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
5.5.1 DB2 Optimizer and MQTs for aggregate navigation . . . . . . . . . . . . 212
5.5.2 Indexing for dimension and fact tables . . . . . . . . . . . . . . . . . . . . . . 218
5.6 Handling changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
5.6.1 Changes to data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
5.6.2 Changes to structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
5.6.3 Changes to business requirements. . . . . . . . . . . . . . . . . . . . . . . . . 226

Chapter 6. Dimensional Model Design Life Cycle . . . . . . . . . . . . . . . . . . 227
6.1 The structure and phases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
6.2 Identify business process requirements . . . . . . . . . . . . . . . . . . . . . . . . . 229
6.2.1 Create and Study the enterprise business process list . . . . . . . . . . 232
6.2.2 Identify business process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
6.2.3 Identify high level entities and measures for conformance . . . . . . . 235
6.2.4 Identify data sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
6.2.5 Select requirements gathering approach . . . . . . . . . . . . . . . . . . . . 237
6.2.6 Requirements gathering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
6.2.7 Requirements analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
6.2.8 Business process analysis summary . . . . . . . . . . . . . . . . . . . . . . . 244
6.3 Identify the grain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
6.3.1 Fact table granularity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
6.3.2 Multiple, separate grains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.3.3 Fact table types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.3.4 Check grain atomicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
6.3.5 High level dimensions and facts from grain . . . . . . . . . . . . . . . . . . 255
6.3.6 Final output of the identify the grain phase . . . . . . . . . . . . . . . . . . . 256
6.4 Identify the dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.4.1 Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
6.4.2 Degenerate dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266

Contents v

6.4.3 Conformed dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
6.4.4 Dimensional attributes and hierarchies . . . . . . . . . . . . . . . . . . . . . . 269
6.4.5 Date and time granularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
6.4.6 Slowly changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
6.4.7 Fast changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
6.4.8 Cases for snowflaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
6.4.9 Other dimensional challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
6.5 Identify the facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
6.5.1 Facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
6.5.2 Conformed facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
6.5.3 Fact types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
6.5.4 Year-to-date facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
6.5.5 Event fact tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
6.5.6 Composite key design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
6.5.7 Fact table sizing and growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
6.6 Verify the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
6.6.1 User verification against business requirements. . . . . . . . . . . . . . . 305
6.7 Physical design considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
6.7.1 Aggregations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
6.7.2 Aggregate navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
6.7.3 Indexing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
6.7.4 Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
6.8 Meta data management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
6.8.1 Identifying the meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
6.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330

Chapter 7. Case Study: Dimensional model development . . . . . . . . . . . 333
7.1 The project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
7.1.1 The background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
7.2 The company. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
7.2.1 Business activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
7.2.2 Product lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
7.2.3 IT Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
7.2.4 High level requirements for the project . . . . . . . . . . . . . . . . . . . . . . 340
7.2.5 Business intelligence - data warehouse project architecture . . . . . 341
7.2.6 Enterprise data warehouse E/R diagram . . . . . . . . . . . . . . . . . . . . 343
7.2.7 Company structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
7.2.8 General business process description . . . . . . . . . . . . . . . . . . . . . . 344
7.2.9 Developing the dimensional models . . . . . . . . . . . . . . . . . . . . . . . . 351
7.3 Identify the requirements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
7.3.1 Business process list. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
7.3.2 Identify business process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
7.3.3 High level entities for conformance . . . . . . . . . . . . . . . . . . . . . . . . . 359

vi Dimensional Modeling: In a Business Intelligence Environment

7.3.4 Identification of data source systems . . . . . . . . . . . . . . . . . . . . . . . 361
7.3.5 Select requirements gathering approach . . . . . . . . . . . . . . . . . . . . 362
7.3.6 Gather the requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
7.3.7 Analyze the requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
7.3.8 Business process analysis summary . . . . . . . . . . . . . . . . . . . . . . . 372
7.4 Identify the grain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
7.4.1 Identify fact table granularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
7.4.2 Identify multiple separate grains . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
7.4.3 Identify fact table types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
7.4.5 Identify high level dimensions and facts . . . . . . . . . . . . . . . . . . . . . 378
7.4.6 Grain definition summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
7.5 Identify the dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
7.5.1 Identify dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
7.5.2 Check for existing conformed dimensions . . . . . . . . . . . . . . . . . . . 384
7.5.3 Identify degenerate dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
7.5.4 Identify dimensional attributes and hierarchies . . . . . . . . . . . . . . . . 385
7.5.5 Identifying the hierarchies in the dimensions . . . . . . . . . . . . . . . . . 393
7.5.6 Date and time dimension and granularity . . . . . . . . . . . . . . . . . . . . 399
7.5.7 Handling slowly changing dimensions . . . . . . . . . . . . . . . . . . . . . . 400
7.5.8 Handling fast changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . 400
7.5.9 Identify cases for snowflaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
7.5.10 Handling other dimensional challenges . . . . . . . . . . . . . . . . . . . . 405
7.5.11 Dimensional model containing final dimensions . . . . . . . . . . . . . . 409
7.6 Identify the facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
7.6.1 Identify facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
7.6.2 Conformed facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
7.6.3 Identify fact types (additivity and derived types) . . . . . . . . . . . . . . . 414
7.6.4 Year-to-date facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
7.6.5 Event facts, composite keys, and growth . . . . . . . . . . . . . . . . . . . . 420
7.6.6 Phase Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
7.7 Other phases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
7.8 Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424

Chapter 8. Case Study: Analyzing a dimensional model. . . . . . . . . . . . . 425
8.1 Case Study - Sherpa and Sid Corporation . . . . . . . . . . . . . . . . . . . . . . . 426
8.1.1 About the company . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
8.1.2 Project definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
8.2 Business needs review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
8.2.1 Life cycle of a product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
8.2.2 Anatomy of a sale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
8.2.3 Structure of the organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
8.2.4 Defining cost and revenue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429

Contents vii

8.2.5 What do the users want? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
8.2.6 Draft dimensional model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
8.3 Dimensional model review guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
8.3.1 What is the grain? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
8.3.2 Are there multiple granularities involved? . . . . . . . . . . . . . . . . . . . . 434
8.3.4 Review granularity for date and time dimension . . . . . . . . . . . . . . . 435
8.3.5 Are there degenerate dimensions? . . . . . . . . . . . . . . . . . . . . . . . . . 435
8.3.6 Surrogate keys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
8.3.7 Conformed dimensions and facts . . . . . . . . . . . . . . . . . . . . . . . . . . 437
8.3.8 Dimension granularity and quality . . . . . . . . . . . . . . . . . . . . . . . . . . 437
8.3.9 Dimension hierarchies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
8.3.10 Cases for snowflaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
8.3.11 Identify slowly changing dimensions . . . . . . . . . . . . . . . . . . . . . . . 441
8.3.12 Identify fast changing dimensions . . . . . . . . . . . . . . . . . . . . . . . . . 442
8.3.13 Year-to-date facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
8.4 Schema following the design review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443

Chapter 9. Managing the meta data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
9.1 What is meta data? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
9.2 Meta data types according to content . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
9.2.1 Business meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
9.2.2 Structural meta data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
9.2.3 Technical meta data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
9.2.4 Operational meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
9.3 Meta data types according to the format . . . . . . . . . . . . . . . . . . . . . . . . . 453
9.3.1 Structured meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
9.3.2 Unstructured meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
9.4 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
9.4.1 Meta data strategy - Why? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
9.4.2 Meta data model - What? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
9.4.3 Meta data repository - Where? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
9.4.4 Meta data management system - How? . . . . . . . . . . . . . . . . . . . . . 456
9.4.5 Meta data system access - Who? . . . . . . . . . . . . . . . . . . . . . . . . . . 456
9.5 Data standards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
9.5.1 The contents of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
9.5.2 The format of the data - domain definition . . . . . . . . . . . . . . . . . . . 458
9.5.3 Naming of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
9.5.4 Standard data structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
9.6 Local language in international applications . . . . . . . . . . . . . . . . . . . . . . 470
9.7 Dimensional model meta data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
9.8 Meta data data model - an example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 480
9.9 Meta data tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481

viii Dimensional Modeling: In a Business Intelligence Environment

9.9.1 Meta data tools in business intelligence . . . . . . . . . . . . . . . . . . . . . 481
9.9.2 Meta data tool example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482

Chapter 10. SQL query optimizer: A primer . . . . . . . . . . . . . . . . . . . . . . . 497
10.1 What is a query optimizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
10.2 Query optimizer by example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
10.2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
10.2.2 The environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
10.2.3 Problem identification and decomposition. . . . . . . . . . . . . . . . . . . 506
10.2.4 Experiment with the problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
10.3 Query optimizer example solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
10.3.1 The total solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
10.3.2 Additional comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
10.4 Query optimizer example solution update . . . . . . . . . . . . . . . . . . . . . . . 534
10.4.1 Reading the new query plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
10.4.2 All table access methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
10.4.3 Continue reading the new query plan . . . . . . . . . . . . . . . . . . . . . . 540
10.4.4 All table join methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
10.4.5 Continuing the list of all table join methods. . . . . . . . . . . . . . . . . . 546
10.5 Rules and cost-based query optimizers . . . . . . . . . . . . . . . . . . . . . . . . 551
10.5.1 Rule 1: Outer table joins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
10.5.2 Rule 2: (Non-outer, normal) table joins . . . . . . . . . . . . . . . . . . . . . 556
10.5.3 Rule 3: (Presence and selectivity of) Filter columns . . . . . . . . . . . 559
10.5.4 Rules 4 and 5: Table size and table cardinality. . . . . . . . . . . . . . . 567
10.6 Other query optimizer technologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
10.6.1 Query rewrite. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
10.6.2 Multi-stage back-end command parser . . . . . . . . . . . . . . . . . . . . . 579
10.6.3 Index negation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
10.6.4 Query optimizer directives (hints) . . . . . . . . . . . . . . . . . . . . . . . . . 584
10.6.5 Data distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
10.6.6 Fragment elimination, multidimensional clustering . . . . . . . . . . . . 593
10.6.7 Query optimizer histograms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
10.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598

Chapter 11. Query optimizer applied. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
11.1 Software development life cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
11.1.1 Issues with the life cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602
11.1.2 Process modeling within a life cycle . . . . . . . . . . . . . . . . . . . . . . . 603
11.2 Artifacts created from a process model. . . . . . . . . . . . . . . . . . . . . . . . . 607
11.2.1 Create the SQL API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
11.2.2 Record query plan documents . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
11.3 Example of process modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
11.3.1 Explanation of the process example . . . . . . . . . . . . . . . . . . . . . . . 614

Contents ix

11.4 An SQL DML example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
11.4.1 Explanation of DML example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
11.5 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624

Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627

Abbreviations and acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633

Related publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
IBM Redbooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
Other publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
How to get IBM Redbooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
Help from IBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639

x Dimensional Modeling: In a Business Intelligence Environment

Notices

This information was developed for products and services offered in the U.S.A.

IBM may not offer the products, services, or features discussed in this document in other countries. Consult
your local IBM representative for information on the products and services currently available in your area.
Any reference to an IBM product, program, or service is not intended to state or imply that only that IBM
product, program, or service may be used. Any functionally equivalent product, program, or service that
does not infringe any IBM intellectual property right may be used instead. However, it is the user's
responsibility to evaluate and verify the operation of any non-IBM product, program, or service.

IBM may have patents or pending patent applications covering subject matter described in this document.
The furnishing of this document does not give you any license to these patents. You can send license
inquiries, in writing, to:
IBM Director of Licensing, IBM Corporation, North Castle Drive Armonk, NY 10504-1785 U.S.A.

The following paragraph does not apply to the United Kingdom or any other country where such provisions
are inconsistent with local law: INTERNATIONAL BUSINESS MACHINES CORPORATION PROVIDES
THIS PUBLICATION "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESS OR IMPLIED,
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF NON-INFRINGEMENT,
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Some states do not allow disclaimer
of express or implied warranties in certain transactions, therefore, this statement may not apply to you.

This information could include technical inaccuracies or typographical errors. Changes are periodically made
to the information herein; these changes will be incorporated in new editions of the publication. IBM may
make improvements and/or changes in the product(s) and/or the program(s) described in this publication at
any time without notice.

Any references in this information to non-IBM Web sites are provided for convenience only and do not in any
manner serve as an endorsement of those Web sites. The materials at those Web sites are not part of the
materials for this IBM product and use of those Web sites is at your own risk.

IBM may use or distribute any of the information you supply in any way it believes appropriate without
incurring any obligation to you.

Information concerning non-IBM products was obtained from the suppliers of those products, their published
announcements or other publicly available sources. IBM has not tested those products and cannot confirm
the accuracy of performance, compatibility or any other claims related to non-IBM products. Questions on
the capabilities of non-IBM products should be addressed to the suppliers of those products.

This information contains examples of data and reports used in daily business operations. To illustrate them
as completely as possible, the examples include the names of individuals, companies, brands, and products.
All of these names are fictitious and any similarity to the names and addresses used by an actual business
enterprise is entirely coincidental.

COPYRIGHT LICENSE:
This information contains sample application programs in source language, which illustrates programming
techniques on various operating platforms. You may copy, modify, and distribute these sample programs in
any form without payment to IBM, for the purposes of developing, using, marketing or distributing application
programs conforming to the application programming interface for the operating platform for which the
sample programs are written. These examples have not been thoroughly tested under all conditions. IBM,
therefore, cannot guarantee or imply reliability, serviceability, or function of these programs. You may copy,
modify, and distribute these sample programs in any form without payment to IBM for the purposes of
developing, using, marketing, or distributing application programs conforming to IBM's application
programming interfaces.

© Copyright IBM Corp. 2006. All rights reserved. xi

Trademarks
The following terms are trademarks of the International Business Machines Corporation in the United States,
other countries, or both:

Redbooks (logo) ® DB2 OLAP Server™ MQSeries®
pSeries® DB2® Rational®
AIX® DRDA® Red Brick™
Cube Views™ Informix® Redbooks®
Database 2™ Intelligent Miner™ RS/6000®
Distributed Relational Database IBM® WebSphere®
Architecture™ IMS™

The following terms are trademarks of other companies:

SAP, and SAP logos are trademarks or registered trademarks of SAP AG in Germany and in several other
countries.

Oracle, JD Edwards, PeopleSoft, Siebel, and TopLink are registered trademarks of Oracle Corporation
and/or its affiliates.

MetaStage, MetaBroker, MetaArchitect, DataStage, Ascential, are trademarks or registered trademarks of
Ascential Software Corporation in the United States, other countries, or both.

Co-Author, and Portable Document Format (PDF) are either registered trademarks or trademarks of Adobe
Systems Incorporated in the United States, other countries, or both.

EJB, Java, JDBC, JDK, JRE, JVM, J2EE, and all Java-based trademarks are trademarks of Sun
Microsystems, Inc. in the United States, other countries, or both.

Excel, Microsoft, Visual Basic, Windows NT, Windows, and the Windows logo are trademarks of Microsoft
Corporation in the United States, other countries, or both.

UNIX is a registered trademark of The Open Group in the United States and other countries.

Linux is a trademark of Linus Torvalds in the United States, other countries, or both.

Other company, product, or service names may be trademarks or service marks of others.

xii Dimensional Modeling: In a Business Intelligence Environment

Preface

Business intelligence (BI) is a key driver in the business world today. We are now
deep into the information age, and things have changed dramatically. It has long
been said that information is power, and we can now understand that statement
ever more clearly.

Business is moving at a much faster pace. Management is looking for answers to
their questions, and they need these answers much more quickly. Time is money,
and real-time information is fast becoming a requirement. These directions,
movement, and initiatives force major changes in all the business processes, and
have put a sharper focus on the whole area of data management.

To support all this requires integrated, accurate, current, and understandable
sources of data that can be accessed very quickly, transformed to information,
and used for decision-making. Do you have the data infrastructure in place to
support those requirements? There can be a good number of components
involved, such as:
A well-architected data environment
Access to, and integration of, heterogeneous data sources
Data warehousing to accumulate, organize, and store the data
Business processes to support the data flow
Data federation for heterogeneous data sources
Dashboarding for proactive process management
Analytic applications for dynamic problem recognition and resolution

It should be quite clear that data is the enabler. And the first point in the above list
is about having a good data architecture. What would you say is the cornerstore
building block for creating such an architecture? You are correct! It is the data
model. Everything emanates from the data model. More precisely, in a data
warehousing and business intelligence environment, the dimensional model.
Therefore, the subject of this IBM® Redbook.

Are you ready for these changes? Planning for them? Already started? Get
yourself positioned for success, set goals, and quickly move to reach them.

Need help? That is one of the objectives of this IBM Redbook. Read on.

Acknowledgement
Before proceeding, we would like to acknowledge Dr. Ralph Kimball for his work
in data warehousing and dimensional data modeling. He is well known in the
industry, is a leading proponent of the technology, and is generally acknowledged

© Copyright IBM Corp. 2006. All rights reserved. xiii

to be the originator of many of the core concepts in this subject area. When you
think of subjects such as data warehousing, data marts, and dimensional
modeling, one of the first names that comes to mind is Dr. Kimball. He has
written extensively about these and related subjects, and provides education and
consulting offerings to help clients as they design, develop, and implement their
data warehousing environments. We have listed a few of his publications about
dimensional modeling and data warehousing, published by John Wiley & Sons,
in “Related publications” on page 637. We consider these to be required reading
for anyone who is interested in, and particularly for those who are implementing,
data warehousing.

The team that wrote this redbook
This IBM Redbook was produced by a team of specialists from around the world
working at the International Technical Support Organization, San Jose Center.

Some team members worked locally at the International Technical Support
Organization - San Jose Center, while others worked from remote locations. The
team members are depicted below, along with a short biographical sketch of
each:

Chuck Ballard is a Project Manager at the International
Technical Support organization, in San Jose, California. He
has over 35 years experience, holding positions in the areas
of Product Engineering, Sales, Marketing, Technical
Support, and Management. His expertise is in the areas of
database, data management, data warehousing, business
intelligence, and process re-engineering. He has written
extensively on these subjects, taught classes, and presented
at conferences and seminars worldwide. Chuck has both a
Bachelors degree and a Masters degree in Industrial Engineering from Purdue
University.

Daniel M. Farrell is an IBM certified Professional and
Pre-sales Engineer, from Denver, Colorado. In 1985, Daniel
began working with Informix® software products at a national
retail and catalog company, a job that would set his direction
for the next twenty or so years. He joined IBM as a result of
the acquisition of Informix. Daniel has a Masters in Computer
Science from Regis University and is currently pursuing a
Ph.D. from Clemson in Adult Education and Human
Resource Studies.

xiv Dimensional Modeling: In a Business Intelligence Environment

Amit Gupta is a Data Warehousing Consultant from IBM,
India. He is a Microsoft® Certified Trainer, MCDBA, and a
Certified OLAP Specialist. He has about seven years of
experience in the areas of databases, data management,
data warehousing, and business intelligence. He teaches
extensively on dimensional modeling, data warehousing,
and BI courses in IBM India. Amit has also been a
Co-Author® for a previous IBM Redbook on Data Mart
Consolidation. He holds a degree in Electronics and
Communications from Delhi Institute of Technology, Delhi University, New Delhi,
India.

Carlos Mazuela is a data warehousing specialist, data
modeler, and data architect from Zurich, Switzerland. He has
been with IBM for eight years, and is currently working on
data warehousing projects in the insurance and banking
sectors. Carlos is also a DB2/UDB specialist, primarily in the
areas of security, performance, and data recovery. He has
skills and experience in the business intelligence
environment, and has used tools such as Business Objects,
Brio Enterprise, and Crystal Reports. Carlos holds a degree
in Computer Science from the Computer Technical School of the Politechnical
University of Madrid, Spain.

Stanislav Vohnik is an IT Architect from Prague, Czech
Republic. He holds a Masters Degree in Mathematics and
Physics from the Charles University, in Prague, Czech
Republic. He has more than 10 years of experience in data
management, and has been a programmer, IT specialist, IT
adviser, and researcher in Physics. His areas of expertise
include database, data warehousing, business intelligence,
database performance, client/server technologies, and
internet computing.

Other Contributors:
Thanks to the following people for their contributions to this project:

Sreeram Potukuchi: We would like to give special acknowledgement to
Sreeram, who is a data warehouse architect and manager. He works for Werner
Enterprises, an IBM Business Partner. Sreeram contributed content, and worked
along with the team in determining the objectives and topics of the redbook.

A special thanks also to Helena, Marketa and Veronika.
From IBM Locations Worldwide

Preface xv

Melissa Montoya - DB2® Information Management Skills Segment Manager,
Menlo Park, CA.
Christine Shaw - Information Management Senior IT Specialist, Denver, CO.
Stepan Bem - BI Solutions Expert, IBM Global Services, Prague, Czech
Republic.
Milan Rafaj - Data Management Senior IT Specialist, IBM Global Services,
Prague, Czech Republic.

From the International Technical Support Organization, San Jose Center
Mary Comianos - Operations and Communications
Deanna Polm - Residency Administration
Emma Jacobs - Graphics
Leslie Parham - Editor

Become a published author
Join us for a two- to six-week residency program! Help write an IBM Redbook
dealing with specific products or solutions, while getting hands-on experience
with leading-edge technologies. You'll team with IBM technical professionals,
Business Partners and/or customers.

Your efforts will help increase product acceptance and customer satisfaction. As
a bonus, you'll develop a network of contacts in IBM development labs, and
increase your productivity and marketability.

Find out more about the residency program, browse the residency index, and
apply online at:
ibm.com/redbooks/residencies.html

Comments welcome
Your comments are important to us!

We want our Redbooks® to be as helpful as possible. Send us your comments
about this or other Redbooks in one of the following ways:
Use the online Contact us review redbook form found at:
ibm.com/redbooks
Send your comments in an e-mail to:
redbook@us.ibm.com

xvi Dimensional Modeling: In a Business Intelligence Environment

1

Chapter 1. Introduction
Within the scope of business computing and information technology (IT), it has
been said that the 1980s were about performing your business, the 1990s were
about analyzing your business, and the late 1990s and beyond are about new
routes to market and bringing your business to the World Wide Web. To build
further upon these categorizations, we offer the following points:
Performing your business
– Online Transaction Processing (OLTP) was the supporting technology.
– Companies could increase net revenue by:
• Lowering cost of sales through computerized automation, and less
intensive labor. The internal cost and percent of (human) labor is
lowered.
• Better utilization of (capital) inventory, and increased visibility of on
hand inventory. A customer new order call center can view and sell
products from numerous and remote company-owned distribution
centers.
• Faster turnaround on order processing, faster order assembly and
shipment, and faster bill generation and collection. The number of days
outstanding, an important accounting metric, is lowered.
– Companies could grow in size beyond previously known limits.
– OLTP data is modeled using Entity Relationship (E/R) modeling,
sometimes referred to as Third Normal Form (3NF).

© Copyright IBM Corp. 2006. All rights reserved. 1

– Why business or IT cares: OLTP business systems are designed and
delivered to support the business and operational goals. While IT is a cost
center, IT support systems deliver cost savings on operational activities.
Analyzing your business
– Business Intelligence (BI) systems provide the information that
management needs to make good business decisions.
– Increase company net revenue and decrease operating margins (internal
cost) by:
• Lowering customer service. BI aids in identifying high value customers,
delivering customer reward programs, and identifying causes of
customer loss through data analysis.
• Analysis of markets (product and customer demographic data) enables
more efficient application of (target) marketing programs. BI systems
support increases in market share by enabling better understanding
and execution of the business plan to enable increased sales.
• Better operational efficiencies through better understanding of
operational data.
– Allows companies to compete with the most efficient operating margins.
– BI data is modeled using a small amount of E/R (Entity/Relationship), such
as OLTP systems, but a larger percentage of businesses uses
dimensional modeling.
– Why business or IT cares: At this point in history, most companies are
expected to deliver and execute competent OLTP/operational type
systems for efficiency. However, BI systems, when executed properly, can
improve their effectiveness and offer a distinct and strategic competitive
advantage. While technically still a cost center, IT moves more into the
strategic side of business planning and execution. IT is now, or should be,
viewed as moving from simply being necessary for operations to being a
strategic requirement for success.
Bringing your business to the World Wide Web
– While OLTP and BI are types, or categorizations, of business application
systems, the Web is a standard computing platform. OLTP and BI
systems could be delivered via the Web, two-tier client/server systems, or
possibly even simple ASCII (green screen) terminals. In this context, the
Web is merely a delivery platform for OLTP and BI systems.
– You can increase company gross revenue and net revenue by the
following, for example:
• Allow for new routes to market (new sales channels). For example, a
local retailer can offer to sell products nationally and even

2 Dimensional Modeling: In a Business Intelligence Environment

internationally via their commercial Web site. Larger and global
markets offer the opportunity to increase sales revenue.
• Customers can self-service their account through a Web site or Web
Portal, make inquiries, and place, manage and track orders at any time
of day they choose. Customer self-service lowers the cost of sales. The
ability for customer self-service at the time of their choosing raises
customer satisfaction and lowers customer churn.
• Shared inventory visibility between retailer and manufacturer (extranet
application) lowers cost of sales, and can lower, or even eliminate,
inventory levels. A customer new order call center, or Web-based retail
site, can sell and deliver products from numerous and remote
manufacturer-maintained distribution centers.
– Allow companies to sell and compete globally, and at lower costs.
– While the Web is a standard computing platform, meaning standard
communication protocols, there are a few (at least) standard computing
infrastructure platform choices. Microsoft’s [dot].Net and the open source
Java/J2EE are two examples. This IBM Redbook gives attention and
preference to open source, cooperative computing, implying Java/J2EE™.
– Why business or IT cares: IT can become a profit center. That is, it can
become the virtual, automated sales agent, or customer self-service
agent. OLTP and BI systems are delivered remotely or via the Web, and
should be designed and constructed to leverage this and future platforms.

This is not to say that these categorizations imply an ending to one and
beginning of another. They are all ongoing and in a constant state of
improvement. For example, BI is all about analyzing your business. And, it is still
not a mature category. The fact that we are now in the process of bringing
business to the Web does not mean we have finished with data analysis and
performing our business. It is just a continuation of the evolution of business
computing and information technology.

Chapter 1. Introduction 3

1.1 Scope of this redbook
Business intelligence (BI) environments are changing, and quite dramatically. BI
is basically comprised of a data warehousing infrastructure, and a query,
analysis, and reporting environment. In this redbook, we focus on the data
warehousing infrastructure, but primarily a specific element of it termed the data
model. Or, more precisely in a data warehousing and business intelligence
environment, the dimensional model. We consider this the base building block of
the data warehouse. The focus then is on the data model. Or, more precisely, the
topic of data modeling and its impact on the business and business applications.

We discuss data modeling techniques and how to use them to develop flexible
and highly performant data models. We refer to the two primary techniques
businesses use, as EAR or E/R (Entity Attribute Relationship or sometimes as
simply Entity Relationship) data modeling and dimensional modeling.

In this redbook, we take a specific focus on dimensional modeling. There is a
detailed overview of dimensional modeling, along with examples to aid in
understanding. We also provide best practices for implementing and maintaining
a dimensional model, for converting existing data models, and for combining
multiple models.

Acknowledgement
Before proceeding, we would like to acknowledge Dr. Ralph Kimball for his work
in data warehousing and dimensional data modeling. He is well known in the
industry, is a leading proponent of the technology, and is generally acknowledged
to be the originator of many of the core concepts in this subject area. When you
think of subjects such as data warehousing, data marts, and dimensional
modeling, one of the first names that comes to mind is Dr. Kimball. He has
written extensively about these and related subjects, and provides education and
consulting offerings to help clients as they design, develop, and implement their
data warehousing environments. We have listed a few of his publications about
dimensional modeling and data warehousing, published by John Wiley & Sons,
in “Related publications” on page 637. We consider these to be required reading
for anyone who is interested in, and particularly for those who are implementing,
data warehousing.

Objective
Once again, the objective is not to make this redbook a treatise on dimensional
modeling techniques, but to focus at a more practical level. That is to relate the
implementation and maintenance of a dimensional model to business
intelligence. The primary purpose of business intelligence is to provide answers
to your business questions, and that requires a robust data warehousing
infrastructure to house your data and information objects. But, it also requires a


query, analysis, and reporting environment to get the information out of the data
warehouse and to the users. And to get it to those users of the system with
acceptable performance.

We also provide a discussion of three current business intelligence initiatives,
which are business performance management, real-time business intelligence,
and data mart consolidation. These are all initiatives that can help you meet your
business goals and objectives, as well as your performance measurements. For
more detailed information about these specific initiatives, refer to the following
IBM Redbooks:
Business Performance Management...Meets Business Intelligence,
SG24-6340
Preparing for DB2 Near-Realtime Business Intelligence, SG24-6071
Data Mart Consolidation: Getting Control of Your Enterprise Information,
SG24-6653

To get these, and other IBM Redbooks, see “How to get IBM Redbooks” on
page 638.

Programming and the data model
This IBM Redbook also has a focus on business application programming. More
specifically, we dedicated it to business application programming for business
intelligence systems, including such elements as data marts, data warehouses,
and operational data stores, and related technologies using relational database
servers that utilize industry standard Structured Query Language (SQL).

In general, there is systems programming and there is business application
programming. Systems programming involves creating the next great operating
system, spreadsheet, or word processor. Business application programming
would involve creating such things as an employee time and attendance system,
a customer new order entry system, a market productivity analysis and reporting
system, or something similar. While both systems programming and business
application programming make use of data modeling, systems programming
does so sparingly. Systems programming is typically more about algorithms,
program function, the user interface, and other similar things. Systems
programming uses data modeling merely as a means to accomplish its task,
which is to deliver such things as word processing functionality or a spreadsheet
program. Business application programming is typically all about processing
data, and is almost entirely dependent on data modeling. Regardless of the
architecture, a business application systems always sits on top of a persistent
data model of some type.


1.2 What this redbook includes
It seems that the informational technology industry is always chasing the next
great thing. At this time, the current next great thing seems to be
service-oriented architecture (SOA); a specific subtopic or sub-capability within a
Web-based architecture. So why do we want to provide a new publication on
business intelligence and dimensional modeling? Well, consider the following:
We focus specifically on the combined topics of dimensional modeling and
business intelligence. This redbook is not intended to be an academic
treatise, but a practical guide for implementing dimensional models oriented
specifically to business intelligence systems.
Business intelligence is a strategic type of information technology that can
deliver a significant contribution to the net and operating revenues of a
company. While the Web and initiatives such as SOA are in high demand,
they are merely the architectures with which business intelligence may be
delivered.
This redbook is a strategic and comprehensive dimensional modeling
publication.
We included best practices, as well as specific procedures to deliver systems
of this type more quickly, and with measurable and increased success rates.
Because we want to see you succeed, and we believe this redbook has
information that can help you do that.

Further, this IBM Redbook includes:
Detailed discussion of a dimensional model life cycle (DMDL). This was
developed to help you create functional and highly performant dimensional
models for your BI environment.
An extensive case study about developing a dimensional model by following
the processes and steps in the DMDL.
A detailed analysis of an existing sample dimensional model, along with a
discussion of techniques that you can use to improve it.
Practical and understandable examples. We present business intelligence
concepts by using examples taken from the business environment.
Application of current, practical technologies. We present examples that are
demonstrated using technology in currently available software products,
where applicable.
A common base of knowledge. We assume that you are already familiar with
information technology, and even online transaction processing (OLTP). This
IBM Redbook bridges the transformation and delivery of OLTP systems data
into business intelligence.


1.3 Data modeling and business intelligence
We have discussed the evolution of information technology, and the directions it
has taken over the years. It has served many needs, and particularly those
associated with business intelligence. It is, after all, information that enables
business intelligence. And if we continue looking at the information structure, we
see that at the base level there is data. That data is collected from many sources
and integrated with technology to enhance its usefulness and meaning.

However, to finish our investigations, we must go to one more level. That is the
level that defines and maintains the structure of the data, and is the key enabler
of the usefulness of the data. That level is the data model.

While there are many technologies, techniques, and design patterns, presented
in the pages and chapters that follow, this section demonstrates, by example, a
single and simple online transaction processing (OLTP) business application
system that is then migrated to a business intelligence (BI) system.

1.3.1 SQL, OLTP, and E/R modeling
Structured Query Language (SQL) is a data access command and control
language associated exclusively with relational databases. Databases are simply
sizeable collections of related data; such as facts, transactions, names, dates,
and places. Relational databases are those databases of a given operational
style or design pattern. As computer software languages go, SQL is best referred
to as a declarative language. Declarative languages declare; they tell some
other entity what to do without telling them how to do it. Hyper Text Markup
Language (HTML) is another declarative computer software language. HTML
tells the Web browser what the markup instructions are, but it is the Web browser
that determines how to render the given text. For example, HTML may request a
bold emphasis on a text string, but the Web browser determines that a given
installation desires a heavyweight character font and 12 point type face when
bold is requested. SQL is the command language to read, write, and define the
data structures that reside within a database.

The technology for relational database was unveiled in 1970 in a publication by
IBM researcher, E.F. Codd. However, it was not until the early to mid-1980s that
relational databases began to prove their commercial viability. At that time,
independent software vendors, such as Informix Software (originally named
Relational Database Systems), Oracle® Software (originally named Relational
Database Technologies), and Ingres Software began shipping software systems
that would serve data via receipt and processing of SQL commands. A business
application program handled the user interface. Both keyboard input and terminal
output execute some amount of business logic, but then rely on the database
server to read and write data to a shared repository, the relational database.


A project plan to develop an E/R data model
Rather than project plan, perhaps it would be better to say we are reviewing the
software development life cycle in which an online transaction processing
business application system is developed using E/R data modeling. A few more
points to consider here:
As a phrase, software development life cycle represents an abstract concept,
a life cycle.
The Waterfall Method is one implementation of an SDLC. In a Waterfall
Method SDLC, there are generally five to seven development and analysis
stages through which you create business application software. Figure 1-1 on
page 9 displays a Waterfall Method.
– Stage 1 of Figure 1-1 on page 9 is entitled discovery. During this phase,
you determine the requirements of the business application. For an OLTP
business application system, this includes gathering the layouts of all
printed reports and specific data to be collected during data entry.
– Stage 2 of Figure 1-1 on page 9 is entitled data model. During this phase,
you create the data model. How you create an entity relationship (E/R)
data model, and, more specifically, how you create a dimensional model,
is the topic of this IBM Redbook.
– There is a very specific and intentional dashed line drawn between Stage
2, and the remaining stages of this Waterfall Method SDLC. Stage 2 is
normally completed by a data modeler, where Stages 3, 4, and 5 are
normally performed by programmers and programmer management. This
can raise numerous project management and performance issues, which
we address throughout this redbook.
OLTP is a type, or categorization, of a business application, and generally has
the following characteristics:
– Data reads (SQL SELECT statements) return very few data records,
generally in the range of five to one hundred and certainly fewer than
hundreds or thousands of records.
– The filter criteria for a given SQL SELECT is generally well known in
advance of execution; such as to return a customer order by order
number, report revenue by company division, and then by product stock
keeping identifier.
– The database and database server run-time environments are easily
optimized for this workload.


Data Admin Developers

Incorrect
400 Table Mappings of
Data Model Legacy Data

Coding and
1. Discovery 2. Data model 3. debugging 4. System test 5. Cut over

Missing Columns,
Errors in Schema
Errors in Performance
Dependencies Issues

Figure 1-1 Example of Waterfall Method, Software Development Life Cycle

Project plan issues
A number of problems display themselves in a typical project plan, as shown in
Figure 1-1. The issues include:
The people who gather all of the application expertise and business
knowledge to create the data model, are not the same people who author the
programming to access this data model. Optimizations and efficiencies would
be gained if the software development life cycle being employed somehow
encouraged this knowledge to propagate from the data modeling phase to the
programming and testing phases.
The application programmers are those people best enabled to find missing
columns and other errors in the data model as they begin and then complete
their work. This can create a cycle of dependency, and then greater errors as
the programmers rely on the modelers to make changes, which then often
introduces errors in other application programs which had already been
completed.
In the system test phase, further errors are uncovered, as it becomes known
that the programmers misinterpreted the large data model, incorrectly joining
tables, and other similar activities.
Often the first migration of data from the existing system to the new system
occurs in whole only towards the end of the project. This leads to discovery of
errors in the data model, missing attributes, or incorrect data type mappings.


And then last, but certainly not all inclusive, application performance too is
one of the last elements to be tested following data migration from the
existing system.

In addition to data modeling for business intelligence, the redbook addresses the
process of data modeling, not only as a standalone topic, but as a topic to be
integrated into a software development life cycle.

An example OLTP business system with E/R data model
Figure 1-2 on page 12 displays an OLTP data model. Again, an OLTP data model
is a bit of a misnomer; it is actually a style or categorization of business
application system. OLTP systems are almost exclusively associated with E/R
data modeling. As a means to detail dimensional (data) modeling, which is
associated with BI business application systems, both E/R modeling and
dimensional modeling are detailed in the redbook. For now, see the following:
As shown in Figure 1-2 on page 12, E/R data models tend to cascade;
meaning one table flows to the next, picking up volume as it goes down the
data hierarchy. A customer places one or more orders, and an order has one
or more order line items. A single customer could have purchased hundreds
of order line items, as brokered through the dozens of orders placed.
From the example in Figure 1-2 on page 12, a printed copy of a Customer
Order would have to extract (assemble) data from the Customer table, the
Customer Order table, the Order Line Item table, and perhaps others. E/R
data models optimize for writing, and as a result have little or no data
redundancy. The customer contact data, a single record in the customer
table, is listed in one location. If this customer contact data is needed to print
a given customer order, data must be read from both the Customer table and
Customer Order (and perhaps even Order Line Item) table.
E/R data models, and therefore OLTP business applications, are optimized
for writing. E/R data models do not suffer for reading data, because the data
access methods are generally well known; and data is generally located by a
well known record key.
If E/R data models are reasonably performant for writing and reading, why
then do we need dimensional modeling?
– E/R data models represent the data as it currently exists. From the
example in Figure 1-2 on page 12, the current state of the Customer
Account, the current Inventory level (Stock table), the current Item Price
(Stock table), are all things that are known. However, E/R data models do
not adequately represent temporal data. That is, a history of data as it
changes over time. What was the inventory level by product each morning
at 9 AM, and by what percentage did sale revenue change and net profit
rise or fall with the last retail price change. These are examples of the


types of questions that E/R models have trouble answering, but that
dimensional models can easily answer.
– Although E/R models perform well for reading data, there are many
assumptions for that behavior. E/R models typically answer simple
questions which were well anticipated, such as reading a few data
records, and only then via previously created record keys. Dimensional
models are able to read vast amounts of data, in unanticipated manners,
with support for huge aggregate calculations. Where an E/R model might
not perform well, such as with sweeping ad hoc reads, dimensional
models likely can. Where dimensional models might not perform well,
such as when supporting intensive data writes, E/R models likely can.
Dimensional models rely upon a great amount of redundant data, unlike
E/R models, which have little or no redundant data. Redundant data
supports the sweeping ad hoc reads, but would not perform as well in
support of the OLTP writes.


Figure 1-2 Sample OLTP schema

1.3.2 Dimensional modeling
The history of E/R modeling is tied to the birth of relational database technology.
E/R models did perform well and serve OLTP business application systems well.
However the promise of relational database to provide easy access for all to the
corporate database came into dispute because of the E/R model.


An E/R data model for even a corporate division level application can have
100-200 significant data tables and thousands of columns. And all with possibly
cryptic column names, such as “mnth_end_amt_on_hnd”. This was thought to be
too complex an environment for non-IT users. And the required task of joining
three, five, nine, or more tables of data to produce a useful report was generally
considered too complex a task for an everyday user. A new approach was
required, and that approach was using dimensional modeling rather than E/R
modeling.

This new categorization of business application was called Decision Support, or
Business Intelligence, or a number of other names. Specific subtypes of these
business applications were called data warehouses, data marts, and operational
data stores.

Data warehouses and data marts still reside within relational database servers.
They still use the Structured Query Language (SQL) data access command and
control language to register their requests for service. They still place data in
tables, rows, and columns.

How data warehouses and data marts differ from OLTP systems with their E/R
data model, is also detailed, and expanded upon, throughout this redbook. As a
means to set current scope, expectations, and assumptions, in the next section
we take the E/R data model from Figure 1-2 on page 12 and convert it to a
dimensional model.

An example dimensional model
E/R data models are associated with OLTP business application systems. There
is a process you go through to accurately design and validate an E/R data model.
A business intelligence business application system may use an E/R data model,
but most commonly uses a dimensional model. A dimensional model also goes
through a design process so that it can be accurately designed and validated.
Figure 1-3 on page 14 displays a dimensional model that was created from the
E/R data model displayed in Figure 1-2 on page 12.


Figure 1-3 Dimensional model created from that shown in Figure 1-2 on page 12

This data model was created to support a customer new order entry OLTP
business application system. As mentioned, there is a process to create and then
validate a data model.


The following points address the creation of a dimensional model.
Creation of an E/R data model has a semi-rigid and well defined procedure to
follow (there are rules called Normal Forms to which you must adhere);
dimensional modeling is somewhat less formal and less rigid. Certainly there
are design patterns and goals when creating a dimensional model, but
dimensional modeling involves a percent of style. Perhaps all of this
originates from the target audiences that E/R data models and dimensional
models each have. E/R data models serve the IT architects and
programmers, and indirectly serve users, where dimensional models directly
serve users.
The central purpose of the E/R data model in Figure 1-2 on page 12 was to
serve a sales application in a customer new order entry business application
system. A customer order is, by one viewpoint, the central event or whole
point of that application. You could also imagine it to centrally be an inventory
management system, revenue recognition system, or other system. Taking
the viewpoint of the central event being a customer placing a new order, we
see that most of the tables in Figure 1-3 on page 14 point to a table titled
Sales.
A dimensional model may have one or more subjects, but simple dimensional
models have only one subject. In the case of a single subject dimensional
model, this subject is the primary topic, fact, or event of the model, and is
joined by numerous dimensions. A synonym for dimension in this context
might be demographic trait, index into (a given fact), or temporal state.
The (central) fact table in the dimensional model shown in Figure 1-3 on
page 14 is
Sales. Sales is joined by the dimensions; Customer, Geography, Time,
and then Product and Cost.
Where E/R data models have little or no redundant data, dimensional models
typically have a large amount of redundant data. Where E/R data models
have to assemble data from numerous tables to produce anything of great
use, dimensional models store the bulk of their data in the single Fact table,
or a small number of them.
Dimensional models can read voluminous data with little or no advanced
notice, because most of the data is resident (pre-joined) in the central Fact
table. E/R data models have to assemble data and best serve small well
anticipated reads. Dimensional models would not perform as well if they were
to support OLTP writes. This is because it would take time to locate and
access redundant data, combined with generally few if any indexes. Why are
there few indexes in a dimensional model? So often the majority of the fact
table data is read, which would negate the basic reason for the index. This is
referred to as index negation, a topic discussed further in Chapter 10, “SQL
query optimizer: A primer” on page 497.


These are but brief examples so that you may gather an introductory
understanding of the topic of this IBM Redbook. The following is a detailed list of
the remaining contents.

1.4 Redbook contents abstract
In this section, we give a brief description of the topics we present in this IBM
Redbook and how we organize them. We cover a wide spectrum of information,
because that is what is demanded for business intelligence solutions. However,
keep in mind that the more specific focus of this redbook is on dimensional
modeling and how it impacts business intelligence solutions.

The information presented includes topics such as OLTP, business intelligence,
dimensional modeling, program architectures, and database server
architectures. Depending on your specific interest, level of detail, and job focus,
certain chapters may be more relevant to your needs than others. So, we have
organized this redbook to enable selectivity in reading.

Let us get started with a brief overview of the IBM Redbook contents:
Chapter 1, “Introduction” includes:
– A description of the objectives and scope of the redbook, and summarizes
the value to be gained from reading this redbook.
– A summary of the evolution of information technology and describes how it
relates to dimensional modeling.
– Positioning and relationship of data modeling and business intelligence.
– An introduction to dimensional modeling, that includes descriptions and
examples of the types of data modeling.
Chapter 2, “Business Intelligence: The destination”, includes:
– A focus on business intelligence, beginning with an overview.
– A discussion of the IBM Information Pyramid, an important structure to
understand as you develop your data, and data warehousing, architecture.
– A brief discussion of the impact of the World Wide Web on data modeling.
– An overview of three of the key BI initiatives and how data modeling
impacts them.
Chapter 3, “Data Model: The organizing structure”, includes:
– More detailed descriptions of data modeling, along with detailed
discussions about the types of data modeling.


– An description of several data warehousing architectures, discussions
about the implementation of data warehousing architectures, and their
reliance on the data model.
– An introduction to the data modeling life cycle for data warehousing.
Chapter 4, “Data analysis techniques”, includes:
– A description of data analysis techniques, other than data warehousing
and BI. This includes a discussion of the various enterprise information
layers in an organization.
– An overview of BI reporting tool architectures, which includes a
classification of BI users based on their analytical needs. This includes
query and reporting, and multidimensional analysis techniques.
– An overview of querying and reporting tool capabilities that you can use in
your data analysis environment.
Chapter 5, “Dimensional model design life cycle”, includes:
– An introduction to a dimensional modeling design life cycle. Included are
the structure, phases, and a description of the business process
requirements.
– A description of the life cycle phases that can guide you through the
design of a complete dimensional model.
Chapter 6, “Some modeling considerations”, includes:
– A discussion of considerations and issues you may encounter in the
development of a dimensional model. It provides you with examples and
alternatives for resolving these issues and challenges.
– Guidelines for converting an E/R model to a dimensional model.
Chapter 7, “Case Study: Dimensional model development”, includes:
– A case study in the development of a dimensional model. It includes a
description of the project and case study company. The company data
model requirements are described, and the dimensional design life cycle
is used to develop a model to satisfy the requirements. The objective is to
demonstrate how you can start with requirements and use the life cycle to
design dimensional models for your company.
Chapter 8, “Case Study: Analyzing a dimensional model”, includes:
– Another case study. However, this is a study analyzing an existing
dimensional model. The requirements and existing model are analyzed for
compliance. Where the model design does not meet the requirements,
you are provided with suggestions and methods to change the model.
– A case study that can help prepare you to analyze your existing
dimensional models to verify that they satisfy requirements.


Chapter 9, “Managing the meta data”, includes:
– Information on meta data. Having emphasized the importance of the
dimensional model, now take a step back. Meta data is the base building
block for all data. It defines and describes the data, and gives it structure
and content.
– Definitions and descriptions of meta data types and formats. It also
provides a discussion about meta data strategy, standards, and design.
– An overview of tools used to work with meta data.
Chapter 10, “SQL query optimizer: A primer”, includes:
– A review of relational database server disk, memory, and process
architectures, which includes the relational database server multistage
back-end.
– A detailed and real world query optimizer case study. In this section, a
query is moved from executing in five minutes to subsecond. In addition to
the technology used to improve the performance of this example, a
methodology of SQL statement tuning and problem solving is introduced.
– A full review of the algorithms of rules and cost-based query optimizers.
This includes reading query plans, table access methods, table join
methods, table join order, b-tree+ and hash indexes, and gathering
optimizer statistics (including data distributions, temporary objects, and
others).
– A detailed review of other query optimizer technologies, to include; query
rewrite, optimizing SQL command parser use, index negation, query
optimizer directives, table partitioning, and multidimensional clustering.
Chapter 11, “Query optimizer applied”, includes:
– An overview of the Development Life Cycle, specifically the Waterfall
method.
– A discussion of the issues encountered with the life cycle, and the subject
of process modeling within the life cycle.
– A discussion of artifacts created by a process model, such as the SQL
API, and query plan documents.
– An example of process modeling with explanations of the steps and
structure. It also includes an SQL DML example, along with an explanation
of those steps and the structure.

That is a brief description of the contents of the redbook, but it hopefully will help
guide you with your reading priorities.


2

Chapter 2. Business Intelligence: The
destination
In Chapter 1, we gave an overview introduction to the scope, objectives, and
content of the redbook. We also introduced the two primary techniques used in
data modeling (E/R and Dimensional), positioned them, and then stated that in
this redbook we have a specific focus on dimensional modeling.

However, it is good to keep in mind the bigger picture and understand why we are
so interested in the topic of dimensional modeling. Simply put, it is because the
data model is the base building block for any data structure. In our case, the data
structure is a data warehouse. And, the data warehouse is the base building
block that supports business intelligence solutions - which is really the goal we
are trying to achieve. It is our destination.

Since it is our destination, we need to have an understanding of what it is and
why we want to go there.

Here we introduce the topic of business intelligence, discuss BI activities,
describe key business intelligence initiatives, show you several data warehouse
architectures, and then describe the impact that data modeling has on them.

The specific topics we cover are:
An overview of business intelligence
Business intelligence initiatives


– Business Performance Management
– Real-time Business Intelligence
– Data Mart Consolidation
The impact of data modeling on the key BI initiatives


2.1 Business intelligence overview
In the competitive world of business, the survival of a company depends on how
fast they are able to recognize changing business dynamics and challenges, and
respond correctly and quickly. Companies must also anticipate trends, identify
new opportunities, transform their strategy, and reorient resources to stay ahead
of the competition. The key to succeeding is information.

Companies collect significant volumes of data, and have access to even more
data from outside their business. They need the ability to transform this raw data
to actionable information by capturing, consolidating, organizing, storing,
distributing, analyzing, and providing quick and easy access to it. This is the
competitive advantage, but also the challenge. All of this is the goal of business
intelligence (BI). BI helps a company create knowledge from that information to
enable better decision making and to convert those decisions into action.

BI can help with the critical issues of a company, such as finding areas with the
best growth opportunities, understanding competition, discovering the major
profit and loss areas, recognizing trends in customer behavior, determining their
key performance indicators, and changing business processes to increase
productivity. BI analyzes historical business data that is created by business or
derived from external sources, such as climatic conditions and demographic
data, to study a particular function or line of business. Information is used to
understand business trends, strengths, weaknesses, and to analyze competitors
and the market situation. Prior to the existence of BI technologies, many
companies used standard conventional methods to transform data into
actionable information. This was certainly a laborious process that consumed
enormous amounts of resources and time, not to mention the human error factor.

2.1.1 Information environment
It should be clear that BI is all about information, and the home for that
information is an enterprise data warehouse. It is no longer something that is
built for a particular advantage, it should now be seen as a business requirement.
That requirement is to have a structured and organized information environment.
Such a structure contains a number of different types and organizations of data.
And, those can be organized into a structured environment. We depict this
environment as an information pyramid in Figure 2-1 on page 24.

Chapter 2. Business Intelligence: The destination 23

Decision Making

Static
Reports,
Dashboards
Floor 5

Strategic
Dimensional, Data Marts,
Cubes.
Floor 4

Summarized Data,
Performance and Rolled-up Data.
Floor 3

Near 3rd Normal Form, Subject
Area, Code and Reference tables.
Floor 2 Tactical

Staging, Detail, Denormalized, and Raw Source
Floor 1
Operational
Operational Systems

Floor 0

Figure 2-1 Information pyramid: an example

The information technology organization (IT) has traditionally seen different
types of data in the information pyramid as separate layers, and required that
data to be copied from one layer to another. However, you should look at
different types of data as different views of the same data, with different
characteristics, required to do a specific job. To emphasize that, we have labeled
them as floors, rather than layers, of information.

To move and copy the data between the floors (and typically from the lower to the
higher floors) is no longer the only option available. There are a number of
approaches that enable integration of the data in the enterprise, and there are
tools that enable those approaches. At IBM, information integration implies the
result, which is integrated information, not the approach.

We have stated that the data on each floor has different characteristics, such as
volume, structure, and access method. Now we can choose how best to
physically store and organize the data on the different floors. For example, we
can decide whether the best technology solution is to build the floors separately
or to build the floors together in a single environment. An exception is floor zero,
which, for some time, will remain separate. For example, an OLTP system may


reside in another enterprise, or another country. Though separate, we still can
have access to the data and can move it into our data warehouse environment.

Floors one to five of the information pyramid can be mapped to the layers in an
existing data warehouse architecture. However, this should only be used to
supplement understanding and subsequent migration of the data. The preferred
view is one of an integrated enterprise source of data for decision making — and
a view that is current, or real-time.

2.1.2 Web services
At the start of Chapter 1, we discussed major eras in IT. One of those was about
bringing your business to the World Wide Web. The Web has revolutionized the
information industry, and dramatically changed the way the world does business.

There are many things we could discuss regarding the advantages and advances
made and supported by the Web. However, although the Web has not
significantly impacted the technology of dimensional modeling, it has certainly
resulted in a number of changes. For example, there are many new data
definitions to be included in the meta data and data model. And there has been a
significant increase in the volumes of data captured and transmitted, but this
does not directly impact the dimensional model.

Web services are a key advance that has significantly impacted applications.
The word Web in Web services means that all operations are performed using
the technology and infrastructure of the World Wide Web. The word service
represents an activity or processing performed on behalf of a requestor, such as
a person or application. Web services have existed ever since the Web was
invented. The ability of a Web browser to access e-mail and the ability to order a
product on the Internet are examples of Web services. More recently, however,
Web services increasingly make use of XML-based protocols and standards, and
it is better to think in terms of XML Web services. In this book, for simplicity, we
use the term Web services to signify XML Web services.

The promise of Web services
Web services technology is essentially a new programming paradigm to aid in
the development and deployment of loosely-coupled applications both within and
across enterprises. In the past, developers have developed most of their
applications from the ground up. The term code reuse was used, but this was
often not put into practice because developers typically only trust the code they
develop. As software development has progressed as a discipline, and as
programming languages have also advanced, the ability to reuse application
code has increased. The Java™ programming language, for example, has many
built-in class libraries that developers use.


As applications grow, they must execute in a distributed environment. Distributed
applications provide unlimited scalability and other benefits. Defining an interface
for distributed applications has been a challenge over the years.
Language-independent technologies, such as CORBA (Common Object
Request Broker Architecture), provide a comprehensive and powerful
programming model. Other distributed technologies work well within a single
language environment, such as Java RMI (Remote Method Invocation) and
Microsoft DCOM (Distributed Common Object Model), but are not useful in a
heterogeneous systems environment.

In contrast, Web services provide a simple-to-understand interface between the
provider and the consumer of application resources using a Web Service
Description Language (WSDL). Web services also provide the following
technologies to help simplify the implementation of distributed applications:
Application interface discovery using Universal Description, Discovery, and
Integration (UDDI)
Application interface description, again using UDDI
A standard message format using Simple Object Access Protocol (SOAP),
which is being developed as the XML Protocol specification by W3C

Web services enable any form of distributed processing to be performed using a
set of standard Web- and XML-based protocols and technologies. For more
detailed information about Web services and related topics, see the work effort
by the World Wide Web Consortium at:
http://www.w3c.org

In theory, the only requirements for implementing a Web service are:
A technique to format service requests and responses
A way to describe the service
A method to discover the existence of the service
The ability to transmit requests and responses to and from services across a
network

The primary technologies used to implement these requirements in Web
services are XML (format), WSDL (describe), UDDI (discover), and SOAP
(transmit). There are, however, many more capabilities (authentication, security,
transaction processing, for example) required to make Web services viable in the
enterprise, and there are numerous protocols in development to provide those
capabilities.

It is important to emphasize that one key characteristic of Web services is that
they are platform neutral and vendor independent. They are also somewhat
easier to understand and implement than earlier distributed processing efforts,
such as CORBA. Of course, Web services still need to be implemented in vendor


specific environments, and this is the focus of facilities such as IBM Web
Services.

Web services architecture
The Web services architecture is defined in several layers. These layers are
illustrated in Figure 2-2.

Context

Transactions Agreements
BPEL Composition
Routing

Reliability Quality of Service
WSDL
Security Service
XMLP UDDI
Attachments Interface Directory

SOAP XML Schema Inspection

Protocols Descriptions Discovery

BPEL = Business Process Execution Language
XMLP = XML Protocol

Figure 2-2 Web services layered architecture

The underpinnings of the Web services architecture are WSDL and SOAP.
WSDL is an XML vocabulary used to describe the interface of a Web service, its
protocol binding and encoding, and the endpoint of the service. SOAP is a
lightweight protocol for the exchange of information in a distributed environment,
and is used to access a Web service. It is transport-protocol independent. SOAP
messages can be transported over HyperText Transfer Protocol (HTTP), for
example, but other protocols are also supported. Examples include:
SOAP over WebSphere® MQ (Message Queuing)
RMI (Remote Method Invocation) over IIOP (Internet Inter-ORB [Object
Request Broker] Protocol)

At present, the current SOAP standard only defines bindings for HTTP. SOAP is
rightfully seen as the base for Web application-to-application interoperability. The
fast availability of SOAP implementations, combined with wide industry backing,
has contributed to its quick adoption.

SOAP employs a XML-based RPC (Remote Procedure Call) mechanism with a
request/response message-exchange pattern. It is used by a service requestor


to send a request envelope to a service provider. The SOAP request envelope
contains either an RPC method call or a structured XML document. Input and
output parameters, and structured XML documents are described in XML
schema. The service provider acts on a request and then sends back a SOAP
response envelope.

The existence of a Web service can be published and advertised in a public
UDDI registry. Publishing Web services in a public registry allows client
applications to discover and dynamically bind to Web services. UDDI helps
distributed application developers solve the maintenance problem caused by
constantly changing application interfaces. Developers can use internal private
registries, and public UDDI registries (hosted on the Internet by companies, such
as IBM and Microsoft) to publicize their application interfaces (as specified by
WSDL) and to discover other Web services. When a WSDL interface changes, a
developer can republish the new interface to the registry, and subsequent
access to the Web service will bind dynamically to the new interface.

IBM Web services
Two key IBM products for supporting Web services are WebSphere Studio and
the WebSphere Application Server.

WebSphere Studio contains a set of development tools for creating and
maintaining Java applications that use Web services, and it is based on an open
development framework known as Eclipse. For more details, see:
http://www.eclipse.org

WebSphere Studio provides tools for creating WSDL interfaces to Java
applications and DB2 data. You can publish Web services defined using
WebSphere Studio to a UDDI registry directly from the WebSphere Studio
environment. WebSphere Studio provides a UDDI browser.

IBM WebSphere Application Server is a J2EE-compliant Java Web Application
Server. It is an ideal platform for hosting DB2 Web service provider applications.
WebSphere Application Server includes the Apache SOAP server. For details,
see:
http://ws.apache.org/soap/

2.1.3 Activity examples
Typically, BI solutions use different technologies for lines of business and
departments, and you can implement them in several ways. The focus of this
redbook is to explore dimensional modeling and its impact on BI
implementations.


Business Intelligence activities include such activities as:
Multidimensional Cube Analysis
Business analysis
Clickstream analysis
Information visualization
Forecasting
Data mining (text, video, and voice)
Trending analysis
Query and reporting
Geo-spatial analysis
Enterprise portal implementation
Digital dashboards

The following are examples of BI usage in a company:
Operations: BI has an active role in helping management meet their
operational performance measurements. For example, each manager in a
retail company can receive a digital dashboard with summaries of key
performance indicators in their particular area of responsibility. This can be
areas such as a store or department, specific merchandise, loss prevention,
risk, or cash flow. Whenever the actual performance falls below a preset
threshold, BI can send alerts to the managers indicating a potential problem.
This gives the manager the ability to monitor performance metrics, analyze
information, make proactive decisions, and act on those decisions. The
manager can visualize the impact of the changes made as a result of the
changing performance indicators on the dashboard.
Finance: BI provides immediate access to financial budgeting and
forecasting data. That enables business decisions to be made based on
current and accurate financial data, such as personalized views of revenue
information by product, customer, merchandise, region, store, and time
period. It also enables business decision makers to develop trend revenue
forecasts with accuracy and speed, to compare and contrast revenues with
the goals, and identify the areas in which the company is performing better or
worse. Management at that point is well-armed with accurate information so
they can act effectively and in a timely fashion.
Customer Service: BI helps organizations assess various market segments
and customers, identify potential new business, and retain existing business.
For example, you can directly correlate customer information with the sales
information. Companies that have focused on customer service and customer
relationships have typically realized significant competitive advantage.
Human Resources: BI supports activities such as recruitment, employee
retention, and career development. Organizations can align staffing needs
with strategic goals by mapping those goals to the skills needed to achieve
them. They can then identify the recruitment practices required to bring high


quality individuals on board and identify new potential candidates for
management. BI also provides information critical for areas, such as
compensation planning, employee benefit planning, productivity planning,
and skills rating.
Marketing: One of the most important beneficiaries of BI in any company is
the marketing department. BI helps them identify various trends in business,
analyze revenues on a daily basis, identify high performers, quantify the
impact of price changes, and identify opportunities for growth. For example, a
trucking company can analyze fuel rate increases to determine such things as
the impact on revenue and profitability, identify heavy/low impact segments,
obtain optimized trucking routes, and identify multi-modal routing
opportunities.

2.1.4 Drivers
Business intelligence permeates every area of a business enterprise. It is the
need for more and more information that drives it. Those who have access to
more and accurate information are in a position to enable better
decision-making. Information is a significant business and competitive
advantage.

Business measurements
There is an ongoing and ever demanding need for businesses to increase
revenues, reduce costs, and compete more effectively. Companies today are
under extreme pressure to deploy applications rapidly, and provide business
users with easy and faster access to business information that will help them
make the best possible decisions in a timely manner.

Business complexity
Companies today are offering and supporting an ever growing range of products
and services to a larger and more diverse set of clients than ever before. BI
provides the sophisticated information analysis capabilities required to handle
those business complexities, such as mergers, acquisitions, and deregulation.

Business costs
There is a growing need to reduce the overhead of IT expenditures in every
company. There is increased pressure on IT to leverage their existing systems to
maximum benefit for the business. BI widens the scope of that ability by enabling
access not only to operational and data warehouse data, but also to external
data.


2.2 Key business initiatives
For example, we provide a brief summary of three of the key initiatives. They are
quickly becoming the differentiators as competition becomes more and more
intense. This not only applies to revenues from products and services, but also to
the internal processes and operations that can minimize production delivery
schedules and impact cost. These internal goals are significantly increasing in
importance, because they can dramatically impact the ability of the company to
meet their business performance measurements.

Business measurements and goals have come under more close scrutiny by
management and stakeholders, and the measurement time frames are much
shorter. For example, management must now be in a position to track and report
performance against quarterly goals. Missing those goals can result in a
significant response from the marketplace - and can typically be directly reflected
in a changing stakeholder value assessment.

2.2.1 Business performance management
In the dynamic business environment, increased stakeholder value has become
the main means by which business executives are measured. The ability to
improve business performance is therefore a critical requirement for
organizations. Failure to improve business performance is under close scrutiny
by stakeholders. Their voices are heard through the buying or selling of company
stock. One result of this is increased volatility of stock prices, which creates a
tense roller-coaster ride for executives. Bringing more pressure to bear, is the
fact that business performance measurement time frames are becoming ever
shorter. Quarterly targets have replaced annual ones, and the expectation of
growth and success is there at every quarter end.

To help smooth out the roller-coaster ride, businesses must react quickly to
accommodate changing marketplace demands and needs. Flexibility and
business agility are key to remaining competitive and viable. We need a holistic
approach that enables companies to align strategic and operational objectives in
order to fully manage achievement of their business performance
measurements.

The objective of BPM is to help companies improve and optimize their
operations across all aspects of the business. Business requirements, therefore,
determine what type of BPM environment is necessary. Implementing BPM,
however, is more than just about installing new technology, it also requires
organizations to review the business environment to determine if changes are
required to existing business processes to take advantage of the benefits that
BPM can provide.


To become more proactive and responsive, businesses need the information that
gives them a single view of their enterprise. With that, they can:
Make more informed and effective decisions
Manage business operations and minimize disruptions
Align strategic objectives and priorities both vertically and horizontally
throughout the business
Establish a business environment that fosters continuous innovation and
improvement

The need to continuously refine business goals and strategies, however, requires
an IT system that can absorb these changes and help business users optimize
business processes to satisfy business objectives. BPM assists here by
providing performance metrics, or key performance indicators (KPIs), which
businesses can employ to evaluate business performance. A KPI is a
performance metric for a specific business activity that has an associated
business goal or threshold. The goal or threshold is used to determine whether
or not the business activity is performing within accepted limits. The tracking and
analysis of KPIs provides business users with the insight required for business
performance optimization.

BPM also becomes a great guide for IT departments that are asked to do more
with less. It helps them focus their resources in areas that provide the most
support to enable management to meet their business goals. They can now
prioritize their tasks and focus on those aligned with meeting business
measurements and achieving the business goals.

BPM requires a common business and technical environment that can support
the many tasks associated with performance management. These tasks include
planning, budgeting, forecasting, modeling, monitoring, analysis, and so forth.

Business integration and business intelligence applications and tools work
together to provide the information required to develop and monitor business
KPIs. When an activity is outside the KPI limits, alerts can be generated to notify
business users that corrective action needs to be taken. Business intelligence
tools are used to display KPIs and alerts, and guide business users in taking
appropriate action to correct business problems. To enable a BPM environment,
organizations may need to improve their business integration and business
intelligence systems to provide proactive and personalized analytics and reports.

Simply stated, BPM is a process that enables you to meet your business
performance measurements and objectives. A BPM solution enables that
process. It enables you to proactively monitor and manage your business
processes, and take the appropriate actions that result in you meeting your
objectives.


There are a few words in the previous statement that require you to take action.
Here are a few examples:
Monitor your processes. This means you have well-defined processes, and a
way to monitor them. And the monitoring should be performed on a
continuous basis.
Manage your processes. You must be aware of their status, on a continuous
basis. That means you must be notified when the process is not performing
satisfactorily. In your well-defined process, you need to define when
performance becomes unsatisfactory. Then you must get notified so you can
take action.
Appropriate action. You need the knowledge, flexibility, and capability to take
the appropriate action to correct problems that arise.

BPM places enormous demands on BI applications and their associated data
warehouses to deliver the right information at the right time to the right people. To
support this, companies are evolving their BI environments to provide (in addition
to historical data warehouse functionality) a high-performance and
enterprise-wide analytical platform with real-time capabilities. For many
organizations this is a sizable challenge, but BPM can be the incentive to move
the BI environment to a new level of capability.

IBM has also developed a BPM Framework that enables the assembly of
components encompassing business partner products and IBM foundation
technologies. The framework includes a wide range of capabilities for modeling,
integrating, connecting, monitoring, and managing business operations within an
enterprise and across a value chain of trading partners and customers. The
unifying framework that accommodates the IBM framework is illustrated in
Figure 2-3 on page 34. This framework identifies the functional components
required for the real-time monitoring, analysis, and optimization of business
operations, and their underlying IT infrastructure.


Business performance management capabilities

Enterprise integrated development environment

Common event infrastructure

Figure 2-3 IBM BPM framework

Creating a unified framework is critical to the success of a BPM implementation.
BPM is a paradigm that succeeds from proactively managing the environment,
rather than reactively resolving individual business issues. There is a need,
therefore, for an architected solution that consolidates and integrates data
related to monitoring, events, alarms, and situation-related information across
the enterprise. Fracturing that enterprise view with isolated data silos defeats the
purpose of BPM. That is, it actually inhibits the ability to monitor and manage
enterprise-wide business performance.

You can satisfy the need for a unified solution by integrating the required data in
a DB2 data warehouse and using the data warehouse to feed the BPM
environment. If this is not feasible, use information integration technologies, such
as the IBM WebSphere Information Integrator, to create an integrated business
view of disparate source data. Using dependent data marts, that leverage data
from a data warehouse, is another possible solution. We do not recommend,
however, building a BPM solution using data from independent data marts, or
sourcing the data directly out of the operational systems. These approaches
involve isolated data silos that can lead to inconsistencies and inaccurate results.

IBM BPM solutions are based on DB2 relational databases, as well as a
combination of DB2 relational databases and DB2 OLAP databases. In most
cases, the BPM environment involves a multi-tier approach, consisting of existing


applications, data warehouses that feed the BPM solution, BPM tools, and
packaged applications.

Implementing a BPM system results in making business performance data
available to everyone that needs it. Usually, most of this performance data has
not been available to business users prior to the BPM implementation. A BPM
solution is typically facilitated by providing the proactive distribution of the data
through graphical dashboards, rather than relying on users to search for data
they require. Most users respond positively to graphical dashboards embedded
in enterprise portals, and require little if any training in how to use them.

BPM and BI
Any BI implementation is aimed at turning available data into information and
putting it into the hands of decision makers. It might be easy to conclude
therefore that BI and BPM are the same thing. BPM is focused, however, on a
subset of the information delivered by a BI system. BPM is concerned with
information that shows business performance and indicates business success or
failure. This information subset enables organizations to focus on the important
task of optimizing business performance.

BI enables businesses to access, analyze, and use their data for decision making
purposes. It is used for long-term strategic planning, short-term tactical analysis,
and for managing the daily operational business activities. Key developments in
the use of BI include:
Tactical and strategic BI are moving closer together. This is because strategic
time frames (budgeting and forecasting cycles, for example) are shrinking to
enable companies to become more responsive to business needs and
customer requirements.
Analytic applications are used more and more for proactively delivering
business intelligence to users, rather than requiring them to discover it for
themselves. In many cases, these applications not only deliver information
about business operations, but also put actual business performance into
context by comparing it against business plans, budgets, and forecasts.
Dashboards are the becoming the preferred method for delivering and
displaying business intelligence to users. Dashboards are more visual and
intuitive, and typically provide linkages that can enable people to take
immediate action.
Business rules are a key requirement as companies implement so-called
closed-loop processing to use the results of business intelligence processing
to optimize business operations. This is particularly true when there is a
requirement to support automated decisions, recommendations, and actions.


Close to real-time (near real-time), or low-latency, business information is
becoming an important requirement as organizations increasingly make use
of business intelligence for managing and driving daily business operations.

BPM forms the underpinnings for many of the BI developments we outline.
Specifically, BPM provides the BI to improve operational decision making,
become more proactive and timely, and support a wide range of business
functions.

BPM enables a BI system to tap into business events flowing through business
processes, and to measure and monitor business performance. BPM extends
traditional operational transaction processing by relating measures of business
performance to specific business goals and objectives. User alerts and
application messages can then inform business analysts and applications about
any situations that require attention. The integration of business process
monitoring with operational BI analytics enables a closed-loop solution.

2.2.2 Real-time business intelligence
As business and technology continue to change and improve, a new phenomena
is occurring. The two originally opposite ends of the business intelligence
spectrum, namely tactical and strategic decision-making, are becoming much
more closely aligned.

There is a merging of requirements and a need for current information by
everyone. It is fueling and enabling the acceleration toward another key
capability that is called closed-loop analytics. That means the results of data
warehousing, or business intelligence, analytics are being fed back to the
operational environment. Events can now be acted upon almost immediately,
avoiding costly problems rather than simply trying to minimize their impact. This
is a significant leap forward, and can now be realized with the capability to
support real time.

The Internet is also playing a key role in this movement. For example, investors
now have access to information as never before. And, they can move their money
in and out of investments quickly. If one company is not performing, they can
quickly and easily move their investment to another. Now everyone seems to
have the need for speed. They want results, and they want them now! They want
information, and they want it in real time!

To meet the requirements and support this change in thinking and measurement,
companies are moving towards real time. Change, flexibility, and speed are now
key requirements for staying in business. Typically, tactical (or short-term)
decisions are made from the operational systems. However, that data usually
requires a good deal of analysis and time to provide benefit. And strategic (or


long-term) decisions are made from historical data - which may exist in a data
warehouse. This process must change to support real time.

The direction is to make data for both types of decision-making available, or
accessible, from the data warehousing environment. Then you can combine,
analyze, and present it in a consistent manner to enable more informed
management decision-making. The data warehousing environment is now
fulfilling its destiny as the enterprise data repository. It is the base for all business
intelligence systems. However, that means you must get the data into the data
warehouse in a much more timely manner. Can you do it in real time?

The answer is yes, depending on your definition of real time. Regardless, the
movement to real time has be an evolution, because it requires a gradual
re-architecting of the company infrastructure and business process
re-engineering. The evolution needs to be based on business requirements and
priorities, and a sound business case. Companies will move a step at a time
along this evolution.

The various processes that supply data can work independently or in concert
with one another. One differentiator that impacts the list of processes for Getting
Data Into the data warehouse is identified in the term Continuous Loading that is
depicted in front of these processes in Figure 2-4 on page 38. That is a key
requirement for enabling real time, and brings with it a number of issues and
required product or implementation capabilities.

Getting the data into the data warehouse is, of course, of utmost importance.
Without it, there is no business intelligence. However, the real value of data
warehousing and BI lies on the right side of Figure 2-4 on page 38. That is the
category of Getting Data Out. That data provides the information that enables
management to make more informed decisions. And, the key now is that
management works with data that is genuinely current, that is, real-time, or near
real-time, data.


Integrating the Enterprise

Parallel Concurrent Queries Personalization
Engines with
ETL Engines Data Mining, Consumers

MQSeries ODS Rules Base,
Campaigns
Queues
Continuous
Replication Loading
DB2 Alerts,
Triggers,
Data Warehouse KPIs,
Web Services
Analytics

Dashboards

Getting Data In! Content Getting Data Out!

Figure 2-4 Real-time BI

To make use of the current data, users must be aware of it. Many
implementations create the environment, and then simply enable users with
query capability. To increase the value received, the environment needs to be
proactive. Enable the environment to recognize important events and alert users
automatically so they can take corrective action. Or, the environment must
display current status as changes occur rather than requiring a user to ask for
current status. To enable this capability, you can define key performance
indicators (KPI) with threshold values as one way to trigger such alerts. When a
defined threshold is exceeded, the system could automatically trigger an alert.

This capability will typically require a change in your business processes. Logic
must be included in the process and imbedded in your applications. It is not
sufficient for a user to run queries or look at reports to try to identify when
processes need attention. That decision process should be added to your
application software so it can automatically be ever watchful for those processes
that exceed specific thresholds. And notify you in real time! That is a vital and
very beneficial value of real-time processing. The objective is to avoid problems
rather than just reacting to them. To avoid costly problems and issues rather than
simply trying to minimize their impact.

To implement these added capabilities requires new types of applications, such
as analytic applications and interactive information dashboards. They can
monitor processes, detect values and compare to thresholds, and display or
deliver alerts and associated real-time data to the users for immediate action.
Better still, put logic into the analytic applications to enable them to do much, or
all, of the decision processing and action initiation. This capability would require


a good rules base for automated action taking. For situations requiring manual
intervention, provide guidance. That guidance (or guided analysis) could be
problem solving logic embedded in the analytic application. These are
capabilities that can enable, initiate, and/or facilitate closed-loop processing. And
it is a significant competitive advantage.

Implementing real-time BI
The implementation of near real-time BI involves the integration of a number of
activities. These activities are required in any data warehousing or BI
implementation, but now we have elevated the importance of the element of time.
The traditional activity categories of getting data in and getting data out are still
valid. But, now they are ongoing continuous processes rather than a set of steps
performed independently.

Figure 2-4 on page 38 depicts an overall view of real-time BI. On one side, it
shows the various techniques for getting data into the data warehouse from the
various data sources and integrating it. We see the following examples:
Parallel ETL Engines: Data must be extracted from the operational
environment, cleansed, and transformed before it can be placed in the data
warehouse in order to be usable for query processing and analysis. These
extract, load, transform (ETL) processes have historically been
batch-oriented. Now they must be altered to run on a continuous or
near-continuous basis. Running multiples of these processes in parallel is
another alternative for getting data into the data warehouse as quickly as
possible. Another approach is to extract and load the data, then perform any
required transformations after the data is in the data warehousing
environment. This extract, load, transform (ELT) process can result in
updating the data warehouse in a shorter time frame.
MQSeries® queues: These queues are key in a real-time environment. They
are part of a messaging-oriented infrastructure, in this case delivered by the
WebSphere family of products. Messaging systems assume the responsibility
for delivery of the messages (data) put into their queues. They guarantee
data delivery. This is extremely important because it relieves application
programmers of that responsibility, which can significantly improve the
application programmers’ productivity and translate into reduced
development costs. As a general rule, it is always better to have systems
software do as much as possible in the application development process. The
result can be faster application delivery and at a lower cost.
Replication: WebSphere Information Integrator - Replication Edition delivers
this capability. The capture component of replication takes the required data
from system logs, which eliminates the need for application development to
satisfy the requirement. It is another example of having the system provide
the capability you need rather than developing, and maintaining, it yourself.


Web services: This is a strategy where you acquire and use specific services
by downloading them from the Web. This makes them generally available to
everyone for use in the application development process. It is another
example of providing tested, reusable application modules when and where
they are needed. And, it represents a potential reduction in application
development costs.
Information Integration: Data must also be available from external data
sources, many of which are housed in a heterogeneous data environment.
WebSphere Information Integrator is a product that can more easily enable
you to access and acquire the data from these sources and use them for
updating the data warehouse.
The Operational Data Store (ODS) is another source of input to the data
warehouse, but is unique in that it can concurrently be used by the operational
systems. It typically contains data that comes from the operational transaction
environment, but has been cleansed and prepared for inclusion in the data
warehouse. The ODS is typically a separate sub-schema, or set of tables, and
itself a driver for continuous loading of the real-time data warehouse. As such
it is then considered as part of the data warehousing environment.

To summarize, real-time business intelligence is having access to information
about business actions as soon after the fact as is justifiable based on the
requirements. This enables access to the data for analysis and its input to the
management business decision-making process soon enough to satisfy the
requirement.

2.2.3 Data mart consolidation
A data mart is a construct that evolved from the concepts of data warehousing.
The implementation of a data warehousing environment can be a significant
undertaking, and is typically developed over a period of time. Many departments
and business areas were anxious to get the benefits of data warehousing, and
reluctant to wait for the natural evolution.

The concept of a data mart exists to satisfy the needs of these departments.
Simplistically, a data mart is a small data warehouse built to satisfy the needs of
a particular department or business area. Often the data mart was developed by
resources external to IT, and paid for by the implementing department or
business area to enable a faster implementation.

The data mart typically contains a subset of corporate data that is valuable to a
specific business unit, department, or set of users. This subset consists of
historical, summarized, and possibly detailed data captured from transaction
processing systems (called independent data marts), or from an existing
enterprise data warehouse (called dependent data marts). It is important to


realize that the functional scope of the data mart’s users defines the data mart,
not the size of the data mart database.

Figure 2-5 depicts a data warehouse architecture, with both dependent and
independent data marts.

Operational
System

Extract, Transform, ETL
and Load meta data

Operational
Data Store meta data

Enterprise
Data Warehouse
meta data

Line of Business
Data Marts
Independent
Dependent Data Marts Data Marts

Figure 2-5 Data warehouse architecture with data marts

As you can see, there are a number of options for architecting a data mart. For
example:
Data can come directly from one or more of the databases in the operational
systems, with few or no changes to the data in format or structure. This limits
the types and scope of analysis that can be performed. For example, you can
see that in this option, there may be no interaction with the data warehouse
meta data. This can result in data consistency issues.
Data can be extracted from the operational systems and transformed to
provide a cleansed and enhanced set of data to be loaded into the data mart
by passing through an ETL process. Although the data is enhanced, it is not
consistent with, or in sync with, data from the data warehouse.
Bypassing the data warehouse leads to the creation of an independent data
mart. It is not consistent, at any level, with the data in the data warehouse.
This is another issue impacting the credibility of reporting.
Cleansed and transformed operational data flows into the data warehouse.
From there, dependent data marts can be created, or updated. It is key that
updates to the data marts are made during the update cycle of the data
warehouse to maintain consistency between them. This is also a major
consideration and design point, as you move to a real-time environment. At


that time, it is good to revisit the requirements for the data mart, to see if they
are still valid.

However, there are also many other data structures that can be part of the data
warehousing environment and used for data analysis, and they use differing
implementation techniques. These fall in a category we are simply calling
analytic structures. However, based on their purpose, they could be thought of
as data marts. They include structures and techniques, such as:
Materialized query tables (MQT)
Multidimensional clustering (MDC)
Summary tables
Spreadsheets
OLAP (Online Analytical Processing) databases
Operational data stores
Federated databases

Although data marts can be of great value, there are also issues of currency and
consistency. This has resulted in recent initiatives designed to minimize the
number of data marts in a company. This is referred to as data mart
consolidation (DMC).
Data mart consolidation may sound simple at first, but there are many things to
consider. A critical requirement, as with almost any project, is executive
sponsorship, because you will be changing many existing systems on which
people have come to rely, even though the systems may be inadequate or
outmoded. To do this requires serious support from senior management. They
will be able to focus on the bigger picture and bottom-line benefits, and exercise
the authority that will enable making changes.

To help with this initiative, we constructed a data mart consolidation life cycle. I
Figure 2-6 on page 43 depicts a data mart consolidation life cycle.


Assess Plan Design Implement Test Deploy
Investigate Existing DMC Project EDW Schema Target EDW or
Analytic Structures Scope,Issues, and Architecture Schema
Risks Involved Construction
Analyze Data Standardization
Quality and List of Analytical of Business rules
ETL Process

DMC Assessment Findings Report
and definitions

Implementation Recommendation
Consistency Structures to be Development
consolidated
Standardization

Deployment Phase
Analyze Data
of Metadata Modifying or

Testing Phase
Redundancy Choose
Consolidation for Creating New
Identify Facts User Reports
Business/Technical each Approach
and Dimensions
Metadata of existing
to be Conformed
data marts Identify Data Standardizing
Integration and Source to Target Reporting
Existing Reporting Cleansing Effort Environment
Mapping and ETL
Needs and
Design
Environment Identify Team Standardizing
User Reports, Other BI Tools
Hardware/Software Prepare DMC Environment, and
and Inventory Plan Acceptance Tests
Project Management Project Management Project Management
Project Management

Continuing the Consolidation Process
Figure 2-6 Data mart consolidation life cycle

In the following section, we give you a brief overview of the DMC process. The
data mart consolidation life cycle consists of the following activities:
Assessment: During this phase, we assess the following topics:
– Existing analytical structures
– Data quality and consistency
– Data redundancy
– Source systems involved
– Business and technical meta data
– Existing reporting needs
– Reporting tools and environment
– Other BI tools
– Hardware, software, and other inventory

Note: Based on the assessment phase, create the “DMC Assessment
Findings” report.

Planning: Key activities in the planning phase include:
– Identifying business sponsor
– Identifying analytical structures to be consolidated
– Selecting the consolidation approach
– Defining the DMC project purpose and objectives
– Defining the scope


– Identifying risks, constraints, and concerns
– In the planning phase above, based on the DMC Assessment Findings
report, create the Implementation Recommendation report.
Design: Key activities involved in this phase are:
– Target EDW schema design
– Standardization of business rules and definitions
– Meta data standardization
– Identify dimensions and facts to be conformed
– Source to target mapping
– ETL design
– User reports
Implementation: The implementation phase includes the following activities:
– Target schema construction
– ETL process development
– Modifying or adding user reports
– Standardizing reporting environment
– Standardizing other BI tools
Testing: This may include running in parallel with production.
Deployment: This will include user acceptance testing.
Loopback: Continuing the consolidation process, which loops you back to
start through some, or all, of the process again.

When you understand all the variables that enter into a data mart consolidation
project, you can see how it can become a rather daunting task. There are not
only numerous data sources with which to deal, there is the heterogeneity. That
is, data and data storage products from numerous different vendors. Consider
that they all have underlying data models that must be consolidated and
integrated. It can be a significant integration (consolidation) project.

2.2.4 The impact of dimensional modeling
We have discussed three key BI initiatives. They are all powerful and can provide
significant benefits to the implementing companies. The common thread among
all of them is the need for data. Well, the actual need is information, but that
comes from the data. And each of those initiatives not only use data, but they
also provide it.

That data typically comes from, and goes into, the enterprise data warehouse.
That is the one common and consistent source of data in an enterprise. And as
we all know, BI is built on, integrated with, and dependent on, the data
warehouse. Therefore, it should not be a surprise when we state that all


companies need a robust and well designed data warehouse if they want to be
successful, and survive.

And one thing you should understand from reading this redbook, is that the
unifying structure of the data warehouse is the data model. It is the glue that
holds everything together, and so is deserving of significant attention. And to get
the information from the data warehouse, you need business intelligence
solutions.

You can begin to get a better appreciation for the impact of data modeling as you
look closer at data mart consolidation. This initiative involves merging data (and
the associated data models) from multiple data marts. And these data marts can
exist on many heterogeneous platforms and reside in many different databases
and data structures from many different vendors. As you look at the formats, data
types, and data definitions from these heterogeneous environments, you quickly
see the beginning of a complex task, and that is the integration of all these
heterogeneous components and elements, along with all the associated
applications and queries.

From all this it should become clear that the role of the data modeler is highly
significant and of utmost importance in the quest for the integrated real-time
enterprise.


3

Chapter 3. Data modeling: The
organizing structure
Data modeling is important because it specifies the data structure, which can
impact all aspects of data usage. For example, it can have a significant impact on
performance. This is particularly true with data warehousing. And, the data
warehouse is the primary structural element in business intelligence.

Key to business intelligence is the ability to analyze huge volumes of data,
typically by means of query processing and analytic applications. And for this,
performance is critical. We provide more detail on this topic in this chapter.

In addition, we discuss the topics of:
The primary data modeling techniques (E/R and dimensional modeling)
Data warehouse (DW) architecture choices and the data models involved:
– Enterprise data warehouse
– Independent data marts
– Dependent data marts
The data modeling life cycle for the data warehouse, which includes both
logical and physical modeling


3.1 The importance of data modeling
Generally speaking, a model is an abstraction and reflection of the real world.
Modeling gives us the ability to visualize what we cannot yet realize. It is the
same with data modeling. The primary aim of a data model is to make sure that
all data objects required by the business are accurately and fully represented.

From the business perspective, a data model can be easily verified because the
model is built by using notations and language which are easy to understand and
decipher.

However, from a technical perspective the data model is also detailed enough to
serve as a blueprint for the DBA when building the physical database. For
example, the model can easily be used to define the key elements, such as the
primary keys, foreign keys, and tables that will be used in the design of the data
structure.

Different approaches of data modeling
Data models are about capturing and presenting information. Every organization
has information that is typically either in the operational form (such as OLTP
applications) or the informational form (such as the data warehouse).

Traditionally, data modelers have made use of the E/R diagram, developed as
part of the data modeling process, as a communication media with the business
analysts. The focus of the E/R model is to capture the relationships between
various entities of the organization or process for which we design the model.
The E/R diagram is a tool that can help in the analysis of business requirements
and in the design of the resulting data structure.

However, the focus of the dimensional model is on the business. Dimensional
modeling gives us an improved capability to visualize the very abstract questions
that the business analysts are required to answer. Utilizing dimensional
modeling, analysts can easily understand and navigate the data structure and
fully exploit the data. Actually, data is simply a record of all business activities,
resources, and results of the organization. The data model is a well-organized
abstraction of that data. So, it is quite natural that the data model has become the
best method for understanding and managing the business of the organization.
Without a data model, it would be very difficult to organize the structure and
contents of the data in the data warehouse.

E/R and dimensional modeling, although related, are extremely different. Of
course, all dimensional models are also really E/R models. However, when we
refer to E/R models in this book, we mean normalized E/R models. Dimensional
models are denormalized.


There is much debate about which method is best and the conditions under
which you should select a particular technique. People use E/R modeling
primarily when designing for highly transaction-oriented OLTP applications.
When working with data warehousing applications, E/R modeling may be good
for reporting and fixed queries, but dimensional modeling is typically better for ad
hoc query and analysis.

For the OLTP applications, the goal of a well-designed E/R data model is to
efficiently and quickly get the data inside (Insert, Update, Delete) the database.
However, on the data warehousing side, the goal of the data model
(dimensional) is to get the data out (select) of the warehouse.

In the following sections, we review and define the modeling techniques and
provide selection guidelines. We also define how to use the data modeling
techniques (E/R and Dimensional Modeling) together or independently for
various data warehouse architectures. We discuss those architectures in 3.3,
“Data warehouse architecture choices” on page 57.

3.2 Data modeling techniques
In Chapter 1, “Introduction” on page 1, we briefly discussed E/R and dimensional
data modeling techniques. In this section, we discuss these important data
modeling techniques in more detail, to understand the advantages and
disadvantages of each.

3.2.1 E/R modeling
E/R modeling is a design technique in which we store the data in highly
normalized form inside a relational database. Figure 3-1 on page 50 shows a
visualization of a normalized E/R model. It is simply to depict how the various
tables in an E/R model connect and interrelate. It is called a normalized
structure. Normalization basically involves splitting large tables of data into
smaller and smaller tables, until you have tables where no column is functionally
dependent on any other column, each row consists of a single primary key and a
set of totally independent attributes of the object that are identified by the primary
key. This type of structure is said to be in third normal form (3NF).

Chapter 3. Data modeling: The organizing structure 49

Figure 3-1 A typical E/R model

The objective of normalization is to minimize redundancy by not having the same
data stored in multiple tables. As a result, normalization can minimize any
integrity issues because SQL updates then only need to be applied to a single
table. However, queries, particularly those involving very large tables, that
include a join of the data stored in multiple normalized tables may require
additional effort and programming to achieve acceptable performance.

Although data in normalized tables is a very pure form of data and minimizes
redundancy, it can be a challenge for analysts to navigate. For example, if an
analyst must navigate a data model that requires a join of 15 tables, it may likely
be difficult and not very intuitive. This is one issue that is mitigated with a
dimensional model, because it has standard and independent dimensions.

We strongly recommend third normal form for OLTP applications since data
integrity requirements are stringent, and joins involving large numbers of rows
are minimal. Data warehousing applications, on the other hand, are mostly read
only, and therefore typically can benefit from denormalization. Denormalization
is a technique that involves duplicating data in one or more tables to minimize or
eliminate time consuming joins. In these cases, adequate controls must be put in
place to ensure that the duplicated data is always consistent in all tables to avoid
data integrity issues.


Note: A table is in third normal form (3NF) if each non-key column is
independent of other non-key columns, and is dependent only on the key.
Another much-used shorthand way of defining third normal form is to say, it is
the key, the whole key, and nothing but the key.

The E/R model basically focuses on three things, entities, attributes, and
relationships. An entity is any category of an object in which the business is
interested. Each entity has a corresponding business definition, which is used to
define the boundaries of the entity — allowing you to decide whether a particular
object belongs to that category or entity. Figure 3-2 depicts an entity called
product.

Entity Relationship

Product Retail Store

• Product Type • Address of Store
• Product Name • Area
• Weight • Manager

Attributes
Figure 3-2 Example E/R model showing relationships

Product is defined as any physical item that may be stocked in one or more of the
retail stores in the company. Whether this definition is appropriate or not depends
on the use to which the model is put. In this sense, an entity may be quite specific
at one extreme, or very generic at the other extreme. Each entity has a number of
attributes associated with it. An attribute is any characteristic of an entity that
describes it and is of interest to the business.

A relationship that exists between the entities in a model describes how the
entities interact. This interaction is usually expressed as a verb. In the example,
the relationship between Product and Retail Store is defined as the retail store
stocks product.

In summary, here are advantages of the E/R modeling technique:
It eliminates redundant data, which saves storage space, and better enables
enforcement of integrity constraints.


The INSERT, UPDATE, and DELETE commands executed on a normalized
E/R model are much faster than on a denormalized model because there are
fewer redundant sources of the data, resulting in fewer executions.
The E/R modeling technique helps capture the interrelationships among
various entities for which you are designing the database. In other words, an
E/R model is very good at representing relationships.

A disadvantage of an E/R model is that it is not as efficient when performing very
large queries involving multiple tables. In other words, an E/R model is good at
INSERT, UPDATE, or DELETE processing, but not as good for SELECT
processing.

3.2.2 Dimensional modeling
To overcome performance issues for large queries in the data warehouse, we
use dimensional models. The dimensional modeling approach provides a way
to improve query performance for summary reports without affecting data
integrity. However, that performance comes with a cost for extra storage space.
A dimensional database generally requires much more space than its relational
counterpart. However, with the ever decreasing costs of storage space, that cost
is becoming less significant.

What is a dimensional model?
A dimensional model is also commonly called a star schema. This type of model
is very popular in data warehousing because it can provide much better query
performance, especially on very large queries, than an E/R model. However, it
also has the major benefit of being easier to understand. It consists, typically, of
a large table of facts (known as a fact table), with a number of other tables
surrounding it that contain descriptive data, called dimensions. When it is drawn,
it resembles the shape of a star, therefore the name. Figure 3-3 on page 53
depicts an example star schema.


PRODUCT CUSTOMER
Product_ID Customer_ID
Product_Desc Customer_NAME
Customer_Desc

FACT
Product_ID
Customer_ID
Region_ID
Year_ID
Month_ID
REGION Sales TIME
Profit Year_ID
Region_ID
Country Month_ID
State Week_ID
City Day_ID

Figure 3-3 A star schema or a dimensional model

The dimensional model consists of two types of tables having different
characteristics. They are:
Fact table
Dimension table

The following sections provide more detail for understanding the two types of
tables. Figure 3-4 on page 54 depicts an example of the fact table structure.

Fact table characteristics
The fact table contains numerical values of what you measure. For example,
a fact value of 20 might mean that 20 widgets have been sold.
Each fact table contains the keys to associated dimension tables. These are
called foreign keys in the fact table.
Fact tables typically contain a small number of columns.
Compared to dimension tables, fact tables have a large number of rows.
The information in a fact table has characteristics, such as:
– It is numerical and used to generate aggregates and summaries.
– Data values need to be additive, or semi-additive, to enable
summarization of a large number of values.
– All facts in Segment 2 must refer directly to the dimension keys in
Segment 1 of the structure, as you see in Figure 3-4 on page 54. This
enables access to additional information from the dimension tables.


Segment 1 Segment 2
Dimension keys
259 1239 169 97 75219 $15.06 $74.43 1,132 47.1 0.0436
156 499 74 88 3599 $21.64 $95.21 304 96.5 0.0039

Time Dimension
Extended Cost
Customer Dimension
Extended Unit Price
Product Dimension
Sent Units
Promotion Dimension Total Package Weight
Salesperson Dimension Unit Weight

Figure 3-4 Fact table structure

– We have depicted examples of a good fact table design and a bad fact
table design in Figure 3-5. The bad fact table contains data that does not
follow the basic rules for fact table design. For example, the data elements
in this table contain values that are:
• Not numeric. Therefore, the data cannot be summarized.
• Not additive. For example, the discounts and rebates are hidden in the
unit price.
• Not directly related to the given key structure, which means they
cannot be not additive.

“bad” “good”
fact table fact table
dim_Time dim_Time
dim_Customer dim_Customer
dim_Product dim_Product
non-numeric fields dim_Promotion dim_Promotion
new dimensions
dim_Salesperson
Salesperson dim_Status
non-additive fields
Type Status full additive –
wrong granularity Unit Price quantity sold gross margin can
(non daily) Gross Margin extended list price be computed
Daily Sales total allowances
total discounts
YearToDate Sales
extended net price
Last year YTD Sales

Figure 3-5 Good and bad fact table

We provide a more detailed discussion about fact table design in Chapter 6,
“Dimensional Model Design Life Cycle” on page 227. We now look at the second
component, the dimension table characteristics.


Dimension table characteristics
Dimension tables contain the details about the facts. That, as an example,
enables the business analysts to better understand the data and their reports.
The dimension tables contain descriptive information about the numerical
values in the fact table. That is, they contain the attributes of the facts. For
example, the dimension tables for a marketing analysis application might
include attributes such as time period, marketing region, and product type.
Since the data in a dimension table is denormalized, it typically has a large
number of columns.
The dimension tables typically contain significantly fewer rows of data than
the fact table.
The attributes in a dimension table are typically used as row and column
headings in a report or query results display. For example, the textual
descriptions on a report come from dimension attributes. Figure 3-6 depicts
an example of this.

Report Labels and Textual Data
comes from Dimension Attributes

Dimensional Model Sales $
Analysis Month Product Employee Revenue

January Choco-1 Amit 400

February Choco-2 Stanislav 1200

March Choco-3 Carlos 1600
April Choco-4 Daniel 2000

Figure 3-6 The textual data in the report comes from dimension attributes

Types of dimensional models
There are three basic types of dimensional models, and they are:
Star model
Snowflake model
Multi-star model

Figure 3-7 on page 56 depicts these models:


Star Schema Snowflake Schema Multi-Star Schema

Figure 3-7 Types of dimensional models

In the following section, we give a brief summary of the three types of
dimensional models:
Star model: Star schemas have one fact table and several dimension tables.
The dimension tables are not denormalized.
Snowflake model: Further normalization and expansion of the dimension
tables in a star schema result in the implementation of a snowflake design. A
dimension is said to be snowflaked when the low-cardinality columns in the
dimension have been removed to separate normalized tables that then link
back into the original dimension table. Figure 3-8 depicts this.

Population Market
Pop_id Market Customer
Pop_Alias Sales
Market_id Customer_id
Pop_id Market_id Customer_Name
Region_id Product_id Customer_Desc
Region Market Customer_id
State Date_id
Region_id
Director Sales Time
Profit
Year
Month
Family Product Week
Family_id Date_id
Product
Intro_date
Product_id
Family_id
Product

Figure 3-8 Snowflake schema

In this example, we expanded (snowflaked) the Product dimension by
removing the low-cardinality elements pertaining to Family, and putting them
in a separate Family dimension table. The Family table is linked to the Product
dimension table by an index entry (Family_id) in both tables. From the Product
dimension table, the Family attributes are extracted by, in this example, the
Family Intro_date. The keys of the hierarchy (Family_Family_id) are also


included in the Family table. In a similar fashion, the Market dimension was
snowflaked.
For a more detailed discussion of the snowflake schema, refer to 5.3.7,
“Identifying dimensions that need to be snowflaked” on page 171.
Multi-star model: A multi-star model is a dimensional model that consists of
multiple fact tables, joined together through dimensions. Figure 3-9 depicts
this, showing the two fact tables that were joined, which are EDW_Sales_Fact
and EDW_Inventory_Fact.

STORES
EMPLOYEE STOR_ID (PK)
EMPLOYEEKEY (PK) STOR_NAME
EMPLOYEE_Natural STOR_ADDRESS
REPORTS_TO_ID CITY
(FULL NAME) STATE
FIRSTNAME ZIP
(MANAGER NAME) And More ….
DOB
HIREDATE
ADDRESS
EDW_INVENTORY_FACT
CITY CALENDAR
And More …. STORE_ID (FK)
C_DATEID_SURROGATE (PK)
PRODUCTID_ID (FK)
C_DATE)
C_YESR
DATE_ID (FK) Fact
EDW_SALE_FACT
C_QUARTER
SUPPLIER_ID (FK)
QUANTITY_IN_INVENTORY
Table
C_MONTH
Fact PRODUCTKEY (FK) And More ….

Table EMPLOYEEKEY (FK)
CUSTOMERKEY (FK)
SUPPLIERKEY (FK)
DATEID (FK)
(POSTRANSID)
SALESQTY
UNITIPRICE
SALESPRICE VENDOR
DISCOUNT
SUPPLIERKEY (PK)
ST0REID (FK)
SUPPLIERID_Natural
(COMPANYNAME)
(CONTACTNAME)
ADDRESS
CITY
REGION
CUSTOMER
POSTALCODE
CUSTOMERKEY (PK)
COUNTRY
CUSTOMER_Natural
PHONE
COMPANYNAME
CONTACTNAME
And More …. PRODUCT
ADDRESS PRODUCTKEY (PK)
CITY PRODUCTID_NATURAL
REGION PRODUCTNAME
POSTALCODE CATERGORYNAME
COUNTRY CATEGORYDESC
PHONE QUANTITYPERUNIT
FAX And More ….
And More ….

Figure 3-9 Multi-star model

3.3 Data warehouse architecture choices
In this section, we describe three architectural approaches for data warehousing.
The approach, or combination of approaches, you select impact the data
modeling requirement.

The data warehouse architecture will determine, or be determined by, the
locations of the data warehouses and data marts themselves, and where the
control resides. For example, the data can reside in a central location that is
managed centrally. Alternatively, the data can reside in distributed local and/or
remote locations that are either managed centrally or independently.


Here we describe three architectural approaches, listed below and depicted in
Figure 3-10:
Enterprise data warehouse (EDW)
Independent data marts
Dependent data marts

You can also use these architectural choices in combinations. The most typical
example of which is combining the EDW and dependent data marts. In this
instance, the data marts receive data from the EDW, rather than directly from the
staging area.

Source Independent Data Marts
Data

Interconnected (Dependent) Enterprise (EDW)
Data Marts

Figure 3-10 Various data warehouse architectures

3.3.1 Enterprise data warehouse
An enterprise data warehouse is one that will support all, or a large part, of the
business requirement for a more fully integrated data warehousing environment
that has a high degree of data access and usage across departments or lines of
business. That is, the data warehouse is designed and constructed based on the
needs of the business as a whole. Consider it a common repository for
decision-support data that is available across the entire organization, or a large
subset of that data. We use the term Enterprise here to reflect the scope of data
access and usage, not the physical structure.

Figure 3-11 on page 59 shows an architecture diagram for an enterprise data
warehouse.


Staging Area
Source Load
Services: Data
Systems
Cleaning Mart
Combining Dimensional Model
Extract Load
Quality assure
Sorting EDW
Matching
Conform Dimensions
Conform Facts
Verify
More…

Analyst Reports

Figure 3-11 Enterprise data warehouse architecture

This type of data warehouse is characterized as having all the data under central
management. However, centralization does not necessarily imply that all the
data is in one location or in one common systems environment. That is, it is
centralized, but logically centralized rather than physically centralized. When this
is the case, by design, it then may also be referred to as a hub and spoke
implementation. The key point is that the environment is managed as a single
integrated entity.

Note: The enterprise data warehouse may also be called a Hub and Spoke
data warehouse implementation if the control is logically centralized even if the
data is spread out and physically distributed, such as the EDW and data
marts, as shown in Figure 3-11.

3.3.2 Independent data mart architecture
An independent data mart architecture, as the name implies, is comprised of
standalone data marts that are controlled by particular workgroups, departments,
or lines of business. They are typically built solely to meet the particular needs of
that particular workgroup, department, or line of business. Although there could
be, there typically is no connectivity with data marts in other workgroups,
departments, or lines of business. Therefore, these data marts do not share any
conformed dimensions and conformed facts between them.

This is one of the concerns when using an independent data mart. The data in
each may be at a different level of currency, and the data definitions may not be
consistent - even for data elements with the same name.

For example, let us assume that data mart#1 and data mart#2, as shown in
Figure 3-12 on page 60, have a customer dimension. However, since they do not
share conformed dimensions, it means that these two data marts must each
implement their own version of a customer dimension. It is these types of
decisions that can lead, for example, to inconsistent and non-current sources of


data on independent data marts in an enterprise. And this can result in
inconsistent and non-current sources of data that can lead to inaccurate decision
making

Source Staging Area Independent User
Systems Data Mart #1 Reports
Services:
Cleaning
Combining
Quality
Sorting

Non-conformed
Dimensions and Facts

Staging Area
Independent
Services: Data Mart #2
Cleaning
Combining
Quality
Sorting

Figure 3-12 Independent data warehouse architecture

Independent data marts are primary candidates for data mart consolidation for
companies around the world today. The proliferation of such independent data
marts has resulted in issues such as:
Increased hardware and software costs for the numerous data marts
Increased resource requirements for support and maintenance
Development of many extract, transform, and load (ETL) processes
Many redundant and inconsistent implementations of the same data
Lack of a common data model, and common data definitions, leading to
inconsistent and inaccurate analyses and reports
Time spent, and delays encountered, while deciding what data can be used,
and for what purpose
Concern and risk of making decisions based on data that may not be
accurate, consistent, or current
No data integration or consistency across the data marts
Inconsistent reports due to the different levels of data currency stemming
from differing update cycles; and worse yet, data from differing data sources
Many heterogeneous hardware platforms and software environments that
were implemented, because of cost, available applications, or personal
preference, resulting in even more inconsistency and lack of integration


For more detailed information about data mart consolidation specific initiatives,
refer to the following IBM Redbook:
SG24-6653

3.3.3 Dependent data mart architecture
An interconnected data mart architecture is basically a distributed
implementation. Although separate data marts are implemented in a particular
workgroup, department, or line of business, they are integrated, or
interconnected, to provide a more global view of the data. These data marts are
connected to each other using, for example, conformed dimensions and
conformed facts. For example, look at Figure 3-13. Assume that data mart#1 and
data mart#2 both use a customer dimension. When we say that these data marts
share conformed dimensions, it means that the two data marts implement the
same common version of a customer dimension. Also, each of these data marts
typically has a common staging area. At the highest level of integration, the
combination of all dependent data marts could be thought of as a distributed
enterprise data warehouse.

Figure 3-13 shows the architecture for dependent data marts. In the
implementation previously mentioned, where the EDW and dependent data mart
architectures are combined, the Staging Area is basically replaced by the EDW.

Staging Area Data Mart #1
Source (or EDW)
Systems Load
Services:
Extract
Cleaning
Combining
Sorting Conformed Dimensions
Source Transformation and
Systems Sorting Conformed Facts
Matching
Extract More…
Load

Data Mart #2

Figure 3-13 Dependent data mart architecture

3.4 Data models and data warehousing architectures
Now that we have an understanding of the various data warehouse architectures,
let us spend time investigating the components of each of the implementations.


The primary focus here is to understand how the data models are used jointly or
independently to design these data warehouse architectures.

3.4.1 Enterprise data warehouse
Figure 3-14 shows the primary components of the enterprise data warehousing
architecture. It also shows the type of data model (E/R or Dimensional) on which
each of the components is based.

Staging Area EDW
Data Mart

Source Load
Systems

Dimensional
Extract Load Model

Normalized/Denormalized Normalized User
Reports
ER Model ER Model

Figure 3-14 Data models in the enterprise data warehousing environment

Source systems
The source systems are the operational databases of the enterprise. As such,
they are typically in 3NF and represented by E/R models. However, they can also
be dimensional models or other file structures. They are typically created and
used by the applications that capture the transactions of the business. In most
cases, they are highly normalized for fast performance on transactions, using the
typical insert, update, and delete processing. The priorities of a
transaction-oriented source system are performance and high availability.

Data staging area
The staging area is the place where the extracted and transformed data is placed
in preparation for being loaded into the data warehouse. As you see in
Figure 3-14, the data staging area is primarily in 3NF, and represented by an E/R
model. However, the staging area may also contain denormalized models.

Enterprise data warehouse
As shown in Figure 3-14, the EDW is primarily in 3NF and based on an E/R data
model. Key here is that we differentiate between the data warehouse and the
data warehousing environment. That is, within the data warehousing


environment, there can also be dimensional models - in the form of dependent
data marts. The analysts can query the data warehouse directly, or through a
data mart which is populated from the data warehouse. The latter may improve
the ad hoc query performance and availability depending on the configuration.

Data mart
The data marts in the enterprise data warehousing environment are based on a
dimensional model, and could be any of the following types of schema:
Star
Snowflake
Multi-Star

And, those data marts employ conformed dimensions and conformed facts.

3.4.2 Independent data mart architecture
Figure 3-15 shows the primary components in the independent data mart
architecture. It also shows the type of data model (E/R or Dimensional) upon
which each component is based.

Staging Area Independent User
Source Data Mart #1
Systems Reports

Dimensional
Model

Normalized
ER Model
Non-conformed
Dimensions and Facts

Staging Area
Independent
Data Mart #2

Dimensional
Model
Normalized
ER Model

Figure 3-15 Data models in an independent data mart architecture


The descriptions of the components of the independent data mart architecture
are as follows:

Source systems
The source systems are typically in 3NF and based on an E/R model. However,
they can include dimensional models and other file structures. They are the data
stores used by the applications that capture the transactions of the business. The
source systems are highly normalized for fast performance on transactions,
using the typical insert, update, and delete processing. The priorities of a
transaction-oriented source system are performance and high availability.

Data staging area
The staging area is the place where the extracted and transformed data is placed
in preparation for being loaded into the data warehouse. As shown in Figure 3-15
on page 63, the data staging area is primarily in 3NF, and represented by an E/R
model. However, the staging area may also contain denormalized models. In
case of independent data warehouse architecture, there are separate staging
areas for each data mart and therefore no sharing of data between these
disparate staging areas.

Independent data marts
As shown in Figure 3-15 on page 63, the data marts are based on a dimensional
model which can be any of the following types of schemas:
Star
Snowflake
Multi-Star

The data marts in the independent data warehouse are not connected to each
other because they are not designed using conformed dimensions and
conformed facts. Here again, this can result in inconsistent and old, out of date
sources of data that can lead to inaccurate decision making.

3.4.3 Dependent data mart architecture
The dependent data mart, as the name implies, is dependent on something. And,
that something is the enterprise data warehouse.

That means that the data for the dependent data mart comes from the EDW.
Therefore, the data loaded into the dependent data mart is already transformed,
cleansed, and consistent. The primary concern when using a data mart is the
currency of the data, or how fresh it is. That is, what is the date and time of the
last extract of data from the EDW that was loaded into the dependent data mart.


And, when data is used from multiple data marts, care must be taken to assure
that the freshness of the data is consistent across them.

Figure 3-16 on page 65 shows the components of a dependent data mart
architecture, and the type of data model upon which it is based.

EDW
User
Data Mart #1 Reports

Conformed Dimensions
and
Conformed facts

Dimensional
Model
Normalized E/R Model

Figure 3-16 Data models in a dependent data mart architecture

The descriptions of the components of a dependent data mart architecture are as
follows:

Source systems
The source system for the dependent data mart is the EDW.

Data staging area
Since data is coming from the EDW, there is no requirement for a data staging
area.

Dependent data marts
The data marts are based on a dimensional model. The dimensional model can
be any of the following schemas:
Star
Snowflake
Multi-Star

The data marts are based on conformed dimensions and conformed facts.


3.5 Data modeling life cycle
In this section, we describe a data modeling life cycle. It is a straight forward
process of transforming the business requirements to fulfill the objectives for
storing, maintaining, and accessing the data within IT systems. The result is a
logical and physical data model for an enterprise data warehouse.

We describe a specific life cycle to design a dimensional model (star schema) in
Chapter 6, “Dimensional Model Design Life Cycle” on page 227.

3.5.1 Modeling components
The goal of the data modeling life cycle is primarily the creation of a storage area
for the business data. That area comes from the logical and physical data
modeling stages, as depicted in Figure 3-17.

Data Modeling Life Cycle

Business Fulfilled Business
Logical Data Physical Data Requirements
Requirements Modeling Modeling for Data Storage

Figure 3-17 A generic data modeling life cycle

In the context of data warehousing, data modeling is a critical activity. The
complexity of the business and its needs for business intelligence are a real
challenge for data modelers. We start now with a brief overview of the life cycle
components:
Logical data modeling: This component defines a network of entities and
relationships representing the business information structures and rules. The
entities are representations of business terms of relevance to the business,
such as: involved party, location, product, transaction, and event. The
relationships are representations of associations between entities. An entity
characterizes attributes, such as name, description, cost, sale price, and
business code. Figure 3-18 on page 67 is an example of a logical model.


cd Logical Model

Involved Party
Involved Party Product
Involved Party
* PK Involved Party ID: * PK Product ID:
Name: Name:
Description: Description:
Organization: Classification:
Address:
+ (PK) PK_Product()
+ (PK) PK_Involved Party()

Transaction
Involved Party
* PK Transaction ID:
Involved Party ID:
Product ID:
Transaction Type:
Net Cash Flow Amount
+ (PK) PK_Transaction()

Figure 3-18 A logical model representation

Physical data modeling: This component maps the logical data model to the
target database management system (DBMS) in a manner that meets the
system performance and storage volume requirements. The physical
database design converts the logical data entities to physical storage (such
as tables and proprietary storage) of the target DBMS or secondary storage
medium. The physical database design is composed of the Data Definition
Language (DDL) that implements the database, an information model
representing the physical structures and data elements making up the
database, and entries in the data dictionary documenting the structures and
elements of the design. An example of the DDL for the table named
transaction (TXN) is in Example 3-1.
Example 3-1 Sample DDL
create table TXN
( TXN_ID INTEGER not null,
PPN_DTM TIMESTAMP,
SRC_STM_ID INTEGER,
UNQ_ID_SRC_STM CHAR(64),
MSR_PRD_ID INTEGER,
TXN_TP_ID INTEGER,
ENVT_TP_ID INTEGER,
UOM_ID INTEGER,
TXN_VAL_DT DATE,
TXN_BOOK_DT DATE,
NET_CASH_FLOW_AMT NUMERIC(14,2),


constraint P_TXNXPK primary key
(TXN_ID)

3.5.2 Data warehousing
In the 3.3, “Data warehouse architecture choices” on page 57, we describe data
warehousing concepts and possible architectures. In Figure 3-19, we depict an
example enterprise data warehouse, where the arrows show the data flow
among components.

Data Warehouse Environment
Data Warehouse
Staging
(System of Record) Data
Area
Full history in 3rd Normal Form Marts
No user access

Source Summary Area Analytical Area
systems Full history MD model
User access Full history
User access

Figure 3-19 Enterprise data warehouse environment

The DW components differ not only by content of data but also by the way they
store the data and by whom it can be accessed.
Staging area: For handling data extracted from source systems. There can be
data transformations at this point and/or as the data is loaded into the data
warehouse. The structure of the staging area depends on the approach and
tools used for the extract, transform, and load (ETL) processes. The data
model design affects not only performance, but also scalability and ability to
process new data without recreating the entire model.
Data warehouse: This is the area, also called the system of record (SOR),
that contains the history data in 3NF and is typically not accessed for query
and analysis. Use it for populating the summary area, analytical areas, and
the dependent data marts.
Summary area: This area contains aggregations. Structures are usually
derived from the data warehouse where one or more attributes are at the
higher grain (less detail) than in the data warehouse. These are constructed
for high performance data analysis where low level detail is not required.


Analytical area: Contains multidimensional (MD) structures, such as the star
schema, snowflakes, or multi-star schemas, constructed for high performance
data analysis.

3.5.3 Conceptual design
The conceptual design represents an early basis for design reviews, including
confirmation that the business requirements are sufficiently described and that
there is an available solution. From this point starts the logical data modeling
which transforms the business requirements into the context of the
data/information necessary to be stored, accessed, and maintained.

3.5.4 Logical data modeling
In this section, we focus on the modeling of data warehouse. First we look at the
logical data model. The primary purpose of logical data modeling is to document
the business information structures, processes, rules, and relationships by a
single view - the logical data model.

The logical data model helps to address the following:
Validation of the functional application model against business requirements
The product and implementation independent requirements for the physical
database design (Physical Data Modeling)
Clear and unique identification of all business entities in the system along with
their relations, by the logical model

Note: What happens if we do not have the logical data model?

Without the logical data model, the stored business information is described
by a functional model or application (such as ETL or OLAP). Without the
logical data model, there is no single view of all data, and data normalization is
impossible. In this case, the physical data model has to be designed from a
functional model. This will potentially cause performance problems, and data
inconsistency and redundancies, which can result in an inefficient physical
design.

Design steps in logical modeling
The design activity of logical data modeling of the data warehouse consists of the
following steps:
1. Identification of entities, attributes, and relationships
2. Normalization and identification of entities
3. Merging the ETL functional model with the logical data model


4. Validation of the logical model against business requirements

We now discuss each of the tasks in detail.

One - Identification of entities, attributes, and relationships
This step consists of the following activities:
1. Review available documentation for the project, including the scope of the
project and information about the source systems from where data is loaded.
Look for business requirements, process models, profiles, architectural
design, and data models.
2. Create a list of nouns representing general categories of information required
to be stored in the data warehouse. Review this entity list with subject matter
experts. The nouns should represent standalone concepts rather than
attributes or subsets of something larger.
3. From the noun list, identify the entities. An entity is a generalization of the
concepts, involved parties, products, arrangements, locations, or events
about which the system store informations. Entities may occur directly in the
list or may be names for collections of nouns in the list. Each entity name
must be unique and meaningful. Nouns representing out-of-scope concepts
and implementation concepts are not logical data model entities.
4. Once each entity has been defined, determine its relationships with other
entities. Each entity may have multiple relationships with other entities.
However, a single relationship is only held between two entities. When
analyzing the relationships between entities, it is important that the
relationships are analyzed from the context of the business view. Each
relationship is considered bidirectional and granted names for each side of
the relationship.
5. Identify the associated cardinalities (numbers of occurrences of one entity
relative to another entity) and describe it. Cardinality refers to the minimum
and maximum numbers of occurrences implied by the existence of two
entities participating in a relationship.
6. Identify the attributes or characteristics of the entities that are relevant to the
business, and the primary key for each entity. The primary key is made up of
is the subset of attributes that uniquely identify each entity. Attributes must be
atomic. That is, they cannot be further decomposed to simpler elements.
7. Relationships, cardinalities, and inter-attribute dependencies are usually
derived from business rules.
8. For each entity and attribute record, define a text description in the data
dictionary that clearly represents the element from a business point of view.


Two - Normalization and identification of entities
Normalizing the model means removing structural redundancies and
inconsistencies. The recommended approach is to proceed in taking the model
to: first normal form, then second normal form, and finally to third normal form.
The third normal form is used for the data warehouse. This is where necessary
detail and historical data is kept, typically in an E/R model.

The primary reason for a data warehouse is to enable data query and analysis to
provide information for decision making. However, as the volume of data grows in
an E/R model, performance issues may arise. Thus, one of the reasons for the
dimensional model.

Dimensional modeling is a technique that can provide the required performance
for query and analysis on huge volumes of data. It has since become the defacto
technique for data warehousing analytics, and that is why it is the primary subject
of this redbook. There is more detailed information about dimensional modeling
in Chapter 6, “Dimensional Model Design Life Cycle” on page 227.

To prepare for dimensional modeling and analytics, the data in the data
warehouse must be formatted properly. The following list briefly describes
activities involved in this formatting:
1. Repeating Groups: The repeating attribute groups must be removed because
they are indicative of the existence of another entity. Each set is an instance
of the new entity and must contain the primary key of the original entity as a
foreign key. After these sets are removed and made separate entities, the
model is said to be in first normal form.
2. Functional dependencies: Any partial functional dependencies among
attributes in the entities must be removed. In entities that have primary keys
comprised of more than one attribute, non-key attributes may be a function of
the entire key or of part of the key. In the former case, the attribute is said to
be fully functionally dependent. In the latter case, the attribute is said to be
partially functionally dependent. Partially functionally dependent attributes are
indicative of the existence of another entity and should be removed from the
original entity. The primary key of the new entity is that part of the primary key
of the original entity that was involved in the partial dependence. At this point,
the model is said to be in second normal form.
3. Transitive dependencies: Transitive dependencies among attributes in the
entities must be removed. Mutually dependent non-key attributes are
indicative of the existence of another entity. One of the dependent attributes
is said to be dependent on the primary key of the original entity. The other
mutually dependent attributes are said to be transitively dependent on the
primary key. The dependent attribute is left in the original entity as a foreign
key. The transitively dependent attributes are removed. A new entity is
formed whose primary key is the dependent attribute in the original entity and


whose other attributes are the transitively dependent attributes of the original
entity. The model is now said to be in third normal form.
4. Primary key resolution: Instances where multiple entities have the same
primary key and these instances must be resolved. During the normalization
process, it may be found that multiple entities have the same primary keys.
These entities should only be merged if all the joined attributes are valid for all
instances of every entity. If this is not the case, a super/subtype structure
should be developed where the shared attributes are placed in the super-type
and the unique attributes are in separate subtypes.

Three - Merging the ETL functional model with the logical model
In this step, we do the following:
1. Entities in the logical data model should be mappable to source systems.
2. Attributes in the entities are to be created, read, updated, and deleted by data
flows in the ETL functional model and analytic applications.
3. The processes of the ETL functional model should maintain the relationships
and cardinalities of the E/R model and the values of the attributes.

Four - Validation of the logical model against business requirements
Consider the following points for validation and verification of the data model:
1. Each entity should represent an involved party, location, product, transaction,
or event relevant to the business.
2. The logical data model should include all information the data warehouse
needs to store about the business.
3. Each entity should have a name, a primary key, and one or more attributes,
and enter into one or more relationships with other entities.
4. Each relationship should have correct cardinalities that reflect the needs of
the business.
5. Each entity should be properly normalized.
6. Each entity and attribute should be accounted for in the data warehouse, and
related to functions or processes such as ETL, real-time DW, and DW
housekeeping.


Note: Important guidelines for developing the model for the DW are:
Focus on staying within the scope of the data warehouse system being
developed.
Use any existing data models that you have as a starting point. However,
do not just accept it as is. If required, modify it and enhance it, so that it
provides an accurate representation of business data requirements.
The high-level data model should include representation for potential future
data and relationships. This requires good knowledge of company plans for
the future.
It is often useful to first assemble a draft entity relationship diagram early
after creation of the noun list and then review it with client staff through
successive interviews or group sessions. By doing so, review sessions will
have more focus. Following the initial first draft, continuous client
involvement in data model development is critical for success.

3.5.5 Physical data modeling
In this section, we focus on physical data modeling of the main storage of the
data warehouse. The data warehouse provides a reliable single view of the data
at the required level of detail, along with the necessary history data for the
enterprise. Data modeling for the analytical area (dimensional model) or the
summary area (see Figure 3-19 on page 68) is discussed in more detail in
Chapter 6, “Dimensional Model Design Life Cycle” on page 227.

Introduction to physical modeling
The objective of physical data modeling is the mapping of the logical data model
to the physical structures of the RDBMS system hosting the data warehouse.
This involves defining physical RDBMS structures, such as tables and data types
to use when storing the data. It may also involve the definition of new data
structures for enhancing query performance. However, you must do it without
changing the meaning of the logical data model schema.

Note: What happens if there is no physical database design?
We generate DDL from the logical data model.
The final database is fully normalized without additional tables or attributes
for improving performance.


Other important factors to consider:
There are important factors to consider while designing the physical model. For
example:
You need tools for physical data modeling, such as:
– Case tools for maintaining the physical data model (as well as the logical
data model)
– Database performance tools
– Meta data management tools
Scalability of the design, and the physical RDBMS
Queries, ETL, and other applications that use the data warehouse
Use of an abstracted data model for performance
Operation/Maintenance of DW (housekeeping)

Note: How much does physical data modeling in OLTP differ from physical
modeling for the data warehouse? The answer is, not much. At the conceptual
model level, it differs primarily in performance design. The key difference is
that in OLTP, we are primarily concerned with data and transaction volumes,
where with the DW we must focus on load performance, for population of the
analytical areas and the summary tables by batch/real-time applications, and
on performance of the analytical queries.

Physical design activities for the data warehouse
We divide the physical design activity for the data warehouse into the following
steps:

One - Physical modeling of entities and attributes
In this phase, we do the following:
1. For each entity in the logical data model, we define an RDBMS table. We
document this activity and assign the name compatible with the common
DBMS and according to your company naming convention.
2. For each attribute in the entity, we identify a column and assign the name
compatible with the common DBMS naming syntax. We define RDBMS
specific data types, such as character, varchar, integer, float, and decimal.
3. We define Primary and Foreign keys of the entities to the tables.

Note: Before starting any data warehousing project, it is a good practice to
establish within the company a common naming convention for business and
technical objects, along with recommended data types in the data directory.


Two - Build the DDL
The following is a list of activities for this step:
1. Create target database.
2. Define the target database vendor.
3. Connect to target database.
4. Generate DDL by a preferred Case tool.
5. Implement the DDL code using Case tool or a script.

Three - Performance design and tuning
In this step, we comment briefly about how to proceed with the performance
design and tuning for the data warehouse. We discuss tuning the population of
the E/R, rather than optimizing query performance against the data warehouse.
Query performance against the data warehouse is typically supported by using
summary tables, analytical data marts, and derived data marts, as we have seen
in Figure 3-19 on page 68.

The two primary processes for populating the E/R model of the warehouse are
batch and real time. In batch mode, the E/R model is typically populated using
custom applications, ETL tools, or native database utilities that deliver good
performance. It is much the same with real time, but with modifications.

In a real-time environment, you typically must change process models. The
objective is to move to more of a continuous load/update scenario. It cannot be
achieved by simply moving data faster. The processes must be changed to
enable the data to be made available faster. New techniques should also be
considered. For example, rather than the typical ETL process many are moving
to an ELT process. That is, the data is extracted and loaded. Then, the required
transformations are performed. This can enable a performance improvement.

Note: Performance is dependent on the physical data structures of RDBMS.
Altering or adding more appropriate physical structures may possibly improve
the performance of query/extraction/replication. However, it may also increase
the load time of the data warehouse. Performance tuning is a cost
minimization issue. For example, performance can always be improved by
adding more CPU and I/O resources. But the objective is to find a compromise
between acceptable performance and total cost of the system.

Four - Verification of physical data modeling process
At the end of physical design, check the following list for completed activities:
1. Generated physical model DDL script should properly define the physical
structures along with performance enhancements.
2. Good documentation of the physical design should exist in the case tool used.


3. Each entity of logical design should represent a physical table with the
appropriate attributes and relations.
4. Each relationship should describe correct cardinalities (one to one, one to n,
and n to n).
5. Each entity and attribute should be properly described in the data dictionary.
6. All capacity estimates should be validated.

In this section, we have discussed the data modeling techniques for data
warehousing. In Chapter 6, “Dimensional Model Design Life Cycle” on page 227,
we discuss in detail the technique for designing a dimensional model.


4

Chapter 4. Data analysis techniques
In this chapter, we focus on data analysis techniques used in data warehousing.
This is, after all, the primary reason for a data warehouse. We do not intend this
to be an all-inclusive treatise on the subject, but an overview so you have a good
general understanding.

We discuss the following topics:
The information pyramid and associated reporting that you can perform
BI reporting tool architectures
Classification of BI users based on analytical needs
Query and reporting
Multidimensional analysis techniques:
– Slice and dice
– Pivoting
– Drill-up, drill-down, and drill-across
– Roll-up and roll-down
Query and reporting tools


4.1 Information pyramid
Every enterprise produces regulatory, statutory, and internal reports on a regular
basis. The reports are predefined documents composed from tables, summaries,
or charts, containing business information consisting of measures and tables
with a description of columns and rows. The target users are auditors,
governmental regulatory institutions, shareholders, internal users from areas
such as finance, or from the top management of a company. This says that there
is a need for information in the form of analytics, dashboards, queries, and
reports by all levels of people in an organization. The focus of the IT department
is to provide for the analysis and business reporting needs of all company
decision makers.

Gone are the days when you could plan and manage business operations
effectively by using monthly batch reports, and when IT organizations had
months to implement new applications. Today companies need to deploy
informational applications rapidly, and provide business users with easy and fast
access to business information that reflects the rapidly changing business
environment. In short, the pressure on the IT department to deliver the
information to the business continues to dramatically increase.

In this section, we focus on different information environments in the
organization, and discuss how to accomplish data analysis and reporting from
each of them.

4.1.1 The information environment
Consider the information pyramid in Figure 4-1 on page 79. The information
pyramid depicts several environments or levels within an organization where the
information may reside, in what form, and for what duration. Traditionally, IT has
seen these levels of data as separate, with the requirement of copying data from
one level to another for appropriate usage.

However, these different levels should be thought of simply as different views of
the same data, although individual users may only focus on a particular level to
perform their specific job. But, to emphasize this difference in perspective, we
have named these levels as floors of data. While data copying may continue
between the floors, this approach is no longer the only one possible.

The data on the different floors of the pyramid does have different
characteristics, such as volumes, structures, and access methods. Because of
that we can choose the best way to physically instantiate the data. And the
pyramid emphasizes that today the requirement is to access data from all floors
within a single activity or process.


We now describe each floor in a bit more detail, along with the part it plays in the
reporting environment. From this discussion, we can see the advantages,
disadvantages, and how the decision-making processes should be architected.

Decision Making

Static
reports,
Dashboards

Floor 5

Dimensional, Data Mart, Strategic
Cubes. Duration: Years
Floor 4

Summarized Data
Performance and Rolled-up Data.
Duration: Years
Floor 3
Near 3rd Normal Form, Subject Area,
Code and Reference tables.
Duration: Years
Floor 2 Tactical

Staging, detail, denormalized, Raw Source
Duration: 60, 120, 180, . . . Days
Floor 1
Operational
Operational Systems
Duration: 60, 120, 180, . . . Days
Floor 0

Figure 4-1 Information pyramid

The concept behind the information pyramid is to map the delivery of data needs
to your applications and organization needs. This is to help determine the data
flow and delivery requirements. We call this right time data delivery. The lower
floors of the architecture offer the freshest, and most detailed data. This is the
source of data for making operational decisions on a day to day basis. As you
move up the floors, the data becomes more summarized and focused for specific
types of analysis and applications. For example in floors 3, 4, and 5, the data is
used for more strategic analysis.

Now we take a look at each of the floors and discuss them in detail.

Floor 0
This floor represents data from the operational or source systems. The data in
the source systems is typically in 3NF and represented in an E/R model. The
source systems are the applications that capture the transactions of the
business. The data is highly normalized so that the transactions performing

Chapter 4. Data analysis techniques 79

inserts, updates, and deletes can be executed quickly. The primary requirements
for these source systems are transactional performance and high availability.
The source systems are periodically purged, so they typically store relatively little
history data. With source systems, you can make operational, day to day
decisions, but they are not formatted or architected for long-term reporting
solutions.

The advantages of querying from the source systems (floor 0) are:
No need for additional hardware.
Quick start.
Less difficult to accomplish.

The disadvantages of querying from the source systems (floor 0) are:
They are not optimized for query processing.
Executing queries or reports against a source system can negatively impact
performance of the operational transactions.
Insufficient history data for general query and reporting.
They need new data structures (such as summary tables) created and
maintained.

Note: Floors 1-5 can broadly be mapped to the layers in existing data
warehouse architectures. These layers are based on the same fundamental
data characteristics that provided the basis for separating the different types of
data when defining the architecture of the data warehouse.

Floor 1
This floor represents the staging area and the denormalized source systems.
The staging area is not appropriate for reporting because the data in the staging
area is prepared for consumption by the business users. On the other hand, a
denormalized source system, which is also often part of the floor 1, may be used
for limited reporting.

The advantages of querying from floor 1 are:
Reporting can be obtained from the denormalized source systems.

The disadvantages of querying from floor 1 are:
Data in the staging area cannot be used for reporting purposes because it
may not have been cleansed or transformed, and therefore, it is not ready to
be used.


Floor 2
As shown in Figure 4-1 on page 79, floor 2 represents the 3NF data warehouse
and associated reference tables. In Figure 4-2, we show an architecture where
users are querying a data warehouse designed in a 3NF E/R model. A
normalized database is typically difficult for business users to understand
because of the number of tables and the large number of relationships among
them.

Source System A
Staging Area EDW

Source System B

Load

User
Source System C Reports

Normalized/De-normalized Normalized
ER Model ER Model

Figure 4-2 Querying a data warehouse

The advantages of querying from floor 2 are:
There are single sources of data.
Floor 2 contains history and detail data.
Floor 2 is optimized for storage of large volumes of data.

The disadvantages of querying from floor 2 (the data warehouse) are :
There is no structure for dimensional analysis, making it difficult to
understand the data relationships.
Complex project, difficult to deliver.

Floor 3
Floor 3 represents summarized data, which is created directly from the
warehouse.

The advantages of querying from the summarized warehouse are:
The history data is easily available.
It is optimized for a set of data.

The disadvantages of querying from the data warehouse are:


You may need to create summaries for several business processes, and
maintaining such summaries becomes a maintenance issue.

Floor 4
On floor 4, there are dimensional data marts stored in relational databases, and
data cubes.

The advantages of querying from the dimensional data mart or cubes are:
History data is easily available.
Highly optimized data is available for each business process.
Queries are easy to understand by the business users.

The disadvantages of querying from the dimensional data mart or cubes are:
You may need to create and maintain several dimensional models for each
business process. This adds to the overall cost of the data warehouse.
For querying cubes, you may need to purchase specialized software.
For creating and storing cubes, you need additional software.
There will be additional storage requirements.
The additional effort to populate data marts.

Source System A
Staging Area EDW Dimensional
Model

Load

Source System B

Load

Source System C

User
Reports
Normalized/De-normalized Normalized
ER Model ER Model

Figure 4-3 Querying from a dimensional model

Floor 5
Floor 5 represents the reporting environment and infrastructure that is used for
making static reports, dynamic ad hoc reports, and dashboards. It is the delivery
platform that supports all the reporting needs.


4.2 BI reporting tool architectures
The different data analysis tools for reporting typically fall into two main
architectures as shown in Figure 4-4. The architectures are:
2-tier: In this architecture, the BI reporting tool is installed on the client
machines and these tools (clients) directly access the data warehouse or the
data marts.
3-tier: In this architecture, the BI reporting server software is installed on a
server machine. All the clients access the data warehouse or the data marts
by using a browser. No special tool must be installed on the clients.

Clients
2-tier
DB2 BI Tool
mart EDW Installed

mart
Servers
mart
3-tier
Clients

DB2 Browser
mart EDW GUI

mart BI Tools Servers
mart and BI Repository

Figure 4-4 BI reporting tool architectures

4.3 Types of BI users
When buying BI reporting tools, it is important to match them with the type of
user who uses them. Each product tends to excel at certain tasks, but may lag in
others. This typically means that the client ends up with 3-4 BI tools to meet the
requirements. Then there is a need for additional skills, education, maintenance,
and licenses, which are all additional expenses. Soon there is a movement to
consolidate those tools to save money. In this case, you must take care to make
sure the business users still have the specific capabilities they need to do their
jobs. Figure 4-5 on page 84 shows classifications of users according to their BI
reporting needs.


Analytic Complexity
SDKs, open standards, data mining,
high IT performance, data sources

What will happen?
Interactive queries, OLAP,
Power
analytic functions, statistics,
Users simulations

Why did it happen?
Static reports, report
Business Users distribution, scorecards,
XL, dashboards

What
happened? Static reports, ease of
Casual Users use, portlets

low
portlets, extranet
Enterprise Users, Consumers, Partners access to company
Web sites

Population size

Figure 4-5 BI user types and their requirements

Based on the analytical needs, we classify the BI users within an organization as
follows:
Enterprise users, consumers, and partners: These users have the lowest
reporting needs. These users typically access data in the form of portlets or
company Web sites which have been given extranet access.
Casual users: Casual users mainly use static reports and portlets. They
require ease of use, and typically prefer to work with standalone applications,
such as spreadsheets.
Business users: Now we are getting to users who typically belong to middle
and executive management. They are more interested in dashboards and
other static types of reports. Business users are involved with both tactical
and strategic decision making processes.
Power users: These users require higher functionality compared with other
user types. They understand the environment, are familiar with the data
analysis tools, and are comfortable with complex applications.
IT users: IT users are responsible for creating the reports for the business. IT
users work with advanced reporting applications, such as software
development kits (SDK) and data mining software. They understand the IT
environment and are comfortable interacting directly with several
heterogeneous data sources.


4.4 Query and reporting
Query analysis and reporting are the processes for posing questions to answer,
retrieving relevant data from the data warehouse, transforming it into the
appropriate context, and displaying it in a readable format. Query analysis and
reporting are primarily driven by analysts who are quite familiar with posing such
queries to determine the answers to their questions.

Traditionally, queries have dealt with two dimensions, or two factors, at a time.
For example, you might ask, "How much of that product has been sold this
week?" Subsequent queries would then be posed to perhaps determine how
much of the product was sold by a particular store. Figure 4-6 depicts the
process flow in query and reporting.

Query definition is the process of taking a business question or hypothesis and
translating it into a query format that can be used by a particular decision support
tool. When the query is executed, the tool generates the appropriate language
commands to access and retrieve the requested data, which is returned in what
is typically called an answer set. The data analyst then performs the required
calculations and manipulations on the answer set to achieve the desired results.
Those results are then formatted to fit into a display or report template that has
been selected for ease of understanding by the user. This template could consist
of combinations of text, graphic images, video, and audio. Finally, the report is
delivered to the user on the desired output medium, which could be printed on
paper, visualized on a computer display device, or presented audibly.

Query
Definition

Data Access
and Retrieval

Calculation
Manipulation

Report
Preparation

Report
Delivery

Figure 4-6 Process of querying and reporting

Users are primarily interested in processing numeric values, which they use to
analyze the behavior of business processes, such as sales revenue and
shipment quantities. They may also calculate, or investigate, quality measures


such as customer satisfaction rates, delays in the business processes, and late
or wrong shipments. They might also analyze the effects of business
transactions or events, analyze trends, or extrapolate their predictions for the
future. Often the data displayed will cause the user to formulate another query to
clarify the answer set or gather more detailed information. This process
continues until the desired results are reached.

A more detailed discussion on different types of reporting and querying tools in
available in 4.6, “Query and reporting tools” on page 94.

4.5 Multidimensional analysis techniques
Multidimensional analysis has become a popular way to extend the capabilities
of query and reporting. That is, rather than submitting multiple queries, data is
structured to enable fast and easy access to answers to the questions that users
typically ask. For example, the data would be structured to include answers to
the question, "How much of each of our products was sold on a particular day, by
a particular salesperson, in a particular store?" Each separate part of that query
is called a dimension. By precalculating answers to each subquery within the
larger context, many answers can be readily available because the results have
been precalculated for each query; they are simply accessed and displayed. For
example, by having the results to the above query, one would automatically have
the answer to any of the subqueries. That is, we would already know the answer
to the subquery, "How much of a particular product was sold by a particular
salesperson?" Having the data categorized by these different factors, or
dimensions, makes it easier to understand, particularly by business-oriented
users of the data. Dimensions can have individual entities, or a hierarchy of
entities, such as region, store, and department.

Multidimensional analysis enables users to look at a large number of
interdependent factors involved in a business problem and to view the data in
complex relationships. Users are interested in exploring the data at different
levels of detail, which is determined dynamically. The complex relationships can
be analyzed through an iterative process that includes drilling down to lower
levels of detail or rolling up to higher levels of summarization and aggregation.

Figure 4-7 on page 87 demonstrates that the user can start by viewing the total
sales for the organization, then drill-down to view the sales by continent, region,
country, and finally by customer. Or, the user could start at customer and roll-up
through the different levels to finally reach total sales. Pivoting in the data can
also be used. This is a data analysis operation where the user takes a different
viewpoint than is typical on the results of the analysis, changing the way the
dimensions are arranged in the result. Like query and reporting, multidimensional
analysis continues until no more drilling down or rolling up is performed.


Figure 4-7 Drill-down and roll-up analysis

Multidimensional analysis enables you to look at the business problem by large
number of interdependent factors describing the matter. In other words,
multidimensional analysis enables you to view the information at different levels
of detail or to analyze complex relationships.

The following are multidimensional techniques that we discuss in more detail:
Slice and dice
Pivoting
Drill-down, drill-up, and drill-across
Roll-down and roll-up

4.5.1 Slice and dice
We start by discussing slice and dice analysis as individual activities.

Slice
The term slice in multidimensional terminology is used to define a member or a
group of members that are separated (from ALL other dimensions) and then
evaluated across all the dimensions. A member of a dimension means a value
inside a column. Slicing is slightly difficult to understand on a two-dimensional
paper. In order to understand the slicing concept, consider a dimensional model
example. Assume that we have only three dimensions named product, store,
and date in a simple dimensional model. In this simple dimensional model, we
just have one fact table with a fact called sales.


Assume that we isolate three members from the product dimension. The three
members we isolated for the product dimension are soda, milk, and juice. This is
shown in Figure 4-8. If we measure the SUM of sales quantity for ALL stores and
for ALL dates across one or more members of one dimension (product in our
case), then this concept is called slicing. The arrow in Figure 4-8 shows that the
sum is across all dates and all stores.

This slice of the product dimension lets us to select our concerned members
(soda, milk, and juice) from the product dimension. The slicing of the members
allows us to focus only on these three members across all other dimensions.
This concept is called slicing.

(For ALL Stores and Dates)

Product Sales in USD
Soda 2,530
Milk 3,858
Juice 15,396
Total 21,784

Figure 4-8 Slice for product

The slice in Figure 4-8 shows that soda generates the smallest sales amount,
milk second, and juice third.

Note: When you slice, you choose one or more members of a dimension and
consolidate (or summarize) across all other dimensions (in our example, the
other dimensions were store and date.)

Dice
The dicing concept means that you put multiple members from a dimension on
an axis and then put multiple members from a different dimension on another
axis. This allows you to view the interrelationship of members from different
dimensions.

Dicing is analysis of interrelationships among different dimensions or their
members. Figure 4-9 on page 89 and Figure 4-10 on page 89 show examples of
dicing.
In Figure 4-9 on page 89, we see multiple members listed vertically for the store
dimension in one axis. These members are CA, OR, and LA. Similarly, we have
multiple members for the date dimension which are listed horizontally. We are
able to view the interrelationship of members from different dimensions. In other


words, we are able to see the relationship between CA and dates 1/1/2005,
1/2/2005, 1/3/2005, and vice versa.

LA

Figure 4-9 Dice for store and date

Another example of dicing is shown in Figure 4-10.

Milk

LA

Figure 4-10 Dice of store and product dimension

In this example, we can see the interrelationship between the members of the
store and product dimensions. Here we analyze:
How each store contributes to total sales amounts for each product (Soda,
Milk, and Juice).
How a particular product contributes to total sales for each store location.

Note: You dice when you choose one or more members of same dimension
on one axis and on the other axis you choose a member or members from
another dimension. Now you can analyze interrelationships of those
dimensions.


4.5.2 Pivoting
Pivoting in multidimensional modeling means exchanging rows with columns
and vice versa. Figure 4-11 on page 90 shows an example of pivoting. We
exchange the store rows with columns of the product dimension members. It is
simply a quick way to view the same data from a different perspective.

Milk

LA

Piv
o t

LA

Milk

Figure 4-11 Pivoting

Note: You pivot when you exchange the axes of the report.

4.5.3 Drill-down and drill-up
Drilling in multidimensional terminology means going from one hierarchy level to
another. In other words, drill-down can be defined as the capability to browse
through information, following a hierarchical structure.

In the example shown in Figure 4-12 on page 91, we show drilling down through
a simple three level hierarchy present in the product dimension. The hierarchy is


‘Group Class’ → ‘Group’ → ‘Product’. When we drill-down the Group Class
attribute, we reach the Group. Finally by drilling down on the Group attribute, we
reach the lowest detail present inside the product dimension (which is the
individual product) as shown in Figure 4-12 on page 91.

LA

Beverage ->Group
LA

Beverage ->Group -> Pop
LA

Milk

Figure 4-12 Drill-down on product dimension

Note: We consider drilling up and drilling down when we want to analyze the
subject at different levels of detail. Drilling is possible if a dimension contains a
multiple level hierarchy.

Another example of drill-down is shown in Figure 4-13 on page 92. Here we drill
down from total sales in the US to sales at a particular store in Buffalo.


Figure 4-13 Drill-down example

Drill-up is exactly the opposite of drill-down.

4.5.4 Drill-across
Drill-across is a method where you drill from one dimension to another. You must
define the drill-across path. This function is often used in ROLAP. In Figure 4-14,
you see the result of drill-across from store CA to the product dimension. The first
chart depicts the sales in stores in three different states. And, in particular, we
have focused on CA (California).

LA

CA Product

Milk

Figure 4-14 Drill-across result


By drilling across to the product dimension, we can see the details about which
products comprised the sales for the store CA.

4.5.5 Roll-down and Roll-up
Roll-down and roll-up are OLAP functions that give the higher or lower aggregate
over whole dimension at a given hierarchy level.

In the example in Figure 4-15, we roll-down the product dimension from level 3,
to level 2, and to level 1. This is done through the product hierarchy level: “Group
class” → “Group” → “Product”.

The roll-down concept is the same as a drill-down.

3

2

1
Roll-Down

Roll-Up

Milk

Figure 4-15 Roll-down, Roll-up

Roll-up is exactly opposite to roll-down. The arrows in Figure 4-15 show the roll
directions. The roll-up concept is the same as drill-up.


4.6 Query and reporting tools
In this section, we discuss various query and reporting tools you can use to
generate reports. We categorize various tools that can help in creating reports as
follows:
SQL query language using select statement, views, or stored procedures
Spreadsheets
Reporting Applications (Client-Server and Web-based)
Dashboard and scorecard applications
Data Mining tools

4.6.1 SQL query language
Select
In the RDBMS environment, the basic query reporting language is the structured
query language (SQL) which enables you to query data in a database. In
Example 4-1, we show an SQL select returning daily sales values for Product
with subtotals and totals grouped by Date, Product Group, and Products.

Example 4-1 SQL select
select D.DAY_AD as DAY,P.SUB_PRODUCT_GROUP as GROUP,P.PRODUCTS,
sum(S.VALUE) Sale_Amount
from SALE S
join PRODUCT P
on (S.P_ID = P.P_ID)
join DATE D
on (S.D_ID = D.D_ID)
where P.SUB_PRODUCT_GROUP in ('Pop')
group by rollup(D.DAY_AD,P.SUB_PRODUCT_GROUP,P.PRODUCTS)
order by 1,2,3,4;

The output of the Select statement is shown in Example 4-2.

Example 4-2 Output of select
DAY GROUP PRODUCTS SALE_AMOUNT
---------- -------- --------- -------------
2005-01-01 Pop Milk 1141
2005-01-01 Pop Juice 2580
2005-01-01 Pop - 3721
2005-01-01 - - 3721
2005-01-02 Pop Milk 1431
2005-01-02 Pop Juice 8449
2005-01-02 Pop - 9880
2005-01-02 - - 9880


2005-01-03 Pop Milk 1286
2005-01-03 Pop Juice 4367
2005-01-03 Pop - 5653
2005-01-03 - - 5653
- - - 19254
13 record(s) selected.

Views
When an SQL select is complex and used very often in users’ queries or by many
applications, consider embedding it in an SQL view, as shown in Example 4-3.

Example 4-3 SQL view
create view sales_by_date_Pop as (
select D.DAY_AD as DAY,P.SUB_PRODUCT_GROUP as GROUP,P.PRODUCTS,
sum(S.VALUE) as Sale_Amount
from SALE S
join PRODUCT P
on (S.P_ID = P.P_ID)
join DATE D
on (S.D_ID = D.D_ID)
where P.SUB_PRODUCT_GROUP in ('Pop')
group by rollup(D.DAY_AD,P.SUB_PRODUCT_GROUP,P.PRODUCTS)
)

Basically the views are helpful in hiding the complexity of the SQL statement.
The user only needs to select the output from the view (select * from
sales_by_date_pop) instead of writing the long SQL statement, such as the one
in Example 4-3.

RDBMS stored procedures
Stored procedures are programs that use data from RDBMS and can be
executed from the RDMS environment. There are generally two types of RDBMS
procedures, PL SQL-based and External (written in a higher programing
language, such as C, Cobol, or BASIC). In Example 4-4, we show a sample DB2
PL SQL-stored procedure using an SQL select code with parameter “Group”. In
Example 4-4, we use the create procedure command and pass the procedure
with an input parameter specifying a product group. The procedure returns sales
amount for all dates for the specific product group defined by parameter. The
syntax of command is: call sales_by_date('Pop').

Example 4-4 Stored procedure
CREATE PROCEDURE sales_by_date (IN i_GROUP char(35))
RESULT SETS 1
LANGUAGE SQL


SPECIFIC SELECT_LIST
READS SQL DATA
DETERMINISTIC
BEGIN
DECLARE c1 CURSOR WITH RETURN FOR
select D.DAY_AD as DAY, P.SUB_PRODUCT_GROUP as GROUP, P.PRODUCTS, sum(S.VALUE)
as Sale_Amount
from SALE S
join PRODUCT P on (S.P_ID = P.P_ID)
join DATE D on (S.D_ID = D.D_ID)
where P.SUB_PRODUCT_GROUP in (i_GROUP)
group by rollup(D.DAY_AD, P.SUB_PRODUCT_GROUP,P.PRODUCTS) order by 1,2,3,4;
open c1;
END@

And the result we get is shown in Example 4-5.

Example 4-5 PL SQL stored procedure output for parameter “Pop”
Result set 1
--------------
---------- -------- --------- -------------
2005-01-01 Pop Milk 1141
2005-01-01 Pop Juice 2580
2005-01-01 Pop - 3721
2005-01-01 - - 3721
2005-01-02 Pop Milk 1431
2005-01-02 Pop Juice 8449
2005-01-02 Pop - 9880
2005-01-02 - - 9880
2005-01-03 Pop Milk 1286
2005-01-03 Pop Juice 4367
2005-01-03 Pop - 5653
2005-01-03 - - 5653
- - - 19254
Return Status = 0

A different SQL command call sales_by_date('Water') gives us a different
result set for Product group Water as shown in Example 4-6.

Example 4-6 PL SQL stored procedure output for parameter Water
Result set 1
--------------

---------- -------- --------- -----------


2005-01-01 Water Soda 1017
2005-01-01 Water - 1017
2005-01-01 - - 1017
2005-01-02 Water Soda 928
2005-01-02 Water - 928
2005-01-02 - - 928
2005-01-03 Water Soda 585
2005-01-03 Water - 585
2005-01-03 - - 585
- - - 2530


Return Status = 0

Note: Views and stored procedures are also used in the query and reporting
environment to control data access for tables and their rows.

4.6.2 Spreadsheets
One of the most widely used tools for analysis is the spreadsheet. It is a very
flexible and powerful tool; and therefore, you can find it in almost every
enterprise around the world. This is a good-news and bad-news situation. It is
good, because it empowers users to be more self-sufficient. It is bad, because it
can result in a multitude of independent (non-integrated and non-shared) data
sources that exist in any enterprise.

Here are a few examples of spreadsheet use:
Finance reports, such as a price list or inventory
Analytic and mathematical functions
Statistical process control, which is often used in manufacturing to monitor
and control quality

4.6.3 Reporting applications
There are many analytical query and reporting applications in the market which
work with different relational databases or with special multidimensional
structures, such as cubes. These applications include:
IBM Alfablox
DB2 OLAP Server™
Microstrategy


Hyperion Essbase
Brio
BusinessObjects
Cognos Impromptu
Crystal Reports

In Table 4-1, we list important criteria for you to consider when you are selecting
reporting applications for your company.

Table 4-1 Selection criteria for reporting applications
Report authoring and Drag and drop creation, multiple sources, mixed tables,
formatting graphs, tabular formats, style sheets, WYSIWYG, print
control, sorting, invoices, and labels

Report distribution Time or event scheduled; table of contents navigation;
formats, such as PDF, HTML, and Excel®; alerts

Analytical functions Running totals, percent of total, Euro conversion, ranking,
highlight exceptions, mining, and binning

Query interaction Ease of use, shield user from SQL and database
navigation complexity, and modify and reuse existing
queries

OLAP functionality Hierarchical summaries, drill-down, drill-up, and view
pivoting

SDKs and APIs Data sources, language libraries, embedded reporting,
MDX support, and performance

Security Report element security

Aggregates Aggregate awareness

4.6.4 Dashboard and scorecard applications
The most common mechanisms for viewing performance data are dashboards
and scorecards. Top and middle management are not the target users for
analytical tools. Their requirements are for viewing performance data at high
levels. Dashboards provide the management with a high level view of the data.

A dashboard provides a graphical user interface that can be personalized to suit
the needs of the user. A dashboard graphically displays scorecards that show
performance measurements, together with a comparison of these measurements
against business goals and objectives.

Note: Top executives are the target users for dashboard applications.


In Figure 4-16 we show a business process management (BPM) dashboard
example from the insurance industry. It is not critical to read the values on the
dashboard, but to understand the concept. It gives management critical data on a
number of strategic elements that require special focus and monitoring. For
example, it shows new business growth by category. Management can monitor
these elements to make sure they are in line with the business goals and
strategy. If not, management can take immediate action.

The dashboard also gives a current status on a number of projects with
appropriate alerts.

Figure 4-16 Insurance dashboard

In Figure 4-17 on page 100, we show an example retail dashboard. It shows a list
of the key business areas and the appropriate alert status for each area. There is
also a summary of the current business financial status (shown in the inset). With


this type of information, management now has the capability to not just monitor,
but also to impact the measurements.

Figure 4-17 Retail Dashboard

Note: The usage of dashboards to help in business performance
management is discussed in more detail in the IBM Redbook, Business
Performance Management...Meets Business Intelligence, SG24-6340.

4.6.5 Data mining applications
Data mining is a relatively new data analysis technique. It is very different from
query and reporting and multidimensional analysis in that it uses what is called a
discovery technique. That is, you do not ask a particular question of the data but
rather use specific algorithms that analyze the data and report what they have
discovered. Unlike query and reporting and multidimensional analysis where the
user has to create and execute queries based on hypotheses, data mining
searches for answers to questions that may have not been previously asked.
This discovery could take the form of finding significance in relationships
between certain data elements, a clustering together of specific data elements,


or other patterns in the usage of specific sets of data elements. After finding
these patterns, the algorithms can infer rules. These rules can then be used to
generate a model that can predict a desired behavior, identify relationships
among the data, discover patterns, and group clusters of records with similar
attributes.

Data mining is most typically used for statistical data analysis and knowledge
discovery. Statistical data analysis detects unusual patterns in data and applies
statistical and mathematical modeling techniques to explain the patterns. The
models are then used to forecast and predict. Types of statistical data analysis
techniques include:
Linear and nonlinear analysis
Regression analysis
Multi-variant analysis
Time series analysis

Knowledge discovery extracts implicit, previously unknown information from the
data. This often results in uncovering unknown business facts.

Data mining is data driven (see Figure 4-18). There is a high level of complexity
in stored data and data interrelationships in the data warehouse that are difficult
to discover without data mining. Data mining provides new insights into the
business that may not be discovered with query and reporting or
multidimensional analysis alone. Data mining can help discover new insights
about the business by giving us answers to questions we might never have
thought to ask.

Analyst Analyst Data
Driven Assisted Driven

Query
Multidimensional Data
And
Analysis Mining
Reporting

Figure 4-18 Data Mining focuses on analyzing the data content rather than responding to
questions

Some data mining tools available in the market are:
The IBM product for data mining is DB2 Intelligent Miner™, which includes IM
Modeling, IM Scoring, and IM Visualization.


SAS Institute’s Enterprise Miner.
Microsoft Analysis Services.
Oracle Data mining (separately purchased option within Oracle 10g
Enterprise Edition).


5

Chapter 5. Modeling considerations
In this chapter, we discuss considerations, and challenges that may arise, when
designing dimensional models. As examples:
Converting an E/R model to dimensional model. How do you identify fact and
dimension tables from an E/R model? And, how do you convert an E/R model
to a dimensional model?
Identifying the grain.
Working with degenerate dimensions, dimension hierarchies, time as a fact or
dimension, slowly changing dimensions, fast changing dimensions,
identifying and handling snowflakes, identifying garbage dimensions,
handling multi-valued dimensions, use of bridge tables, handling
heterogeneous products, and handling hot swappable dimensions (also
referred as profile tables).
Working with additive and semi-additive facts, composite key design, and
event fact tables.
What about physical design activities, such as indexing?
Working with changes to data, structure, and requirements.


5.1 Converting an E/R model to a dimensional model
In this section we describe how to convert an E/R model to a dimensional model.
A dimensional model can be created from the enterprise data warehouse or
directly from OLTP source systems. For additional information, refer to:
Enterprise data warehouse: “Data warehouse architecture choices” on
page 57.
OLTP Source systems: “Data modeling: The organizing structure” on page 47
in the following sections:
– “Independent data mart architecture” on page 59
– “Dependent data mart architecture” on page 61

The following are the steps for converting an E/R model to a dimensional model:
Identify the business process from the E/R model.
Identify many-to-many tables in the E/R model to convert to fact tables.
Denormalize remaining tables into flat dimension tables.
Identify date and time from the E/R model.

The steps are explained in detail in the following sections.

5.1.1 Identify the business process from the E/R model
It is important to understand that an E/R model can be segmented into
multiple dimensional models. An E/R model (which may be an enterprise data
warehouse or an OLTP source system) consists of several business
processes. This is depicted in Figure 5-1 on page 105. For example, an E/R
model for an ERP system includes several business processes, such as retail
sales, order management, procurement, inventory, and store and warehouse
inventory management.


Business
Process #1

Business Business
Process #3 Process #2

Figure 5-1 E/R model consists of several business processes

5.1.2 Identify many-to-many tables in E/R model
Once the business processes are separated, the next step is to identify the
many-to-many tables (many-to-many relationships) in the E/R model and convert
them to dimensional model fact tables. These many-to-many relationships
contain numeric and additive non-key facts which generally become facts inside
the fact table.

The idea behind this step is to identify the transaction-based tables that serve to
express many-to-many relationships inside an E/R model.

Every E/R model consists of transaction-based tables which constantly have
data inserted, or are updated with data, or have data deleted from them. Some of
these tables also express a many-to-many relationship. For example, in an ERP
database, there are transaction tables, such as Invoice and Invoice_Details,
which are constantly inserted and updated because they are transaction-based
tables. However, tables such as Employee and Products in an E/R model may
be fairly static.

Description of many-to-many relationships
Many-to-many (m:n) relationships add complexity and confusion to the model
and to the application development process. The key to resolving m:n
relationships is to separate the two entities and create two one-to-many (1:n)

Chapter 5. Modeling considerations 105

relationships between them with a third intersect entity. The intersect entity
usually contains attributes from both connecting entities.

To resolve an m:n relationship, analyze the business rules again. Have you
accurately diagrammed the relationship? The telephone directory example has a
m:n relationship between the name and fax entities, as Figure 5-2 depicts. The
business rules say, “One person can have zero, one, or many fax numbers; a fax
number can be for several people.” Based on what we selected earlier as our
primary key for the voice entity, an m:n relationship exists.

A problem exists in the fax entity because the telephone number, which is
designated as the primary key, can appear more than one time in the fax entity;
this violates the qualification of a primary key. Remember, the primary key must
be unique.

Figure 5-2 Telephone directory diagram to show many to many relationship

To resolve this m:n relationship, you can add an intersect entity between the
name and fax entities, as depicted in Figure 5-3 on page 107. The new intersect
entity, fax name, contains two attributes, fax_num and rec_num.


Figure 5-3 Many-to-many relationship

Another example for many-to-many relationship is shown in Figure 5-4.

Employee Has Projects

Programming
Skills

Figure 5-4 Many-to-many relationship

In Figure 5-4, the many-to-many relationship shows that employees use many
programming skills on many projects and each project has many employees with
varying programming skills.

5.1.3 Denormalize remaining tables into flat dimension tables
The final step involves taking the remaining tables in the E/R model and
denormalizing them into dimension tables for the dimensional model. The
primary key of each of the dimensions is made a surrogate (non-intelligent,
integer) key. This surrogate key connects directly to the fact table.


5.1.4 Identify date and time dimension from E/R model
The last step generally involves identifying the date and time dimension. Dates
are generally stored in the form of a date timestamp column inside the E/R
model. You will observe that date and time-related columns are generally found
in the transaction-based tables.

We explain the process of converting an E/R model to a dimensional model in the
section below.

Example: An E/R model conversion
We convert the E/R model shown in Figure 5-5 to a dimensional model using the
following steps:
1. Identify the business process: We discussed this step in detail in “Identify
the business process from the E/R model” on page 104. The business
process we identified for our example is retail sales. The E/R model for this
retail sales schema is shown in Figure 5-5.

Figure 5-5 E/R model for retail sales business process

2. Identify the many-to-many tables in the E/R model to convert them to fact
tables. After identifying the business process as retail sales, and identifying
the E/R model as shown in Figure 5-5, the next step is to identify the


many-to-many relationships that exist inside the model. In order to find this,
we must segregate the tables inside the E/R model into two types.
Transaction-based tables and Non-Transaction based tables, as shown in
Table 5-1.

Note: Fact tables in a dimensional model express the many-to-many
relationships between dimensions. This means that the foreign keys in the fact
tables share a many-to-many relationship.

What is a transaction-based table? In an E/R model, it is one which is
generally involved in storing facts and measures about the business. Such
tables generally store foreign keys and facts, such as quantity, sales price,
profit, unit price, and discount. In transaction tables, records are usually
inserted, updated, and deleted as and when the transactions occur. Such
tables also in many ways represent many-to-many relationships between
non-transaction-based tables. Such tables are larger in volume and grow in
size much faster than the non-transaction-based tables.
What is a non-transaction-based table? This is an E/R model which is
generally involved in storing descriptions about the business. Such tables
describe entities such as products, product category, product brand,
customer, employees, regions, locations, services, departments, and
territories. In non-transaction tables, records are usually inserted and there
are fewer updates and deletes. Such tables are far smaller in volume and
grow very slowly in size, compared to the transaction-based tables.

Table 5-1 Transaction and Non-transaction tables in the E/R model
Transaction tables Non-transaction tables

Store_BILLING Employees: Stores information about
[This transaction table stores the employees.
BILL_NUMBER, Store Billing date, and
time information. The detailed Bill
information is contained inside the
Store_Billing_Details table.]

Store_Billing_Details Department: Stores department
[This transaction table stores the details information for every employee.
for each store bill.]

Store: Stores the name of the store and
type of store information.

Store_Region: Stores the region, county,
state, and country to which the store
belongs.


Transaction tables Non-transaction tables

Suppliers: Stores the supplier name and
supplier manager-related information.

Supplier_Type: Stores supplier type
information.

Products: Stores information relating to
products.

Brand: Stores brand information to which
different products belong.

Categories: Stores categories information
to which different product brands belong.

Packaging: Stores packaging-related
information for each product.

Customers: Stores customer-related
information.

Region: Stores different regions to which
customers belong.

Territories: Stores territories to which
different customers belong.

Customer_Type: Stores customer
classification information.

The transaction tables are identified in Table 5-1 on page 109, and we depict
them in Figure 5-6 on page 111.


Transaction Based Tables

Figure 5-6 Identifying transaction-based tables in the E/R model

The Store_BILLING and Store_Billing_Details tables express a many-to-many
relationship that exists between Employee, Store, Customers, and Products.
Some of the relationships are explained below:
Each Employee can sell many products and each Product could be sold by
many Employees.
Each Store could sell many Products and each Product could be sold by
many Stores.
Each Customer could buy many products and each product could be bought
by many Customers.
Each Employee could sell to many Customers and each Customer could
purchase from many Employees.

Figure 5-7 on page 112 shows that the Store_BILLING and Store_Billing_Details
transaction tables express many-to-many relationships between the different
non-transaction tables.

Note: The many-to-many tables in the E/R model are converted to
dimensional model fact tables. The many-to-many tables in the E/R model are
generally the transaction-oriented tables.


Transaction Based Tables
Supplier

Products
Employee

Store Customer

Figure 5-7 Identifying many-to-many relationships in an E/R model

The Store_BILLING and Store_Billing_Details tables are the tables that identify
the many-to-many relationship between Employee, Products, Store, Customer,
and Supplier tables. The Store_BILLING table stores billing details for an order.

After having identified the many-to-many relationships in the E/R model, we are
able to identify the fact table as shown in Figure 5-8.

Foreign Keys

Degenerate
Dimension

Facts

Note: Store_BILLING and
Store_Billing_Details tables
convert to a Fact table

Figure 5-8 Fact table

The fact table design may change depending upon the grain chosen.
3. Denormalize remaining tables into flat dimension tables: After we
identified the fact table in Step 2, the next step is to take the remaining tables


in the E/R model and denormalize them into dimension tables. The primary
key of each of the dimensions is made a surrogate (non-intelligent, integer)
key. This surrogate key connects directly to the fact table.
Table 5-2 shows the various E/R tables (see E/R model in Figure 5-6 on
page 111) that have been denormalized into dimension tables.

Table 5-2 E/R model to dimension model conversion
Name of tables in E/R model Corresponding Refer to
denormalized figure
dimension table

Customers, Region, Territories, and Customer Figure 5-9
Customer_Type

Products, Brand, Categories, and Packaging Product Figure 5-10

Suppliers and Supplier_Type Suppliers Figure 5-11

Employees and Department Employees Figure 5-12

Store and Store_Region Store Figure 5-13

Figure 5-9 on page 114 shows that the Customers, Region, Territories, and
Customer_Type tables in the E/R model are denormalized to form a Customer
dimension table. The Customer dimension table has a surrogate key.


(Denormalization of Customer
Tables in the E/R Model)

Figure 5-9 Customer dimension table

Figure 5-10 on page 115 shows that the Products, Brand, Categories, and
Packaging tables in the E/R model are denormalized to form a product
dimension table. The product dimension table has a surrogate key.


Supplier is handled as a
Separate Supplier Dimension

(Denormalization of
Product Table in the
E/R Model)

Figure 5-10 Product dimension

Figure 5-11 shows that the Suppliers and Supplier_Type tables in the E/R
model are denormalized to form a Supplier dimension table. The Supplier
dimension table has a surrogate key.

(Denormalization of Supplier
Tables in the E/R Model)

Figure 5-11 Supplier dimension


Figure 5-12 shows that the Employees and Department tables in the E/R
model are denormalized to form a Supplier dimension table. The Supplier
dimension table has a surrogate key.

(Denormalization of Employee
and Department Tables in the
E/R Model)

Figure 5-12 Employee dimension

Figure 5-13 shows that the Store and Store_Region tables in the E/R model
are denormalized to form a Store dimension table. The Store dimension table
has a surrogate key.

(Denormalization of Store
and Store_Region Tables
in the E/R Model)

Figure 5-13 Store dimension table

The resulting dimensional model after we normalize the dimension tables is
shown in Figure 5-14 on page 117.


Figure 5-14 Dimensional model after step 3

4. Add date and time dimensions: The last step involves identifying the date
and time dimension. Dates are generally stored in the form of a date
timestamp column inside the E/R model.
The date and time are stored in the columns called Store_Billing_Date and
Store_Billing_Time of the Store_BILLING table of the E/R model as shown in
Figure 5-15.

Store Billing Date

Store Billing Time

Figure 5-15 Identifying date and time in the E/R model

The E/R models typically have some form of dates associated with them. These
dates are generally stored in the transaction-based tables in the form of a date
timestamp column. There may also be time associated with the business which
is also stored in the form of a date timestamp column present inside the
transaction-based tables.

After the date and time columns are identified in the E/R model, they are usually
modeled as separate date and time dimensions as shown in Figure 5-16 on
page 118.


The final dimensional model
The dimensional model that results from the steps taken is shown in Figure 5-16.

Figure 5-16 The dimensional model

5.2 Identifying the grain for the model
The lowest level of data represented in a fact table is defined as grain. The focus
of this section is to discuss the Identify the grain component in the DMDL, as
depicted in Figure 5-17 on page 119.


Identify Model Components
Requirements Grain Dimensions Facts
Determine All
Document/Study Identify Fact Table Identify Facts
Dimensions
Enterprise Business Granularity
Processes
Identify Degenerate and

(Indexing, Partitioning and Aggregation)
Identify Conformed

Verify Design with User Requirements
Select Business Identify Multiple Facts

Physical Design Considerations
Requirement Gathering Report
Process to Model Separate Grains Identify Dimensional
for a Single Attributes (Granularity)

Grain Definition Report
Identify High level Business Process and Attribute Hierarchies Identify Fact types
Entities and (Additive, Semi Additive,
Measures for Non-Additive, Derived,
Identify Date and Time
Conformance Identify the Fact Textual, Pseudo, or
Granularity
Table Types Fact-less Facts) and
Identify Data (Transaction, Default Aggregate Rules
Identify Slowly Changing
Sources Periodic, and Dimensions
Accumulating)
Select Requirements Year-to-date Facts
Identify Fast Changing
Gathering Approach Dimensions
(Source Driven
Or Check Grain for Event Fact Tables
Atomicity Identify cases for
User Driven) Snowflaking
Composite Key Design
Requirements Identify preliminary Dimensional Challenges
Gathering candidates for (Multi-valued, Garbage,
dimensions and Heterogeneous, Hot Fact Table
Requirements facts from the grain Swappable, Sizing and Growth
Analysis Roleplaying)

Iterate
Metadata Management

Figure 5-17 Dimensional Model Design Life Cycle

In this section we discuss the importance of having the grain defined at the most
detailed, or atomic, level. When data is defined at a very detailed level, the grain
is said to be high. When there is less detailed data, the grain is said to be low. For
example, for date, a grain of year is a low grain, and a grain of day is a high grain.
We also discuss when to consider separate grains for a single business process.

5.2.1 Handling multiple, separate grains for a business process
Typically separate business processes always require separate dimensional
models with unique grain definitions. Each business process consists of several
facts and dimensions which are different from the other business processes. As
discussed in Chapter 6, “Dimensional Model Design Life Cycle” on page 227, we
take the following steps to create a dimensional model:
Identify business process.
Identify grain.
Identify dimensions.
Identify facts.


We explain the concept of multiple fact table grains in the following steps:
1. Business process: Assume that we are creating a dimensional model for the
retail sales business of a big clothing merchandise brand which has stores all
over the U.S. The business is interested in tracking the sales of the goods
from all of its stores. It is also interested in analyzing the reasons for all of its
returned clothing products at all stores and in analyzing all suppliers of
products based on the percentage of defective returned goods.
2. Identify the grain for the retail sales.
There are two separate grain definitions for the retail sales business process:
– For tracking sales in all clothing stores, the grain definition is: One single
clothing line item on a bill.
– For tracking returned clothing goods in all stores, the grain definition is:
Every individual clothing item returned by a customer to any store.
3. Identify the dimensions for the different grains.
– For grain 1 (sales tracking): product, time, customer, date, employee,
supplier, and store
– For grain 2 (returned goods tracking): return date, purchase date,
customer, store purchased, products returned, and reasons for return.
4. Identify the facts for the different grains.
– For grain 1 (sales tracking): unit price, discount, quantity sold, and
revenue
– For grain 2 (returned goods tracking): revenue returned and quantity
returned

Note: A single business process may consist of more than one dimensional
model. Do not force fit the different facts and dimensions which belong to
different dimensional models into a single star schema. We strongly
recommend that separate grains are identified for a business process when
you are not able to fit facts or dimensions in a single star model.

Figure 5-18 on page 121 shows the dimensional model designed for the sales
tracking in the retail sales business process having the grain equivalent to one
single line item on a bill.


Figure 5-18 Retail sales business star schema for sales tracking

Figure 5-19 shows the dimensional model designed for the retail sales business
process having the grain equivalent to every individual item returned by a
customer to any store.

Figure 5-19 Retail Sales Business star schema for tracking returned goods

When to create separate fact tables
When designing the dimensional model for a business process, one or more fact
tables can be created. Here are guidelines to consider when deciding to make
one or more fact tables for designing the dimensional model for the business
process:
Facts that are not true (valid) to any given grain should not be forced into the
dimensional model. Often facts that are not true to a grain definition belong to
a separate fact table with its own grain definition.
Dimensions that are not true (valid) to any given grain should not be forced
into the dimensional model. Often such dimensions belong to a separate
dimensional model with its own fact table and grain.
Separate fact tables (dimensional models) should always be created for each
unique business process.


5.2.2 Importance of detailed atomic grain
The granularity may be defined as the level of detail made available in the
dimensional model. The grain definition is extremely significant from a business,
technical, and data mart project standpoint.

It is extremely important that the grain definition is chosen at the most detailed
atomic level. The atomic grain is important from three broad perspectives as we
discuss below:
From a business perspective:
From a business perspective, the grain of the fact table dictates whether or
not we can easily extend the dimensional model to add new dimensions or
facts as, and when, the business requirements change. A dimensional model
designed at the lowest level grain (detail) is easy to change. New dimensions
or facts can be added to the existing dimensional model without any change
to the fact table grain, which means that new business requirements can be
delivered from existing dimensional models without the need of much change
to an existing one. This is good for businesses whose requirements typically
change.
The dimensional model should be designed at the most detailed atomic level
even if the business requires less detailed data. This way the dimensional
model has potential future capability and flexibility (regardless of the initial
business requirements) to answer questions at a lower level of detail.
To summarize the importance of grain from a business perspective, we look
at an example. Assume your organization wants to analyze customer buying
trends by product line and region so that you can develop more effective
marketing strategies. Consider the following options:
– Customer by Product
The granularity of the fact table always represents the lowest level for
each corresponding dimension. When you review the information from the
business process, the granularity for customer and product dimensions of
the fact table are apparent. Customer and product cannot be reasonably
reduced any further. That is, they already express the lowest level of an
individual record for the fact table. In some cases, product might be further
reduced to the level of product component because a product could be
made up of multiple components.
– Customer by Product by District
Because the customer buying trends that your organization wants to
analyze include a geographical component, you still need to decide the
lowest level for region information. The business process indicates that in
the past, sales districts were divided by city. But now your organization
distinguishes between two regions for the customer base: Region 1 for


California and Region 2 for all other states. Nonetheless, at the lowest
level, your organization still includes sales district data, so district
represents the lowest level for geographical information and provides a
third component to further define the granularity of the fact table.
– Customer by Product by District by Day
Customer-buying trends always occur over time, so the granularity of the
fact table must include a time component. Suppose your organization
decides to create reports by week, accounting period, month, quarter, or
year. At the lowest level, you probably want to choose a base granularity
of day. This granularity allows your business to compare sales on
Tuesdays with sales on Fridays, compare sales for the first day of each
month, and so forth. The granularity of the fact table is now complete.
The decision to choose a granularity of day means that each record in the
time dimension table represents a day. In terms of the storage
requirements, even 10 years of daily data is only about 3,650 records,
which is a relatively small dimension table.
From a technical perspective:
– From a technical perspective, the grain of the fact table has a major
impact on the size of the star schema. The more atomic the grain, the
bigger the size.
– The atomic grain results in a huge dimensional model, which can impact
the operating cost for performing related tasks such as ETL, as well as the
performance.
From a project task lists perspective:
– The most detailed atomic grain means that the project team needs to
represent the data in more detail. This means that the data mart
development team will need to understand more E/R or data
warehouse-related tables and their corresponding attributes.
– The more detailed the grain, the more complex related procedures, such
as ETL. This means designing more complex procedures and also
maintaining more complex meta data.

Note: If the dimensional model is designed with a higher level grain (meanin it
is highly aggregated and therefore contains less detailed data), there will be
fewer dimensions available for use to perform detailed analyses.

Factors to consider when deciding the grain, are as follows:
Current Business Requirements: The primary factor to consider while
deciding the dimensional model grain is the current business requirement.
The basic minimum need of the dimensional model grain is to be able to


answer the current business requirements. It is important to remember that
the primary purpose for developing a dimensional model is to satisfy a set of
business requirements and therefore the grain should be kept at a level of
detail which satisfies those requirements.
Future Business Requirements: Another important factor to always
consider are the future business requirements. For example, if you design the
model to be based on a weekly grain because the business requires only
weekly data for its reports, then you may not be able to get reports based on
daily data if the business requires daily data in the future. It is important to
understand the potential future needs and perhaps design the model at a
lower grain than dictated by the present requirements. However, storing the
data at a more detailed grain means spending more on maintaining the more
atomic grain, but without any present business value. So, it is important that
when you design a model with more detailed grain than currently required,
you understand the additional overhead costs and prepare satisfactory
justification.

How granularity affects the size of the database
The granularity of the fact table also determines how much storage space will be
required for the database. For example, consider the following possible
granularities for a fact table:
Product by day by region
Product by month by region

The size of a database that has a granularity of product by day by region would
be much greater than a database with a granularity of product by month by
region because the database contains records for every transaction made each
day as opposed to a monthly summary of the transactions. You must carefully
determine the granularity of your fact table because too fine a granularity could
result in a huge database. Conversely, too coarse a granularity could mean the
data is not detailed enough for users to perform meaningful queries.

5.2.3 Designing different grains for different fact table types
In the retail sales example, the grain is of the transaction fact table type.

There are three types of fact tables. They are:
Transaction fact table
Periodic fact table
Accumulating fact table

The reason to show different types of fact tables is to emphasize that each
typically has different types of grain associated with it. It is important that the


designer be aware of these fact table types so that the designer can use the most
appropriate type of the fact tables.

There are also differences in ways the inserts and updates occur inside each of
these fact tables. For example, with the transaction and periodic fact tables, only
inserts take place and no rows are updated. With the accumulating fact table, the
row is first inserted, and then subsequently updated as a milestone is achieved
and facts are made available.

We now discuss each of the different fact tables.

Transaction fact table
A transaction-based fact table is a table that records one row per transaction. An
example of a transaction-based fact table is shown in Figure 5-20 on page 126.
Here the fact table records all transactions that happen for an individual account
of a customer of a bank. Assume that a customer named Sherpa makes the
following transactions in the month of August 2005 against his bank account:
Money Withdrawn: $400, Date: August 2, 2005, Time: 4:00AM
Money Deposited: $300, Date: August 4, 2005, Time: 3:00AM
Money Withdrawn: $600, Date: August 5, 2005, Time: 2:00PM

Figure 5-20 on page 126 shows that in a transaction fact table, a single row is
inserted for each bank account transaction (deposit or withdrawal) that Sherpa
makes. Important features about the transaction fact table are:
A single row is inserted for each transaction.
Typically, the date and time dimensions are represented at the lowest level of
detail. For example, the date dimension may be represented at the day level
and the time dimension may be represented at the hour or minute level.
The transaction fact table is known to grow very fast as the number of
transactions increases.


(Deposit or Withdrawal)
Transaction Type

6 Rows inserted (1 row for each Transaction)
Row 1: Withdrawal: $400, Date: 2nd August 2,2005, Time: 4:00AM
Row 2: Deposit: $300, Date: 4th August 4,2005, Time: 3:00AM
Row 3: Withdrawal: $600, Date: 5nd August 5,2005, Time: 2:00PM
Row 4: Withdrawal: $900, Date: 6th August 6,2005, Time: 9:00PM
Row 5: Deposit: $900, Date: 18th August 18,2005, Time: 7:00AM
Row 6 :Deposit: $800, Date: 23rd August 23,2005, Time: 1:00AM

Figure 5-20 Transaction fact table

Periodic fact table
A periodic fact table stores one row for a group of transactions made over a
period of time.

An example of a periodic-based fact table is shown in Figure 5-21 on page 127
where the fact table records a single row per month for all transactions against an
individual account.

Assume that a customer named Sherpa makes the following transactions in the
month of August 2005:
Money Withdrawn: $400, Date: August 2, 2005, Time: 4:00AM

In the transaction-based fact table, we stored six rows for the six transactions
shown above. However, for the periodic-based fact table, we consider a time
period (end of day, week, month, or quarter) for which we need to store the
transactions as a whole.


Figure 5-21 shows a periodic fact table. The grain of the fact table is chosen as
the account balance per customer at the end of each month. Observe that we do
not record each of the transaction dates separately, but instead use the Month
dimension.

Note: It is important to understand that when we design a periodic fact table,
certain dimensions are not defined when compared to a transaction fact table
type. For example, when comparing the transaction fact table and periodic fact
table, we see that certain dimensions such as Transaction_Type, Branch, and
Transaction_Date are not applicable at the Periodic (Monthly) grain.

1 Row for every Customer every Month

Figure 5-21 Periodic fact table

The periodic fact table design consists of the Month table (Monthly grain). The
periodic fact table MONTHLY_ACCOUNT consists of facts, such as
Previous_Balance, Total_Deposit, Total_Withdrawal, and Available_Balance,
which are true at the periodic month-end grain.

Important features of the periodic fact table are:
A single row is inserted for each set of activities over a period of time.
Typically, the date and time dimensions are represented at the higher level of
detail. For example, the date dimension may be represented at the month
level (instead of day) and the time dimension may be represented at the hour
level (instead of seconds or minutes).
The periodic fact table is known to grow comparatively slowly in comparison
to the transaction fact table.

Accumulating fact table
An accumulating fact table stores one row for the entire lifetime of an event. For
example, from the lifetime of a credit card application being sent to the time it is


accepted. Another example could be the lifetime of a job or college application
being sent to the time it is accepted or rejected by the college or the job posting
company.

To understand the concept of an accumulating fact table, consider that a big
recruitment company advertises vacancies in many jobs relating to software,
hardware, networking, apparel, marketing, sales, food, carpentry, plumbing,
housing, house repairs, mechanical, teaching high school, teaching college,
senior management, and working in restaurants. About 100 000 vacancies are
advertised, in all major newspapers every month. The recruitment company
senior management wants to better understand how efficiently their recruitment
staff works in matching potential job candidates with the jobs they seek. The
senior management wants to understand how long it takes for a prospective
candidate to get a job from the time the resume is sent for a particular job
vacancy.

The accumulating fact table is shown in Figure 5-22, where the fact table records
a single row per job vacancy advertised by the recruitment company.

Note: Accumulating fact tables are typically used for short-lived processes
and not constant event-based processes, such as bank transactions.

Figure 5-22 Accumulating fact table


It is important to understand that there are several dates involved in the entire job
application process. The dates are defined in sequential order in Table 5-3 on
page 130. Also after each subsequent date, certain facts become available. This
is also shown graphically in Figure 5-23.

Note: There are several dates associated with accumulating fact tables. Each
of these dates may be implemented with a single date table using views. The
date table, when implemented using views for different dates, is said to have
been involved in Role-playing. We will discuss Role-playing in more detail in
5.3.9, “Role-playing dimensions” on page 179.

Facts Available at Various Dates:
[Quantity of Full-time Vacancies for each Job]
[Quantity of Part-time Vacancies for each Job]
[Quantity of Received Applications (Applied for Full-Time Jobs)]
[Quantity of Received Applications (Applied for Part-Time Jobs)]

[Quantity of Applications Acknowledged (Full-Time Jobs)]
[Quantity of Applications Acknowledged (Part-Time Jobs)]

[Quantity of Accepted Applications (Full-Time Jobs)]
[Quantity of Accepted Applications (Part-Time Jobs)]
[Quantity of Rejected Applications (Full-Time Jobs)]
[Quantity of Rejected Applications (Part-Time Jobs)]

[Quantity of Interviews conducted for Full-Time Jobs]
[Quantity of Interviews conducted for Part-Time Jobs]
[Quantity of Interviews cancelled for Full-Time Jobs]
[Quantity of Interviews cancelled for Full-Time Jobs]

[Quantity of selected candidates for Full-Time Jobs]
[Quantity of selected candidates for Part-Time Jobs]
[Quantity of rejected candidates for Full-Time Jobs]
[Quantity of rejected candidates for Part-Time Jobs]

Figure 5-23 Facts associated with each date milestone in accounts fact table

Table 5-3 on page 130 shows the various dates associated with the accumulating
fact table and the various facts that are made available on each date.


Table 5-3 Activities defined to understand the concept of accumulating facts
Dates Available facts at each corresponding date

Job advertisement The recruitment company advertises several jobs for several
date companies.
Important facts available during this date are:
Number of full-time vacancies for each job
Number of part-time vacancies for each job

Applications received The application applied date is the date at which different
date candidates send the applications for the job vacancies.
Number of received applications (for full-time jobs)
Number of received applications (for part-time jobs)

Applications The application acknowledgement date is the date when each
acknowledgement candidate who sent a job application is acknowledged (by
date e-mail/phone) that their application has been received and is
being processed.
Number of applications acknowledged (for full-time jobs)
Number of applications acknowledged (for part-time jobs)

Applications The application validating date is the date on which the
validating date applications are validated as matching the required
prerequisite requirements for each job. Any application that
does not meet prerequisite job requirements is rejected.
Number of accepted applications (for full-time jobs)
Number of accepted applications (for part-time jobs)
Number of rejected applications (for full-time jobs)
Number of rejected applications (for part-time jobs)

Interview conducting The interview date is the date on which candidates are
date interviewed for the job vacancies.
Number of interviews conducted for full-time jobs
Number of interviews conducted for part-time jobs
Number of cancelled Interviews for full-time jobs
Number of cancelled interviews for part-time jobs


Dates Available facts at each corresponding date

Final results The final results announcement date is the date on which
announcement date announcement of the selected candidates is made.
Number of selected candidates for full-time jobs
Number of selected candidates for part-time jobs
Number of rejected candidates for full-time jobs
Number of rejected candidates for part-time jobs

Comparison between fact tables
Table 5-4 shows a comparison between the various types of fact tables.

Table 5-4 Comparison of fact table types
Feature Transaction Periodic type fact Accumulating type fact
type fact table table table

Grain One row per One row per period. One row for the entire
definition of transaction. For For example, one lifetime of an event. For
the fact table example one row row per month for a example, the lifetime of a
per line item of a single product sold credit card application
grocery bill. in a grocery store. being sent to the time it is
accepted.

Dimensions Involves date Involves date This type of fact table
dimension at the dimension at the involves multiple date
lowest end-of-period dimensions to show the
granularity. granularity.This achievement of different
could be end of day, milestones.
end-of week, end-of
month, or end-of
quarter.



Total number More than Less than Highest number of
of periodic fact transaction fact dimensions when
dimensions type. type. compared to other fact
involved table types. Generally this
type of fact table is
associated with several
date dimension tables
which are based on a
single date dimension
implemented using a
concept of role-playing.
This is discussed in 5.3.9,
“Role-playing dimensions”
on page 179.

Conformed Uses shared Uses shared Uses shared conformed
Dimensions conformed conformed dimensions.
dimensions. dimensions.

Facts Facts are related Facts are related to Facts are related to
to transaction periodic activities. activities which have a
activities. For example, definite lifetime. For
inventory amount at example, the lifetime of a
end of day or week. college application being
sent to the time it is
accepted by the college.

Facts conformed facts. conformed facts. dimensions.

Database Transaction- The size of a Accumulating fact tables
size based fact tables Periodic fact table is are the smallest in size
have the biggest smaller than the when compared to the
size. If the grain Transaction fact Transaction and Periodic
of the table because the fact tables.
transaction is grain of the date and
chosen at the time dimension is
most detailed significantly higher
level, these than lower level date
tables tend to and time dimensions
grow very fast. present in the
transaction fact
table.



Performance Performance is Performance for Performance is typically
typically good. Periodic fact table is good. The select
However, the higher than other statements often require
performance fact table types differences between two
improves if you because data is dates to see the time
chose a grain stored at lesser period in
above the most detailed grain and days/weeks/months
detailed therefore this table between any two or more
because the has fewer rows. activities.
number of rows
decreases.

Insert Yes Yes Yes

Update No No Yes. Only when a
milestone is reached for a
particular activity.

Delete No No No

Fact table Very fast. Slow in comparison Slow in comparison to the
growth to transaction- transaction and periodic
based fact table. fact table.

Need for High need (This None or very few Medium need (This is
aggregate is primarily (This is primarily primarily because the data
tables because the because the data is is stored mostly at the day
data is stored at already stored at a level. However, the data in
a very detailed highly aggregated accumulating fact tables is
level.) level.) less than the transaction
level.)

5.3 Identifying the model dimensions
In this section, we discuss issues relating to dimensions. We discussed the
Dimensional Model Design Life Cycle in detail in Chapter 6, “Dimensional Model
Design Life Cycle” on page 227. Here we focus on the Identify dimensions phase
of the DMDL, as shown in Figure 5-24 on page 134.


Determine All
Dimensions
Processes

Identify Conformed



Granularity
Accumulating)
Select Requirements Identify Fast Changing Year-to-date Facts
(Source Driven

Iterate
Metadata Management


5.3.1 Degenerate dimensions
Before we discuss degenerate dimensions in detail, it is important to understand
the following:

A fact table may consist of the following data:
Foreign keys to dimension tables
Facts which may be:
– Additive
– Semi-additive
– Non-additive
– Pseudo facts (such as 1 and 0 in case of attendance tracking)
– Textual fact (rarely the case)
– Derived facts
– year-to-date facts
Degenerate dimensions (one or more)


What is a degenerate dimension?
A degenerate dimension sounds a bit strange, but it is a dimension without
attributes. It is a transaction-based number which resides in the fact table. There
may be more than one degenerate dimension inside a fact table.

How to identify a degenerate dimension
All OLTP source systems typically consist of transaction numbers, such as bill
numbers, courier tracking numbers, order numbers, invoice numbers, application
received acknowledgements, and ticket numbers. These transaction numbers in
the OLTP system generally define the transaction.

Consider the grocery store example from Chapter 6, “Dimensional Model Design
Life Cycle” on page 227. The dimensional design has a transaction number, such
as the Bill Number# shown in Figure 5-25. This Bill Number# represents several
line items shown on the graphical bill. Now assume that we choose the grain to
be an individual line item on a grocery store bill.

A fact table row is a line item on the bill. And a line item is a fact table row. As
shown in Figure 5-25, we have identified the dimensions that describe each line
item. The dimensions identified are date, time, product, employee, store, and
customer. Other dimensions, such as supplier, are not visible on the graphical bill
but are true to the grain.

Bill Number#
Customer Bill To: Carlos (Degenerate
Invoice #PP0403001
Account No.
Dimension)
Date: 08/29/2005 Date
Store Store=S1394 1600 Hours Time

Description Quantity UP DSC Amount Discount

Grain: 1. Eggs 12 $3 $36
1 Line item 2. Dairy Milk 2 $2 $4
on the Bill Unit
3. Chocolate Powder 1 $9 $9 Price
4. Soda Lime 12 $1.5 $18
Product 5. Bread 2 $4 $8 Quantity

Employee
Submitted By: Amit Total Due: $75 Total Amt

Payment must be received by July 23.

Please return a copy of this invoice with your payment.
Thank you.

Figure 5-25 Grocery store bill


The next question typically is, “what do we do with the Bill Number#?”
We certainly cannot discard it, because the grain definition we chose is a single
line item on a bill.

Should we make a separate dimension for the Bill Number#?

To see whether or not we should make a separate dimension, try to analyze the
Bill Number# information for the grocery store example. The Bill Number# is a
transaction number that tells us about our purchase made at the store. If we take
an old Bill Number# 973276 to the store and ask the manager to find out
information relating to Bill Number# 973276, we may get all information relating
to the bill. For example, assume that the manager replies that Bill Number#
973276 was generated on August 11, 2005. The items purchased were apples,
orange, and chocolates. The manager also tells us the quantity, unit price, and
discount for each of the items purchased. He also tells us about the total price. In
short, the Bill Number # tells us about the following information:
Transaction date
Transaction time
Products purchased
Store from which bill was generated
Customer to which the merchandise was sold
Quantity, unit price, and amount for each purchased product

The important point to note is that we have already extracted all information
relating to the Bill Number# into other dimensions, such as date, time, store,
customer, and product. Information relating to quantity, unit price, and amount
charged is inside the fact table. It would be correct to say that all information that
the Bill Number# represents is stored in all other dimensions. Therefore, Bill
Number# dimension has no attributes of its own; and therefore it cannot be made
a separate dimension.

Should we place the Bill Number# inside the fact table?

The Bill Number# should be placed inside the fact table right after the
dimensional foreign keys and right before the numeric facts as shown in
Figure 5-26 on page 137.


Degenerate
Dimension

Figure 5-26 Degenerate dimension

The Bill Number#, in SQL language, serves as a GROUP BY clause for grouping
together all the products purchased in a single transaction or for a single Bill
Number#. Although to some, the Bill Number# looks like a dimension key in the
fact table. But, it is not. This is because all information (such as date of purchase
and products purchased) relating to the Bill Number# has been allocated to
different dimensions.

Note: A degenerate dimension, such as Bill Number#, is there because we
chose the grain to be an individual line item on a grocery store bill. In other
words, the Bill Number# degenerate dimension is there because the grain we
chose represents a single transaction or transaction line item.

How to identify that a degenerate dimension is missing
The best way to identify a missing or a badly designed degenerate dimension is
to review your dimensional design and look for any dimension table that has
equal or nearly the same number of rows as the fact table. In other words, if, for
every row that you insert in the fact table you also have to pre-insert another row
in any other dimension table, then you have missed a degenerate dimension.

Let us review the dimensional model shown in Figure 5-27 on page 138. The
grain for this dimensional model has been chosen to be an individual line item on
a grocery store bill.


Bill_Number# has been
removed to create a
separate dimension.

IF, Number of Rows in a Dimension table = Number of Rows in Fact Table
Then Degenerate Dimension is missing

Figure 5-27 Missing degenerate dimension (Bad Design)

Note: Dimension tables are static when compared to fact tables, which grow
constantly in size as sales are made on a daily basis and the activity is
recorded inside the fact table. In a dimensional design, if the size of a
dimension table remains equal or nearly equal to that of a fact table, then a
degenerate dimension has been missed. In other words, if for every fact table
row to be inserted, correspondingly, an equal or near equal number of rows
have to be inserted into a dimension table, then there is a design flaw because
a degenerate dimension has been wrongly represented into a new dimension
of its own.

In Figure 5-27, we observe that for every purchase a customer makes, an equal
number of rows are inserted into the dimension and fact tables. This is because
the Bill dimension table has the Bill_Number which is different for each bill. The
Bill dimension is not a static dimension like other dimensions.

When will the Bill_Number# no longer be a degenerate dimension?
When certain data columns belong to the Bill_Number# itself and do not fall
into any other dimensions (such as Date, Time, Product, Customer, Supplier,
and Employee), then Bill_Number# would no longer be a degenerate
dimension. An example of this is explained in a case study in “Identify
degenerate dimensions” on page 384. We see that Invoice Number (type of


transaction number) is handled as a separate dimension and not as a
degenerate dimension inside the fact table.

This decision however needs to carefully be made by studying the Bill_Number#.
If you decide to create a new separate dimension for the Bill_Number#, then you
must make sure that it will not contain equal or near equal rows with the fact
table. If it does, then there has probably been a mistake.

Note: OLTP transaction numbers, such as bill numbers, courier tracking #,
order number, invoice number, application received acknowledgement, and
ticket number, usually produce dimensions without any attributes and are
represented as degenerate dimensions in the fact table.

5.3.2 Handling time as a dimension or a fact
A date dimension typically represents day, week, month, quarter, or year, where
a time dimension represents hours, minutes, and seconds within a day. These
are two physically separate dimension structures within the data warehouse,
each with their own surrogate keys used to join the fact tables. Because of the
unique characteristics of these two dimensions, special considerations can be
made when building them. It is not advised to merge the date and time dimension
into one table because a simple date dimension table which has 365 rows (for 1
year) would explode into 365 (Days) x 24 (Hours) x 60 (Minutes) x 60 (Seconds)
= 31 536 000 rows if we tried storing hours, minutes, and seconds. This is for just
1 year. If we had 10 years of data in our date table, we would have 10 x 365 =
3 650 rows (assuming no leap year). If we now tried to store data at the second,
minute, and hour level, we would expand the simple date table from 3 650 rows
(for 10 years) to about 3 650 x 24 x 60 x 60 = 315 360 000 rows.

Having now made a point to handle the date and time dimensions separately, we
can discuss how to implement the time dimension.

Time in dimensional modeling
Unlike the date dimension, the time-of-day (hour, minute, and second) may be
well expressed as a simple numerical fact rather than as a separate time
dimension, unless there are textual descriptions of certain periods within the day
that are meaningful, such as early morning, late morning, noon, lunch hour,
afternoon, evening, night, and late night shift.
Time of day as a dimension:
Time is expressed as a dimension when the business needs to understand
the sales of its product over a time period which has meaningful textual
names such as:


– Early morning (6 a.m. to 8 a.m.)
– Late morning hours (8 a.m. to 11 a.m.)
– Rush hour (11 a.m. to 1 p.m.)
– Lunch hour (1 p.m. to 2 p.m.)
For expressing such scenarios that describe the time, we handle time as a
dimension. This is shown in Figure 5-28. The grain of this time dimension is
an hour.

Figure 5-28 Handling time as a dimension

In other words, handling time as a dimension is the correct approach if we
need to support the roll-up of time periods into more summarized groupings
for reporting and analysis, such as 30-minute intervals, hours, or a.m./p.m.
(see Table 5-5 which shows sample time dimension rows) or for
business-specific time groupings, such as the week day morning rush period,
week day early mornings, late night shifts, and late evenings.

Table 5-5 Time dimension with sample rows
Timeid Standard time Time of day description Time of day a.m.
p.m. indicator

1 0100 hours Late night shift a.m.

2 0800 hours Morning a.m.

3 1100 hours Rush hour a.m.

4 2000 hours Night p.m.

The other reason you may express time as a dimension is when you want to
represent different hierarchies for the time that you are measuring. Some of
the time hierarchies the business may want are listed as follows:
– Standard Time hierarchy (Hour → Minute → Second)
– Military Time hierarchy (Hour → Minute → Second)


We have seen in Figure 6-21 on page 281 that date dimension can have
multiple hierarchies, such as fiscal and calendar hierarchy inside the same
dimension table.
Time of day as a fact
We express time as a fact inside the fact table if there is no need to roll up or
filter on time-of-day groups, such as morning hour, rush hour, and late night
shift. Then in this scenario we have the option of treating time as a simple
numeric fact which is stored in the fact table. Here the time of day is
expressed as a date-time stamp such as “09-25-2005 10:10:36”.
Figure 5-29 shows time (TIME_OF_SELLING) expressed as a fact inside the
fact table. The TIME_OF_SELLING stores the time as a date-time stamp
which includes the precise time in hours, minutes, and seconds at which the
product sells inside the store.

Time expressed as a
FACT (Date-Time Stamp)
[09-25-2005 10:10:36]

Figure 5-29 Time expressed as a fact

It is important to understand that it is very difficult to generate reports which
require roll-ups for business-specific time groupings, such as the week day
morning rush period, week day early mornings, late night shifts, and late
evenings, using the TIME_OF_SELLING (see Figure 5-29) which is stored as
fact.
If we had stored time as a dimension, then we would be able to generate
reports to support the roll-up of time periods into more summarized groupings
for reporting and analysis, such as 30-minute intervals, hours, or a.m./p.m.


(see Table 5-5 on page 140, which shows sample time dimension rows) or for
business-specific time groupings, such as the week day morning rush period,
week day early mornings, late night shifts, and late evenings.

5.3.3 Handling date and time across international time zones
In businesses that span multiple time zones, both the date and time of day may
need to be expressed both in local time and Greenwich Mean Time (GMT). This
must be done with separate date dimensions for local and GMT, and separate
time-of-day facts as well. This is shown in Figure 5-30.

Local and GMT
Dates
Local and GMT
Time Facts

Figure 5-30 Handling International Time Zones

The separate date dimensions are implemented using views, or a concept called
role-playing. We discuss role-playing in 5.3.9, “Role-playing dimensions” on
page 179.

5.3.4 Handling dimension hierarchies
A hierarchy is a cascaded series of many-to-one relationships and consists of
different levels. Each level in an hierarchy corresponds to a dimension attribute.
Hierarchies document the relationships between levels in a dimension. For
example, a region hierarchy is defined with the levels Region, State, and City.


In other words, a hierarchy is a specification of levels that represents a
relationship between different attributes within a hierarchy. Figure 5-31 shows
sample hierarchies.

Product Date Location

Department Year Level 1
Country

Category Quarter State Level 2

Brand Month Region Level 3

Product Name Level 4

Figure 5-31 Dimension hierarchies

Hierarchies enrich the semantics of data in a dimensional model and
correspondingly improve the class of interesting and meaningful queries that can
be run against it.

Note: In order to generate data for a report, we can drill down or up on
attributes from more than one explicit hierarchy and with attributes that are
part of no hierarchy.

We now discuss implementing three types of hierarchies:
Balanced hierarchy
Unbalanced hierarchy
Ragged hierarchy

Balanced hierarchy
A balanced hierarchy is one in which all of the dimension branches have the
same number of levels. In other words, the branches have a consistent depth.

The logical parent of a level is directly above it. A balanced hierarchy can
represent a date where the meaning and depth of each level, such as Year,
Quarter, and Month, are consistent. They are consistent because each level
represents the same type of information, and each level is logically equivalent.


Figure 5-32 shows an example of a balanced time hierarchy.

Figure 5-32 Balanced hierarchy

How to implement a balanced hierarchy
A balanced hierarchy consists of a fixed number of levels. Assume that we are
designing a date dimension consisting of the hierarchy shown in Figure 5-32. We
design the date dimension as shown Figure 5-33. You should have as many fixed
attributes inside the dimension table as the number of levels in the hierarchy. This
is the number of roll-up levels that you want to track.

Surrogate Key

3 Columns inside date
Dimension for Balanced
Hierarchy

Figure 5-33 Balanced hierarchy for the Date dimension

The date dimension can also have multiple balanced hierarchies in the same
dimension table, as shown in Figure 5-34 on page 145.


Surrogate Key

Calendar Hierarchy

Fiscal Hierarchy

Multiple Balanced Hierarchies in one Dimension

Figure 5-34 Multiple balanced hierarchies in date dimension

A hierarchy is unbalanced if it has dimension branches containing varying
numbers of levels. Parent-child dimensions support unbalanced hierarchies.

An unbalanced hierarchy has levels that have a consistent parent-child
relationship, but have a logically inconsistent level. The hierarchy branches also
can have inconsistent depths. An unbalanced hierarchy can represent an
organization chart. For example, Figure 5-35 on page 146 shows a chief
executive officer (CEO) on the top level of the hierarchy and two people that
branch off below. Those two are the COO and the executive secretary. The COO
hierarchy has additional people, but the executive secretary does not. The
parent-child relationships on both branches of the hierarchy are consistent.
However, the levels of both branches are not logical equivalents. For example,
an executive secretary is not the logical equivalent of a chief operating officer.

We can report by any level of the date hierarchy and see the results across year,
quarter, or month.


Figure 5-35 Unbalanced hierarchy

How to implement an unbalanced hierarchy
It is important to understand that representing an unbalanced hierarchy is an
inherently difficult task in a relational environment. OLAP Cube-based systems
are better suited for reporting when the data has unbalanced hierarchies.
However, if your hierarchies are symmetrical then arguably either type of
technology (OLAP cubes or relational database) is equally capable of providing
the answer.

A common example of an unbalanced hierarchy is an organization chart.
Employees at a given level may have hundreds of employees, while others at the
same level may have none or a few. So, as an example, if you wish to create a
report that computes the sales totals for all employees at a given level, you can
create this report more efficiently with an OLAP reporting system. For an
example of an unbalanced hierarchy, see Figure 5-36 on page 147.

Assume, in this hierarchy, that we want to report the sales for a set of managers.
Each box in the figure represents an employee in the organization, and there are
are four levels. A large organization with thousands of employee can have many
levels. It is important to understand that each employee can play the role of a
parent or a child. The report may require summarizing the sales made by one
employee or summarizing all sales for one manager (and all people in that
management chain). This is equivalent to summarizing the sales revenue to any
node in the organizational tree shown in Figure 5-36 on page 147.


Employee 1 Level 1

Employee 2 Employee 3 Level 2

Employee 4 Employee 5 Employee 6 Level 3

Employee 7 Level 4

Figure 5-36 Unbalanced hierarchies for an organization chart

The general approach for a parent-child relationship in a relational database is to
introduce a child key in the parent table. This is called a recursive pointer, and is
shown in Figure 5-37.

Recursive Pointer
(MANAGER_ID)
to show the
Manager (Parent)
of each employee

Figure 5-37 Recursive pointer to show parent-child relationship

Issues using a recursive pointer for unbalanced hierarchy
The recursive pointer approach (shown in Figure 5-37) does implement the
unbalanced hierarchy (shown in Figure 5-36). However, this approach does not
work well at all with SQL language. Why?

Assume that we want a report which shows the total sales revenue at node
Employee 2. What this means is that we want the report showing the total sales
made by Employee 2, Employee 4, Employee 5, and Employee 7.


We cannot use SQL to answer this question because the GROUP BY function in
SQL cannot be used to follow the recursive tree structure downward to
summarize an additive fact such as sales revenue in the SALES_FACT table.
Because of this, we cannot connect a recursive dimension to any fact table.

The next question is, how do we solve the query using an SQL Group by clause?
And, how will we be able to summarize the sales revenue at any Employee node
(from Figure 5-36 on page 147).

The answer is to use a bridge table between the EMPLOYEE and SALES_FACT
table as shown in Figure 5-38. The bridge table is called a
BRIDGE_Traverse_Hierarchy. The aim of the bridge table is to help us traverse
through the unbalanced hierarchy depicted in Figure 5-36 on page 147. The
bridge table helps us to analyze the sales revenue (SALES_REVENUE in
Figure 5-37 on page 147) at any hierarchical level using the SQL language,
which was not possible with a
simple recursive pointer as shown in Figure 5-37 on page 147.

Note: The beauty of using a bridge table to solve the problem of traversing an
unbalanced hierarchy is that neither the Employee dimension nor the
Sales_FACT table changes in any way. The use of a bridge table is optional,
and may be used only when you need to traverse through an unbalanced
hierarchy as shown in Figure 5-36.

Figure 5-38 Use of a bridge table for descending the unbalanced hierarchy

What does the bridge table contain?
The Employee dimension table has one row for each employee entity at any
level of the hierarchy.
The bridge table contains a single row for every path as shown in Figure 5-36
on page 147. In other words, the BRIDGE_Traverse_Hierarchy table contains


one row for each pathway from an employee entity to each immediate
employee beneath it, as well as a row for the zero-length pathway from an
employee to itself.

Note: A zero-length pathway in an unbalanced hierarchy is a pathway to itself.
For example, in Figure 5-36, for the Employee 1, the zero-length pathway
would signify the distance between Employee 1 and itself.

The BRIDGE_Traverse_Hierarchy table also contains the following other
information:
– NUMBER_OF_LEVELS column: This is the number of levels between the
parent and child.
– BOTTOM_FLAG: This flag tells us that the Employee node is the bottom
most and does not have any employees working under them.
– TOP_FLAG: This flag tells us that the employee is at the top of the
organizational hierarchy and does not have any employees working above
them.

Table 5-6 shows sample rows for the unbalanced hierarchy in Figure 5-38 on
page 148.

Table 5-6 Sample bridge table rows
PARENT_ CHILD_EMPLOYEE NUMBER_OF BOTTOM_ TOP_FLAG
EMPLOYEE _KEY _LEVELS FLAG
_KEY

1 1 0 No Yes

1 2 1 No No

1 3 1 No No

1 4 2 Yes No

1 5 2 No No

1 6 2 Yes No

1 7 3 Yes No

2 2 0 No No

2 4 1 Yes No

2 5 1 No No

2 7 2 Yes No


PARENT_ CHILD_EMPLOYEE NUMBER_OF BOTTOM_ TOP_FLAG
EMPLOYEE _KEY _LEVELS FLAG
_KEY

3 3 0 No No

3 6 1 Yes No

4 4 0 Yes No

5 5 0 No No

5 7 1 Yes No

6 6 0 Yes No

7 7 0 Yes No

How many rows does the bridge table contain?
To get the answer to this question, we need to analyze the unbalanced hierarchy
shown in Figure 5-36 on page 147. The total number of rows in the bridge table
are equal to the total number of paths available inside the unbalanced hierarchy.
The following are the paths:
1. Employee 1 to Employee 1 (Zero-length pathway to itself)
2. Employee 1 to Employee 2
8. Employee 2 to Employee 2 (Zero-length pathway to itself)
10.Employee 2 to Employee 5
12.Employee 3 to Employee 3 (Zero-length pathway to itself)
18.Employee 7to Employee 7 (Zero-length pathway to itself)

Therefore, the total number of rows in the bridge table named
(BRIDGE_Traverse_Hierarchy) is 18. This can be verified from Table 5-6 on
page 149.


Now suppose that we want to find the total sales revenue during January 2005
for Employee 2 (Figure 5-36 on page 147) and all its junior employees (Employee
4, Employee 5, and Employee 7).

In order to descend the hierarchy from Employee 2, we make use of Figure 5-38
on page 148. Example 5-1 shows sample SQL necessary to generate the report:

Example 5-1 Sample SQL to generate the report for descending the hierarchy
Select E.Full_Name, SUM(F.Sales_Revenue)
From
Employee E, Bridge_Traverse_Hierarchy B, Sales_Fact F, Date D
where
E.EmployeeKey= B.Parent_Employee_Key
and
B.Child_Employee_Key= F.EmployeeKey
and
F.DateID= D.DateID
and
E.[FULL NAME]= ‘Employee 2’
// and D.Date= January 2005 Data (Date SQL Code logic depends on Database)
GROUP BY E.Full_Name

Now suppose that we want to find out the total sales revenue during January
2005 for Employee 6 (Figure 5-36 on page 147) and all its senior employees
(Employee 3 and Employee 1). In order to ascend the hierarchy for Employee 6,
we make use of a bridge table, as depicted in Figure 5-39.

Figure 5-39 Use of the bridge table to ascend the unbalanced hierarchy

Example 5-2 shows sample SQL necessary to generate the report.


Example 5-2 Sample SQL to generate the report for ascending the hierarchy
Select E.Full_Name, SUM(F.Sales_Revenue)
From
Employee E, Bridge_Traverse_Hierarchy B, Sales_Fact F, Date D
where
E.EmployeeKey= B.Child_Employee_Key
and
B.Parent_Employee_Key= F.EmployeeKey
and
F.DateID= D.DateID
and
E.[FULL NAME]= ‘Employee 6’
GROUP BY E.Full_Name

Joining a bridge table
The bridge table can be joined between the Employee dimension table and fact
table in two ways. These are shown in Figure 5-40 on page 153. The two ways
are explained as follows:
For descending the hierarchy
The fact table joins to the child employee key of the bridge table and the
employee dimension table joins to the parent employee key of the bridge
table. Descending means selecting an employee and moving down the
hierarchy. We can use the NUMBER_OF_LEVELS column to determine how
deep to go. If the BOTTOM_FLAG=’Yes’, we know that we have reached the
bottom most part of the unbalanced hierarchy.
For ascending the hierarchy
The fact table joins to the parent employee key of the bridge table and the
employee dimension table joins to the child employee key of the bridge table.
The NUMBER_OF_LEVELS column helps us go up the number of levels
desired. The TOP_FLAG helps us determine if we have reached the top of the
hierarchy at the CEO level.


Fact Bridge Employee
parent_id Employee_id
Employee_id child_id Employee 1

Employee 2 Employee 3

(a) Descending the Hierarchy (Moving Down )
Employee 4 Employee 5 Employee 6
Fact Bridge Employee
parent_id Employee_id
Employee_id child_id Employee 7

(c) Organization Hierarchy

(b) Ascending the Hierarchy (Moving Up )

Figure 5-40 Ways of joining a bridge table to fact and dimension tables

Summary: Using a bridge table to navigate an unbalanced hierarchy
There are two join directions that you may use:
Navigate up the hierarchy
– Fact joins to subsidiary customer key
– Dimension joins to parent customer key
Navigate down the hierarchy
– Fact joins to parent customer key
– Dimension joins to subsidiary customer key

Disadvantages of the bridge table approach
The bridge table data is complex to maintain.
The bridge table design is too complex to be used directly by the users.

In what scenarios is the bridge table approach used?
A bridge table is used in scenarios where there is a business need to report
against an unbalanced organizational hierarchy such as the one discussed in
Figure 5-36 on page 147 and Figure 5-40 (c). This approach is typically used in
electronic and manufacturing-related industries dealing with a huge number of
parts that belong to a higher level assembly.


Ragged hierarchy
A ragged dimension contains at least one member whose parent belongs to a
hierarchy that is more than one level above the child (see Figure 5-41). Ragged
dimensions, therefore, contain branches with varying depths.

A ragged hierarchy is one in which each level has a consistent meaning, but the
branches have inconsistent depths because at least one member attribute in a
branch level is unpopulated.

A ragged hierarchy can represent a geographic hierarchy in which the meaning
of each level, such as city or country, is used consistently, but the depth of the
hierarchy varies. Figure 5-41 shows a geographic hierarchy that has Continent,
Country, Province/State, and City levels defined. One branch has North America
as the Continent, United States as the Country, California as the Province or
State, and San Francisco as the City.

However, the hierarchy becomes ragged when one member does not have an
entry at all of the levels. For example, another branch has Europe as the
Continent, Greece as the Country, and Athens as the City, but has no entry for
the Province or State level because this level is not applicable in this example. In
this example, the Greece and United States branches descend to different
depths, creating a ragged hierarchy.

No
State

Figure 5-41 Ragged hierarchy


How to implement a ragged hierarchy in dimensions
A ragged hierarchy as shown in Figure 5-41 on page 154 can be implemented in
a dimension table as shown in Figure 5-42.

Location
Hierarchy

Figure 5-42 Ragged hierarchy implementation in a dimension table

The hierarchy is shown as Continent → Country → State → City. It is possible
that attributes such as State will not be populated for Countries, such as Greece,
that have no states. In this scenario, we can populate a dummy value of “Not
Applicable” or “No States”, rather than null.

5.3.5 Slowly changing dimensions
Handling changes to dimensional data across time can be difficult, and
dimensional attributes rarely remain static. A customer address can change,
sales representatives come and go, and companies introduce new products to
replace older ones.

Changing data in dimension tables can present far-ranging implications when
you view the changes over time. For example, assume a company sells
car-related accessories.The company decides to reassign a sales territory to a
new sales representative. How can you record the change without making it
appear that the new sales representative has always held that territory? If a
customer name or married status changes, how can you record the change and
preserve the old and new version of the name and married status? Designing a
dimensional model that accurately and efficiently handles changes is a critical
consideration when building a data warehouse. After all, the main reason for
building a warehouse is to preserve history.

What are slowly changing dimensions?
A slowly changing dimension is a dimension whose attribute or attributes for a
record (row) change slowly over time.


In a dimensional model, the dimension table attributes are not fixed. They
typically change slowly over a period of time, but can also change rapidly. The
dimensional modeling design team must involve the business users to help them
determine a change handling strategy to capture the changed dimensional
attributes. This describes what to do when a dimensional attribute changes in the
source system. A change handling strategy involves using a surrogate
(substitute) key as its primary key for the dimension table.

We now present a few examples of slowly changing dimension attribute
change-handling strategies using the star schema shown in Figure 5-43. We also
discuss the advantages and disadvantages of each approach.

Figure 5-43 Star schema for discussing slowly changing dimensions

The following are approaches for handling slowly changing dimensions:

Type-1 (overwriting the history)
A Type-1 approach overwrites the existing dimensional attribute with new data,
and therefore no history is preserved. Now consider the star schema shown in
Figure 5-43. Assume that the correct first name of a sales representative is
Monika, but is mistakenly written in the table as Monnica, as shown in Table 5-7.

Table 5-7 Sales_representative table
SALES_REPKEY LASTNAME FIRSTNAME SALESREGION

963276 Agarwal Monnica New York


When we use a Type-1 approach, we overwrite the old (incorrect) value
(Monnica) shown in Table 5-7 on page 156, with the new correct value of Monika,
as shown in Table 5-8.

Table 5-8 Sales_representative table (corrected data)

963276 Agarwal Monika New York

Updating the FIRSTNAME field once and seeing the change across all fact rows
is effective and makes good business sense if the update corrects a misspelling,
such as the first name, for example. But suppose the sales representative now
correctly named Monika is moved to a new sales region in California. Updating
the sales representative dimensional row would update all previous facts so that
the sales representative would appear to have always worked in the new sales
region, which certainly was not the case. If you want to preserve an accurate
history of who was managing which sales region, a Type-1 change might work
for the Sales FIRSTNAME field correction, but not for the “SALESREGION” field.

Note: It is very important to involve the business users when implementing the
Type-1 approach. Although it is the easiest approach to implement, it may not
be appropriate for all slowly changing dimensions since it can disable their
tracking business history in certain situations.

The Type-1 approach is summarized in Table 5-9.

Table 5-9 Review of Type-1 approach
Type-1 approach Description

When to use the Type-1 This may be the best approach to use if the
change handling approach attribute change is simple, such as a correction in
spelling. And, if the old value was wrong, it may
not be critical that history is not maintained.
It is also appropriate if the business does not
need to track changes for specific attributes of a
particular dimension.

Advantages of the Type-1 It is the easiest and most simple to implement.
change handling approach It is extremely effective in those situations
requiring the correction of bad data.
No change is needed to the structure of the
dimension table.



Disadvantages of the All history may be lost if this approach is used
Type-1 change handling inappropriately. It is typically not possible to trace
approach history.
All previously made aggregated tables need to be
rebuilt.

Impact on existing No impact. The table structure does not change.
dimension table structure

Impact on preexisting Any preexisting aggregations based on the old
aggregations attribute value will need to be rebuilt. For
example, as shown in Table 5-8, when correcting
the spelling of the FIRSTNAME of Monika
(correct name), any preexisting aggregations
based on the incorrect value Monnica will need to
be rebuilt.

Impact on database size No impact on database size.

Type-2 (preserving the history)
A Type-2 approach adds a new dimension row for the changed attribute, and
therefore preserves history. This approach accurately partitions history across
time more efficiently than other change handling approaches. However, because
the Type-2 approach adds new records for every attribute change, it can
significantly increase the database's size.

Now consider the star schema shown in Figure 5-43 on page 156. Assume that
the sales representative named Monika works for a sales region in California, as
shown in Table 5-10. After having worked for fifteen years in California, the region
is changed to New Jersey.

Table 5-10 Sales representative table data

963276 Agarwal Monika California

With a Type-2 approach, you do not need to make structural changes to the
Sales_representative dimension table. But you insert a new row to reflect the
region change. After implementing the Type-2 approach, two records appear as
depicted in Table 5-11 on page 159. Each record is related to appropriate facts in
the fact table shown in Figure 5-43 on page 156, which are related to specific
points in time in the time dimension and specific dates of the date dimension.


Table 5-11 Type-2 approach for the sales_representative dimension

963276 Agarwal Monika California

963277 Agarwal Monika New York

This example illustrates two supplementary but important dimensional design
concepts that are the base of a strong dimensional model.

First, the primary key for the Sales_Representative dimension table is a
surrogate key. We highly recommend that you avoid designating a primary key
that has business value, such as the EMPLOYEE_ID (which is a natural id in the
source system). Generally when we design an OLTP database, we use a field
that has meaning as a primary key. However, when designing the primary keys
for our dimension tables, we want to use non-intelligent keys as primary keys for
our dimensions.

This design concept becomes even more important if we are implementing a
change handling technique using the Type-2 approach. To prove this, assume
that we use the EMPLOYEE_ID (which is the natural id in the source system) as
the primary key of the dimension. A Type-2 implementation approach needs to
accommodate multiple rows per sales representative depending upon the
number of changes that need to be preserved for history purposes.

In the example, the EMPLOYEE_ID never changes. The person is still the same
person and has the same employee identification (EMPLOYEE_ID), but his or
her information changes over time. Attempting to use the EMPLOYEE_ID field
would cause a primary key integrity violation when you try to insert the second
row of data for that same person.

The second dimensional model design concept that this example illustrates is the
lack of use of effective date attributes (in Sales_Representative dimension) that
identify exactly when the change in the Sales_Representative dimension
occurred.

A word on effective and expiration date attributes
There is no need to include effective date attributes inside the
Sales_Representative dimension because the Date dimension shown in
Figure 5-43 on page 156 will effectively partition the data over time when it is
related to the fact table. Of course, the effectiveness of this partitioning depends
on the grain of the fact table. In our case the grain tracks data at day level using
the date dimension and at hour level using the time dimension.


Effective and expiration date attributes are necessary in the staging area
because we need to know which surrogate key is valid when loading historical
fact records. This is often a point of confusion in the design and use of type-2
slowly changing dimensions.

However, effective and expiration date attributes are very important in the
staging area of the data warehouse or dimensional model because we need to
know which surrogate key is valid when loading historical fact rows. In the
dimension table, we stated that the effective and expiration dates are not
needed, though you may still use them as helpful extras that are not required for
the basic partitioning of history. It is important to note that if you add the effective
and expiration date attributes in the dimension table, then there is no need to
constrain on the effective and expiration date in the dimension table to get the
correct answer. You can get the same answer by partitioning history using the
date or time dimension of your dimensional designs.

A summary of the type-2 change handling strategy is presented in Table 5-12.

Table 5-12 Summary of type-2 approach concepts

When to use the Type-2 When there is need to track an unknown number
change handling approach of historical changes to dimensional attributes.

Advantages of the Type-2 Enables tracking of all historical information
change handling approach accurately and for an infinite number of changes.

Disadvantages of the Causes the size of the dimension table to grow
Type-2 change handling fast. In cases where the number of rows being
approach inserted is very high, then storage and
performance of the dimensional model may be
affected.
Complicates the ETL process needed to load the
dimensional model. ETL-related activities that are
required in the type-2 approach include
maintenance of effective and expiration date
attributes in the staging area.

Impact on existing No change to dimensional structure needed.
dimension table structure Additional columns for effective and expiration
dates are not needed in the dimension table.

Impact on preexisting There is no impact on the preaggregated tables.
aggregations The aggregated tables are not required to be
rebuilt as with the type-1 approach.



Impact on database size Yes, accelerates the dimensional table growth
because with each change in a dimensional
attribute, a new row is inserted into the dimension
table.

Adding effective and No. This is not necessary in the dimension tables.
expiration dates to
dimension tables However, effective and expiration attributes are
needed in the staging area because we need to know
which surrogate key is valid when we are loading
historical fact rows. In the dimension table, we stated
that the effective and expiration dates are not needed
though you may still use them as helpful extras that
are not required for the basic partitioning of history. It
is important to note that in case you add the effective
and expiration date attributes in the dimension table,
then there is no need to constrain on the effective and
expiration date in the dimension table in order to get
the right answer. You could get the same answer by
partitioning history using the date or time dimension of
your dimensional designs.

Type-3 (preserving one or more versions of history)
The type-3 approach is typically used only if there is a limited need to preserve
and accurately describe history. An example is when someone gets married and
there is a need to retain the original surname of the person.

Consider the last name of the person Suzan Holcomb prior to marriage, as
shown in Table 5-13. The last name for Suzan is Holcomb.

Table 5-13 Customer table
Customer ID First name Old last name New last name Married?

963276 Suzan Holcomb Holcomb No

After Suzan gets married to David Williams, she acquires his last name. For such
changes where we can predict the number of changes that will happen for a
particular dimension attribute (such as last name), we use a type-3 change. This
is shown in Table 5-14.

Table 5-14 Type-3 change for customer table

963276 Suzan Holcomb Williams Yes


In Table 5-14 on page 161, you can preserve one change per attribute. In other
words, you can preserve the last name of Suzan only if she marries once. If she
marries again, to a second person called Brent Donald, then the type-3 change
discards her first last name Holcomb, as shown Table 5-15.

Table 5-15 Type-3 change

963276 Suzan Williams Donald Yes

Table 5-15 shows that now the new last name is Donald and the old last name
column has a value of Williams. The disadvantage with the type-3 approach is
that it can preserve only one change per attribute—old and new or first and last.
In case you want to preserve two changed attributes for Suzan, then you need to
redesign the dimension table as shown in Table 5-16.

Table 5-16 Type-3 two changes
Customer ID First name First old Second old New last Married
last name last name name

963276 Suzan Holcomb. Williams Donald Yes

The more historical changes you want to maintain using the type-3 change, the
higher the number of additional columns needed in the dimensional table.

Instead of inserting a new dimensional row to hold the attribute change, the
type-3 approach places a value for the change in the original dimensional record.
You can create multiple fields to hold distinct values for separate points in time.

Important: The type-3 approach enables us to see new and historical fact
table rows by either the new or prior attribute values.

There is another disadvantage of using the type-3 approach. This has to do with
extracting information from a dimension table which has been implemented using
a type-3. Although the dimension structure that was implemented using type-3
has all the data needed, the SQL required to extract the information can be
complex.

Extracting a specific value is not that difficult, but if you want to obtain a value for
a specific point in time, or multiple attributes with separate old and new values,
the SQL statements needed become long and have multiple conditions. Overall,
a type-3 approach can store the data of a change, but cannot accommodate
multiple changes. Also, summary reporting is very difficult using the type-3
approach.


To summarize the type-3 slowly change handling strategy, we present
Table 5-17.

Table 5-17 Summary of type-3 approach concepts

When to use the type-3 Should only be used when it is necessary for the
change handling data warehouse to track historical changes, and
approach? when such changes will only occur for a finite
number of times. If the number of changes can be
predicted, then the dimension table can be
modified to place additional columns to track the
changes.

Advantages of the type-3 Does not increase the size of the table as
change handling approach compared to the type-2 approach, since new
information is updated.
Allows us to keep part of history. This is
equivalent to the number of changes we can
predict. Such prediction helps us modify the
dimension table to accommodate new columns.

Disadvantages of the Does not maintain all history when an attribute is
type-3 change handling changed more often than the number in the
approach predicted range, because the dimension table is
designed to accommodate a finite number of
changes.
If we designed a dimension table assuming a
fixed number of changes, then needed more,
then we would have to redesign or risk losing
history.

Impact on existing The dimension table is modified to add columns.
dimension table structure The number of columns added depends on the
number of changes to be tracked.

Impact on preexisting You may be required to rebuild the
aggregations preaggregated tables.

Impact on database size No impact is there since data is only updated.

5.3.6 Handling fast changing dimensions
In this section we identify the changing dimensions that cannot be handled using
the Type-1, Type-2, or Type-3 approaches.


What are fast changing dimensions?
A dimension is considered to be a fast changing dimension if one or more of its
attributes changes frequently and in many rows. A fast changing dimension can
grow very large if we use the Type-2 approach to track numerous changes. Fast
changing dimensions are also called rapidly changing dimensions. Consider a
scenario for a customer dimension having 100 000 rows. Assume that we are
handling the changes for this customer using the Type-2 approach. Further
assume that in a year an average of 10 changes occur for each customer.
Therefore in one year the number of rows will increase to 100 000 x 10 = 1 000
000. For some companies, this may still be a small number to manage even
using a Type-2 change. That is, for some, the customer dimension may be a
slowly changing dimension.

Assume that a customer table has 10 million rows. Imagine the same scenario
where, on average, 10 changes occur for a customer each year. This means at
the end of the year, the table would grow to about 100 million rows. This is a huge
growth. Such a customer dimension may be considered a fast growing
dimension. Then, handling such a fast growing dimension using a Type-2
approach is not feasible.

In order to identify the reason for a fast changing dimension, we must look for
attributes that have continuously variable values such as age, test score, size,
weight, credit history, customer account status, or income.

An appropriate approach for handling very fast changing dimensions is to break
off the fast changing attributes into one or more separate dimensions, called
mini-dimensions. The fact table would then have two foreign keys—one for the
primary dimension table and another for the fast changing attributes. These
dimension tables would be associated with one another every time we insert a
row in the fact table.

Figure 5-44 on page 165 shows an example of a fast changing dimension.


Fast Changing
Attributes in a Fast
Changing Dimension

Figure 5-44 Fast changing dimension

The attributes which change rapidly in this fast changing dimension are identified
as:
Age
Income
Test score
Rating
Credit history score
Customer account status
Weight

Solving the fast changing dimensions problem

After having identified the constantly changing attributes, the next step is to
convert these identified attributes individually into band ranges. The concept
behind this exercise is to force these attributes to take limited discreet values.
For example, let us assume that each of the above seven attributes takes on 10
different values, then the Customer_Mini_dimension (Figure 5-45 on page 167)
will have 1 million values. Thus, by creating a mini-dimension table consisting of
band-range values, we have solved the problem of the situation where the
attributes such as age, income, test score, credit history score, customer account
status, and weight can no longer change. These attributes cannot change
because they have a fixed set of band-range values (see Table 5-18 on
page 166) rather than having a large number of values.

Sample rows for this new Mini-dimension, called Customer_Mini_Dimension, are
shown in Table 5-18 on page 166.


Table 5-18 Sample mini-dimension rows
Age Income Test Rating Credit_ Customer_ Weight
Score History_ Account_
Score Status

18 to 25 $800 to 2-4 8-12 7-9 Good 88
$1000

26 to 28 $1000 to 5-9 8-12 10-14 Good 88
$1500

29 to 35 $1500 to 10-17 8-12 15-19 Good 88
$3000

More .... .... .... .... .... ....

What is a mini-dimension?

A mini-dimension is a dimension that usually contains fast changing attributes of
a larger dimension table. This is to improve accessibility to data in the fact table.
Rows in mini-dimensions are fewer than rows in large dimension tables because
we try to restrict the rows in mini-dimensions by using the band range value
concept.

Note: It is important to understand that the mini-dimension created in
Figure 5-45 cannot be allowed to grow too large. If it became a fast changing
dimension, then it must be split into another mini-dimension, and the same
band range technique applied to accommodate the growth.

After identifying the fast changing attributes of the primary customer dimension,
and determining the band ranges for these attributes, a new mini-dimension is
formed called Customer_Mini_Dimension as shown in Figure 5-45 on page 167.
This scenario works reasonably well for a few rapidly changing attributes. The
next question is, “What if there are ten or more rapidly changing attributes?”

Should there be a single mini-dimension for all fast changing attributes? Maybe,
but the growth of this mini-dimension can easily get out of hand and grow into a
huge dimension. The best approach is to use two or more mini-dimensions and
force the attributes of each of these dimensions to take on a set of band range
values.


Fast Changing
Attributes Split
Into a new
Mini-Dimension

Figure 5-45 Fast changing dimension implementation

As a summary, the approach used with fast changing dimensions involves the
following activities:
We analyze all dimensions identified carefully to find which change very fast.
The fast changing dimension is analyzed further to see what the impact is on
the size of the dimension table if we handle the change using the Type-2
approach. If the impact on the size of the dimension table is huge, then we
must avoid the Type-2 approach.
The next step is to analyze the fast changing dimension in detail to identify
which attributes of this dimension are subject to changing fast. Such fast
changing attributes are then separated into one or more mini-dimension
tables whose primary keys are attached to the fact table as a foreign key.
Then there are the following tables:
– One main dimension table, which was the primary original fast changing
dimension, without the fast changing attributes.
– One or more mini-dimensions where each table consists of fast changing
attributes. Each of these mini-dimensions has its own primary surrogate
keys.
Now we study each new mini-dimension and categorize the fast changing
attributes into ranges. For example, age can be categorized into a range,
such as [1-8], [9-15], [16-25], [26-34], [35-60], and [61 and older]. Although
the Type-2 approach should not be needed to track age, age bands are often
used for other purposes such as analytical grouping. Assuming that we have
the customer date of birth, we can easily calculate age whenever needed.


However, for the purpose of designing a fast changing dimension, we cannot
in a vacuum determine the band ranges. Only the business needs should be
the deciding factor in determining which continuously variable attributes are
suitable for converting to bands.

Snowflaking does not solve the fast changing dimension problem
We now discuss snowflaking and fast changing dimensions. As an example,
Figure 5-46 depicts a split Customer dimension and created a Customer
mini-dimension. The mini-dimension is snowflaked and attached as a foreign key
to the Customer dimension. However, this is not a good design. Why?

Mini-Dimension Snowflaked
X Bad design for handling
Fast Changing Dimension

Figure 5-46 Snowflaking a fast changing dimension

Consider the sample rows shown in Table 5-19, for the
Customer_Mini_Dimension. Assume that a Customer named Stanislav Vohnik in
the Customer table is attached to the 1st row in the Customer_Mini_Dimension
table as shown in Table 5-19. At this point, Customer Stanislav Vohnik, is only
attached to one row.

Now assume that in a single year, customer Stanislav has changes to his profile
at least five times, so that he is now attached to all five rows in Table 5-19. That
means the average growth for customer Stanislav Vohnik is about five rows every
year. Imagine, if we have 10 million customers and each has an average of about
five profile changes, the growth will be about 50 million.

Table 5-19 Sample mini-dimension rows
Score Status

18 to 25 $800 to 2-4 8-12 7-9 Good 88
$1000


Score Status

18 to 25 $1000 to 5-9 8-12 10-14 Good 88
$1500

18 to 25 $1500 to 10-17 8-12 15-19 Good 88
$3000

18 to 25 $1000 to 5-9 8-12 16-17 Good 88
$1500

18 to 25 $1000 to 5-9 8-12 21-25 Very Good 88
$1500

We have now introduced the fast changing dimension problem in the Customer
dimension table despite creating a mini-dimension with band range values
because we snowflaked that dimension.

Had we attached the primary key of the Customer_Mini_Dimension table to the
Retail_Sales fact table, as shown in Figure 5-47 on page 170, we could have
solved the problem of the fast changing dimension for the Customer table. In this
scenario, the changes for the fact table record would be handled by the fact table
and the corresponding record of change in the Customer_Mini_Dimension.
There would be still only one record for Customer Stanislav Vohnik in the
Customer table. The appropriate changes would be handled in the fact table by
linking the correct profile of Customer Stanislav Vohnik at the time of purchase of
the product. The appropriate profile would be linked to the
Customer_Mini_Dimension.


Correct Design for handling
Fast Changing Dimension

Figure 5-47 Correct approach for handling fast changing dimension

Snowflaking versus mini-dimensions

Figure 5-48 on page 171 shows the difference between snowflaking and creating
a mini-dimension out of a fast changing customer dimension table. Fast changing
dimensions should be segmented into one or more mini-dimensions and not
snowflaked. Snowflaking is appropriate only in cases where low-cardinality
attributes in the dimension have been removed to separate normalized tables,
and these normalized tables are then joined back into the original dimension
table.


Mini-dimension Snowflaking
Dimension Fact Dimension
Fact Table
Table
(FK) (PK) Snowflake
(FK) (PK)
(FK) (PK)

(FK) (PK)

Mini-dimension
(Fast Changing Attributes)

Figure 5-48 Snowflaking versus mini-dimension

How many mini-dimensions for a fast changing dimension?
Assume that we have 10 fast changing attributes for a fast changing dimension.
Should there be a separate dimension for each attribute? Perhaps, but the
number of dimensions can rapidly get out of control and the fact table will have a
large number of foreign keys for these dimensions.

One approach is to combine several of these mini-dimensions into a single
physical mini-dimension. This is the same technique used to create a garbage
dimension (see5.3.8, “Identifying garbage dimensions” on page 176) that
contains unrelated attributes and flags to get them out of the fact table. It is
important to understand that the fact table must always be involved to relate
customers to their attributes (present in mini-dimension). If the mini-dimension
becomes too large, then split the mini-dimension into two or more
mini-dimensions.

5.3.7 Identifying dimensions that need to be snowflaked
In this section we look at snowflaking to see when it is practical to implement in
dimensional designs. We also look at cases of poor dimensional design where
hierarchies have been snowflaked into separate dimension tables, and as a
result performance and understandability of the dimensional model have
suffered.


What is snowflaking?
Further normalization and expansion of the dimension tables in a star schema
result in the implementation of a snowflake design. In other words, a dimension
table is said to be snowflaked when the low-cardinality attributes in the dimension
have been removed to separate normalized tables and these normalized tables
are then joined back into the original dimension table.

Typically, we do not recommend snowflaking in the dimensional model
environment, because it can impact understandability of the dimensional model
and result in a decrease in performance because more tables will need to be
joined to satisfy queries.

When do you snowflake?
Snowflaking a dimension table can typically be performed under the following two
conditions:
The dimension table consists of two or more sets of attributes which define
information at different grains.
The sets of attributes of the same dimension table are being populated by
different source systems.

To help understand when and why we snowflake, consider the sample customer
table shown in Figure 5-49.

Customer
Attributes

Customer
Country
Attributes

Figure 5-49 Customer dimension


The customer table shows two kinds of attributes. These are attributes relating to
the customer and attributes relating to the customer country. Both set of
attributes represent a different grain (level of detail) and also both sets of
attributes are populated by a different source system.

Such a customer dimension table is a perfect candidate for snowflaking for two
primary reasons:
The customer table represents two different sets of attributes. One set shows
customer attributes and the other set shows customer’s country attributes.
Both of these sets of attributes represent a different level of detail of
granularity. One set describes the customer and the other defines more about
the country. Also, the data for all customers residing in a country is identical.
On detailed analysis, we observe that the customer attributes are populated
by the company CRM source system, where the country-related attributes are
populated from an external demographic firm that has expertise in country
demographic attributes. It may also be possible another source system in the
company may also be supplying some of these attributes.

The snowflaked customer table is shown in Figure 5-50. The customer
dimension table is said to be snowflaked when the low-cardinality attributes
(customer’s country attributes) in the dimension have been removed to separate
a normalized table called country and this normalized table is then joined back
into the original customer dimension table.

Figure 5-50 Snowflaked customer table

When NOT to snowflake?

Typically, we recommend that you avoid the use of snowflaking, or normalization
of dimension tables, unless required and appropriate, as depicted in Figure 5-50.


Dimensional modelers may argue that snowflaking reduces disk space
consumed by dimension tables, but the savings are usually insignificant when
compared with the entire data warehouse. Moreover the disadvantages in ease
of use or query performance far outweigh the space savings achieved by
inappropriate snowflakes.

Figure 5-51 shows an inappropriate use of snowflakes where hierarchies that
belong to a single product dimension table have been split into separate tables.
This has an adverse affect on query performance and understandability.

Figure 5-51 Inappropriate use of snowflaking

Note: Do not snowflake hierarchies of one dimension table into separate
tables. Hierarchies should belong to the dimension table only and should
never be snowflaked. Multiple hierarchies can belong to the same dimension if
the dimension has been designed at the lowest possible detail.

Snowflaking for performance improvement
A snowflake schema is a variation on the star schema, in which very large
dimension tables are normalized into multiple tables. We do not recommend
segmenting hierarchies into snowflake tables, because users typically use
different level members of the hierarchy in viewing reports. If hierarchies are split
into separate tables, performance may be impacted as a larger number of joins
will be required.

In some situations, you may snowflake the hierarchies of a main table. For
example, dimensions with hierarchies can be decomposed into a snowflake
structure to avoid joins to big dimension tables when you are using an aggregate
of the fact table. As an example, if you have brand information that you want to
separate from a product dimension table, you can create a brand snowflake that
consists of a single row for each brand and that contains significantly fewer rows
than the product dimension table.


Figure 5-52 shows a snowflake structure for the brand and product line elements
and the brand_agg aggregate table.

Brand_Agg Aggregate Table (Snowflaked
Brand Table uses this Aggregate Table and
not the Sales Fact Table)

Figure 5-52 Snowflaking to improve performance

By creating an aggregate table that consists of the brand code and the total
revenue per brand, a snowflake schema can be used to avoid joining to the much
larger sales table. This is shown in the example query on the brand and
brand_agg tables, depicted in Example 5-3.

Example 5-3 Query directly on brand table to improve performance
SELECT brand.brand_name, brand_agg.total_revenue
FROM brand, brand_agg
WHERE brand.brand_code = brand_agg.brand_code
AND brand.brand_name = 'Anza'

Without a snowflaked dimension table, a SELECT UNIQUE or SELECT
DISTINCT statement on the entire product table (potentially, a very large
dimension table that includes all the brand and product-line attributes) would
have to be used to eliminate duplicate rows.

While snowflake schemas are unnecessary when the dimension tables are
relatively small, a retail or mail-order business that has customer or product
dimension tables that contain millions of rows can use snowflake schemas to
significantly improve performance.


If an aggregate table is not available, any joins to a dimension element that was
normalized with a snowflake schema must now be a three-way join, as shown in
Example 5-4. A three-way join reduces some of the performance advantages of a
dimensional database.

Example 5-4 Three-way join reduces performance
SELECT brand.brand_name, SUM(sales.revenue)
FROM product, brand, sales
WHERE product.brand_code = brand.brand_code
AND brand.brand_name = 'Alltemp'
GROUP BY brand_name

Disadvantages of snowflaking
The greater the number of tables in a snowflake schema, the more complex the
design.
More tables means more joins, and more joins mean slower queries.
It is best not to use snowflaking to try to save space. Space saved by
snowflaking is extremely small compared to the overall space of the fact table
in the dimensional design.

5.3.8 Identifying garbage dimensions
A garbage dimension is a dimension that consists of low-cardinality columns
such as codes, indicators, and status flags. The garbage dimension is also
referred to as a junk dimension. The attributes in a garbage dimension are not
related to any hierarchy.

We now review the dimensional model (see Figure 5-53 on page 177) for the
grocery store example described in Chapter 6, “Dimensional Model Design Life
Cycle” on page 227. The grain for this dimensional model is a single line item on
the grocery bill or invoice.

To understand the concept of a garbage dimension, assume that we need to add
three additional things to this dimensional model. They are:
Customer feedback about the store, and customer feedback by employee,
categorized as good, bad, and none.
Method used by the customer to pay for the products, categorized as cash or
credit card.
Bag selected to carry the goods, categorized as plastic, paper, or both.


Figure 5-53 Grocery store example

The next question is, how do we accommodate these new attributes in the
existing dimensional design. The following are the options:
Put the new attributes relating to customer feedback, payment type, and bag
type into existing dimension tables.
Put the new attributes relating to customer feedback, payment type, and bag
type into the fact table.
Create new separate dimension tables for customer feedback, payment type,
and bag type.

First, we cannot put the new attributes into existing dimension tables because
each of the new attributes has a different grain than the product, supplier,
employee, time, customer, store, or date table. If we tried to forcibly add these
new attributes, we would create a cartesian product for the dimension tables. For
example, we merge the payment type (Credit card or Cash) attribute with the
product table. Assume that we have 100 product rows in the product table. After
adding the payment type attribute to the product table, we have 200 product
rows. This is because one product would need to be represented with two rows,
for both credit card and cash payment type. Because of this reason, we do not
force adjusting these attributes inside any of the existing dimension tables.

Second, we take the fact table. We do not recommend adding low-cardinality
fields into the fact table because it would result in a very huge fact table.

Note: Fact tables generally occupy 85-95% of the space in most dimensional
designs. To optimize space, recall that fact tables should only contain facts,
degenerate dimension numbers, and foreign keys. Fact tables should not be
used to store textual facts or low-cardinality text fields.

Our third choice was to create a separate dimension for each of new attributes
relating to customer feedback, payment type, and baggage type. The
disadvantage of this approach is that we will have three additional foreign keys


inside the fact table. Having the three foreign keys in the table is equivalent of
increasing the row size in the fact table by 4 x 3 = 12 bytes. A fact table that has,
for example, 100,000 rows inserted into it on a daily basis, would increase about
12 times with an increase of 12 bytes per row. This increased size would impact
the performance of the fact table.

Note: Avoid creating separate dimensions for low-cardinality fields such as
those which hold flags and indicators.

The best way to solve the problem associated with low-cardinality attributes,
such as those relating to customer feedback, payment type, and baggage type,
is to create a single dimension called a garbage dimension. An example is
depicted
in Figure 5-54. A garbage dimension acts as a storage space for the
low-cardinality fields.

Garbage Dimension
consisting of the
Low-cardinality Fields

Figure 5-54 Garbage dimension implementation

Table 5-20 on page 179 shows sample rows inside the Garbage_Dimension
table. The total number of rows inside a garbage dimension is equal to the
number of rows formed after taking the cartesian product of all individual values
that reside inside each of the low-cardinality attributes. Therefore, the number of
rows inside the Garbage_Dimension table is 3 x 3 x 2 x 3 = 54 rows. A new
surrogate key value is also generated for each combination. The following are
the possible values for each of the low-cardinality fields:
Customer feedback: Good, bad, and none
Customer feedback about employee: Good, bad, and none
Payment type: Cash or credit card
Bag type to carry goods: Plastic, paper, or both


Table 5-20 Sample rows inside the garbage dimension
Garbage_ Customer_Store Customer_ Payment_ Bag_Type
Dim-Key _Feedback Employee_ Type
Feedback

1 Good Good Credit card Plastic

2 Good Bad Credit card Plastic

3 Good None Credit card Plastic

Note: If the garbage or junk dimension grows too large, then you must split it
into two or more dimensions. The total number of rows inside a garbage
dimension is equal to the number of rows that are formed after taking the
cartesian product of all individual values that reside inside each of the
low-cardinality attributes.

How should you load the garbage dimension?
There are two ways to load the garbage dimension. They are:
Preload the dimension: In this approach, you identify all the possible
combinations of values that could result from the cartesian product of all
individual values that reside inside each of the low-cardinality attributes. Then
you load the dimension table. This approach works well if the number of rows
that result from the cartesian product are few.
Load at run time: This approach works well if the garbage or junk dimension
has a huge number of rows. However, in this scenario, you might not
pre-populate all combinations.

5.3.9 Role-playing dimensions
A single dimension which is expressed differently in a fact table using views is
called a role-playing dimension.

Figure 5-55 on page 180 shows a dimensional model designed for an order
management business process. There are two date entities (order and received
date) at the same day grain involved.


2 Date Tables
being used at
the same day
level granularity

Figure 5-55 Order management dimensional model

Instead of using two separate date tables, Order_Date and Received_Date with
the same granularity, we can create two views from a single date table (Date) as

The Order_Date_View and Received_Date_View serve as the role-playing
dimensions as depicted in Figure 5-56 on page 181.


Order_Date View
Based on Date Table

Received_Date View
Based on Date Table

Figure 5-56 Role-playing dimension

The Order_Date_View and Received_Date_View dimension tables are
implemented as views, as shown in Example 5-5 and Example 5-6.

Example 5-5 Order_Date_View dimension table implementation
Create View ORDER_DATE_VIEW(ORDER_DateId,Order_Date, Order_Date_Description,
Order_Day_of_Week, Order_Calendar_Month, Order_Calendar_Quarter,
Order_Calendar_Year, Order_Fiscal_Week, Order_Fiscal_Month,
Order_Fiscal_Quarter, Order_Fiscal_Year, Order_Holiday_Flag,
Order_Weekday_Flag, Order_Weekend_Flag, Order_Season, Order_National_Event) as
select * from Date

There is only one date table, but it is projected differently using views.

Example 5-6 Received_Date_View dimension table implementation
Create View Received_DATE_VIEW(Received_DateId,Received_Date,
Received_Date_Description, Received_Day_of_Week, Received_Calendar_Month,
Received_Calendar_Quarter, Received_Calendar_Year, Received_Fiscal_Week,
Received_Fiscal_Month, Received_Fiscal_Quarter, Received_Fiscal_Year,
Received_Holiday_Flag, Received_Weekday_Flag, Received_Weekend_Flag,
Received_Season, Received_National_Event) as
select * from Date


5.3.10 Multi-valued dimensions
We discussed in section 6.4.1, “Dimensions” on page 259 that each dimension
attribute should take on a single value in the context of each measurement inside
the fact table. However, there are situations where we need to attach a
multi-valued dimension table to the fact table. There are situations where there
may be more than one value of a dimension for each measurement. These
cases are handled using multi-valued dimensions.

Consider the design for a dimensional model for a family insurance company.
Assume that Mr. John is the family insurance policy account holder and he has
the policy account number 963276. Mr. John is married to Lisa and they have a
son named Dave. Mr. John later decides to get a joint family policy under the
same policy account number 963276. This means is that the three family
members hold a joint policy for the same account 963276.

Assume that the family insurance company charges a particular amount for the
entire family, which is insured under a common policy account number 963276.

Now assume we are asked to create a data mart for this insurance company and
the business users have the following set of requirements:
Business is interested in analyzing the revenue generated each month after
the policy account holders pay differing amounts for their family insurance.
Business is interested in analyzing the revenue generated for different states
and counties.
Business is also interested in analyzing the data to see individual members
associated with each policy. That is, they want to know how many members,
and their identity, are associated with a single policy number. For example,
the business may be interested to know that for policy account number
963276, Mr. John is the primary policy holder and his wife, Lisa, and their son,
Dave, are associated with it.
To create the dimensional model, we decide the grain of the fact table is one
row for each family policy account at the end of each month. We create a
dimensional model design as depicted in Figure 5-57 on page 183.


Single Account Number
963276 is associated
with 3 Insured persons
(John, Lisa and Dave).

Monthly_Premium_Paid (Fact) is
associated with many values, such
as John, Lisa, and son Dave.
(Multi-valued Concept)
Figure 5-57 Family insurance dimensional model

The multi-valued dimension
As it is shown in Figure 5-57, an individual family insurance policy account can
have one, or more, customers associated with it. For example, for the policy
account number 963276, there are three insured customers.

To represent that there is more than one customer associated with the fact
(Monthly_Premium_Paid), we cannot merely include the customer as an account
attribute. By doing so, we would be violating the granularity of the dimension
table because more than one individual can be associated with a single family
insurance account number.

Similarly, we cannot include an additional customer dimension in the fact table
because we have declared the grain to be one row per family policy account. So,
if we include the customer dimension we would be violating the grain because
there would be one policy account associated to more than one customer. This is
a typical example of a multi-valued dimension.

We can solve this multi-valued dimension problem with the use of an
Account-Insured_Person bridge table as shown in Figure 5-58 on page 184. The
primary key of the Account-Insured_Person bridge table consists of the surrogate
Account and Insured_Person foreign keys. Therefore, using the bridge table, we
are able to associate three customers, namely Mr. John, his wife, Lisa, and son,
Dave, to the same family policy account number without violating the grain
definition.


Bridge Table

Figure 5-58 Use of a bridge table for multi-valued dimension

What is the weighting factor?
A weighting factor is an important column placed inside the bridge table. We
assign a numerical weighting factor to each insured person for each family policy
account number such that the sum of all the weighting factors belonging to a
single group or account number is exactly 1. The weighting factor is simply a way
to allocate the numeric additive facts across the insured persons that are present
in the Insured_Person dimension table.

Approaches for handling multi-valued dimensions
There are situations when we need to attach multiple values for a dimension to a
single fact table row, or when we need to attach a multi-valued dimension table to
the fact table. We described an example of a multi-valued dimension in
Figure 5-58. Another example of a multi-valued dimension is where we associate
many customers to a single bank account. Or, when multiple diagnoses are
associated with single patient. Dimensional modelers usually take one of
following approaches for handling multi-valued dimension attributes:
Choose one particular value and leave the others: This is the most
frequently used technique, but it should not be forced. If we use this
approach, the multi-valued dimension problem is eliminated, but the
dimensional model may still not be useful because of missing dimensions.
Extend the dimensions list to include a fixed number of multi-valued
dimensions: This is not a good approach because it may increase the
number of dimensions for each multi-value. Also, the number of values
(dimensions) we may need to create are not fixed and may vary from case to
case. If we create a fixed number of multi-valued dimensions and if the
number of multi-values increases, then our design will not be flexible enough
to handle it well.


Use a bridge table: The bridge table may be inserted in two ways:
a. Between the fact and dimension table, as shown in Figure 5-59, we chose
the grain as one single line item on each order. We observe that for each
order completion, there are one or more sales representatives that are
involved with the sale and work together to get the order. The weighting
factor column shows the relative contribution of each in getting the order.
Assume that two persons work to get an order (Order Number# 99).
Person A is contributes 90% of the overall time, where Person B only
contributes 10%. There would be one row for the entire order (Order
Number# 99), but there would be two rows for Person A and Person B in
the ORDER-to-SALES REP BRIDGE table. The weighting factor for
Person A is .9, since Person A contributed 90%, and for Person B the
weighting factor is .1 since Person B’s contribution was 10%. There are
two rows in the “SALES_REPRESENTATIVE” table.

Use of Bridge Table to show
that more than 1 Employee
may be involved in 1 Order
Note: Grain is a single entire Order
Figure 5-59 Multi-valued dimension example

b. The other way was depicted in Figure 5-58 on page 184. Here, the bridge
table is inserted between the Account dimension table and
Insured_Person dimension table to show that multiple persons are
associated with the same account.

5.3.11 Use of bridge tables
Bridge tables are used to express a many-to-many relationship between a fact
and a dimension table. In dimensional modeling, the bridge tables are primarily
used for two reasons:
To solve the complex unbalanced hierarchies problem as discussed in
“Unbalanced hierarchy” on page 145.
For solving the multi-valued dimension problem as discussed in “Multi-valued
dimensions” on page 182.


5.3.12 Heterogeneous products
The term heterogeneous means mixed, assorted, or diverse. This is exactly
the state of many businesses. That is, they are selling diverse products to a
common customer base. To better understand the concept, assume that we
work for an insurance company called ABC 979 Insurance Inc., that sells
products in the form of insurance policies. In particular, the company sells
home and car insurance. The concept of heterogeneous products comes to
light when a company sells different products with different unique attributes
to the same customer base.
Moreover, the attributes the business tracks for home insurance are very
different from those collected for car insurance. This is a typical example of a
company selling heterogeneous products.
There are several approaches of handling heterogeneous products within the
company. Several approaches are explained below:
Merge attributes: Merge all the attributes into a single product table and all
facts relating to the heterogeneous attributes in one fact table. This approach
is depicted in Figure 5-60. The disadvantage of this approach is that all
unrelated attributes for home and car insurance are merged into one single
dimension. This makes the insurance dimension very large. Also for every
row inside such a dimension, many columns belonging to home insurance are
NULL for car insurance, because they belong to a different type. For this
approach, a single fact table is created in which all facts for all heterogeneous
products are merged, so the fact table grows very large.
You should avoid this approach because the huge size of the tables can result
in performance issues.

Figure 5-60 Merge all heterogeneous product attributes into one table

Separate tables: Create separate dimension and fact tables for different
heterogeneous products, as depicted in Figure 5-61 on page 187. With this
approach, we create separate dimensions and facts for the heterogeneous
products.


(Home Insurance)

(Car Insurance)

Figure 5-61 Separate dimension and fact tables to handle heterogeneous products

The advantage of having separate dimension and fact tables for such
heterogeneous products is that the specifics about these individual
businesses (Home and Car Insurance) can be analyzed in more detail.
Generic design: Create a generic design to include a single fact and single
product dimension table with common attributes from two or more
heterogeneous products, as depicted in Figure 5-62. In this approach, we
identify common attributes between heterogeneous dimensions and model
them as a single dimension. Also, we identify and create generic facts, such
as Transaction Amount or Revenue, earned for the two heterogeneous
dimensions. This way the business can analyze the two business processes
together for common facts.

Generic Fact (Transaction Amount)
Common Attributes for the Heterogeneous Products
For Heterogeneous
Insurances, such as
Car and Home

Figure 5-62 Generic design to handle heterogeneous products


The disadvantage of this approach is that only common attributes between
the heterogeneous types of insurance (Home and Car) are available for
analysis. Also, there is only a generic fact called the Transaction Amount.
However, for detailed, specific business facts and attributes about each type
of insurance (Car and Home), management can refer to the separate
dimensional models as shown in Figure 5-61 on page 187.

Note: The concept of heterogeneous products applies to situations where the
customer base is common. That is, if the company is selling heterogeneous
products to altogether different customers, then such products would typically
belong to separate data marts and perhaps also separate data warehouses.

5.3.13 Hot swappable dimensions or profile tables
A dimension that has multiple alternate versions of itself that can be swapped
at query time is called a hot swappable dimension or profile table. Each of
the versions of the hot swappable dimension can be of different structures.
The alternate versions of the hot swappable dimensions access the same fact
table, but with different output. The different versions of the primary dimension
may be completely different, including incompatible attribute names and
different hierarchies.
Unlike in a relational database, building hot swappable dimensions in OLAP is
complicated because the joins between the tables cannot be specified at the
query time.

Note: Swappable dimensions may be implemented to improve performance of
the dimensional model. The swappable dimensions also help create more
secure dimensions.

How to implement a swappable dimension
Create separate physical dimensions from the primary dimension, which has
only a subset of data (columns and rows) of interest. The new dimension can
then be swapped instead of using the primary dimension at run time by
joining the key of the swappable dimension to the foreign key inside the fact
table. The fact table remains the same for the primary dimension and all
swappable dimensions.
Create one or more views based on the primary dimension. These views can
be swapped at run time to join to the fact table instead of the primary
dimension table.
Figure 5-63 on page 189 shows a dimensional model for an order
management system, with dimensions such as Product, Customer, Supplier,
and Sales_Channel for which we have created swappable dimensions.


Hot Swappable Dimensions

Employees
Internet

Direct Consumer Catalog
Customer Sales_Channel Local Store
Business Consumer
Warehouse
Facts Business Partner

Date

Goods
Services Product
Yearly Subscriptions Supplier_S1
Monthly Subscriptions
Supplier Supplier_S2
Supplier_S3
Supplier_S4

Hot Swappable Dimensions
Figure 5-63 Hot swappable dimensions

As shown in Figure 5-63, all business users still use the same fact table
“Sales_Fact” which is still stored in a single place. However, each business user
may swap the different versions of dimensions at query time.

Note: An important concept behind using swappable dimensions is that each
of the swappable dimensions can be used in place of the primary dimension to
join to the same fact table. The swappable dimensions have less data when
compared to the primary dimension and so this way fewer rows (of swapping
dimension) join to the huge fact table, and therefore performance is faster.
Also, using swappable dimensions, you can secure data by allowing the user
to access only the swapped dimension which contains restricted data. This
way the user views only the swapped dimension and not the entire primary
dimension.

The description of each primary dimension and its corresponding swappable
dimension is explained in Table 5-21 on page 190.


Table 5-21 Primary dimension and corresponding swappable dimensions
Primary Swappable Description for swappable dimension
dimension dimension

Product Goods The Goods swappable dimension is derived
from product dimension, and consists of all
products which are goods.

Services The Services swappable dimension is derived
from product dimension, which consists of all
products which are offered as services.

Yearly The Yearly Subscription swappable dimension
Subscription is derived from product dimension and consists
of all products which are subscribed to yearly.

Monthly The Monthly Subscription swappable dimension
Subscription is derived from product dimension and consists
of all products which are subscribed to monthly.

Date None None

Sales_Channel Internet The Internet swappable dimension is derived
from Sales_Channel dimension and consists of
all products that are sold through the Web.

Catalog The Catalog swappable dimension is derived
from Sales_Channel dimension and consists of
all products which are sold through the catalog.

Local_Store The Local_Store swappable dimension is
derived from Sales_Channel dimension and
consists of all products that are sold through the
local stores.

Warehouse The Warehouse swappable dimension is
derived from Sales_Channel dimension and
consists of all products that are sold through the
warehouses.

Business_ The Business_Partner swappable dimension is
Partner derived from the Sales_Channel dimension and
consists of all products which are sold through
the business partners.


Primary Swappable Description for swappable dimension
dimension dimension

Supplier Supplier_S1 The Supplier_S1 swappable dimension is
(Goods) derived from Supplier dimension, and consists
of all products (goods) which are supplied by the
supplier “S1”.

Supplier_S2 The Supplier_S2 swappable dimension is
(Goods) derived from Supplier dimension, and consists
of all products (goods) which are supplied by the
supplier “S2”.

(Services) derived from Supplier dimension, and consists
of all products (services) which are supplied by
the supplier “S3”.

(Subscriptions) derived from Supplier dimension, and consists
of all products (all subscriptions) which are
supplied by the supplier “S4”.

Customer Direct_ The Direct_Consumer swappable dimension is
Consumer derived from Customer dimension, and consists
of all direct customers.

Business_ The Business_Consumer swappable dimension
Consumer is derived from Customer dimension, and
consists of all business customers.

5.4 Facts and fact tables
In this section, we discuss challenges that surface because of the particular
attributes being used with facts and fact tables.

5.4.1 Non-additive facts
Non-additive facts are facts which cannot be added meaningfully across any
dimensions. In other words, non-additive facts are facts where the SUM operator
cannot be used to produce any meaningful results. Examples of non-additive
facts include:
Textual facts: Adding textual facts does not result in any number. However,
counting textual facts may result in a sensible number.
Per-unit prices: Adding unit prices does not produce any meaningful
number. For example, unit sales price or unit cost is strictly non-additive


because adding these prices across any dimension will not yield a meaningful
number. If instead we store the unit cost as an extended price (such as
per-unit cost x quantity purchased), it is correctly additive across all
dimensions.
Percentages and ratios:
A ratio, such as gross margin, is non-additive. Non-additive facts are usually
the result of ratio or other calculations, such as percentages. Whenever
possible, you should replace such facts with the underlying calculation facts
(numerator and denominator) so you can capture the calculation in the
application as a metric. It is also very important to understand that when
adding a ratio, it is necessary to take sums of numerator and denominator
separately and these totals should be divided.
Measures of intensity: Measures of intensity such as the room temperature
are non-additive across all dimensions. Summing the room temperature
across different times of the day produces a totally non-meaningful number as
shown in Figure 5-64. However, if we do an average of several temperatures
during the day, we can produce the average temperature for the day, which is
a meaningful number.

Month Avg. Units Price Day Temperature
January $10 Jan 9th (6 AM ) 65 F
February $20 Jan 9th (7 AM ) 65 F
March $30 Jan 9th (9 AM ) 65 F
Total $60 X Total 195 F X
(a) Average Units Price (Non-Additive) (b) Temperature (Non-Additive)

Figure 5-64 Non-additive facts

Note: You may avoid storing non-additive facts, such as unit prices, as facts in
the fact table. Instead you may store fully additive facts such store quantity
sold and amount charged (unit price x quantity). Then to get the unit price of
the product, just divide the amount charged by the quantity.

Averages: Facts based on averages are non-additive. For example, average
sales price is non-additive. Adding all the average unit prices produces a
meaningless result as shown in Figure 5-64.
Degenerate Numbers: Degenerate dimensions are also non-additive.
Number, such as order number, invoice number, tracking number,


confirmation number, and receipt number, are stored inside the fact table as
degenerate dimensions. Counts of such numbers produce meaningful results
such as total number of orders received or total number of line items in each
distinct order.

Note: Non-additive facts, such as degenerate dimensions, may be counted. A
count of degenerate dimensions, such as order numbers, produces
meaningful results.

5.4.2 Semi-additive facts
Semi-additive facts are facts which can be summarized across some dimensions
but not others. Examples of semi-additive facts include the following:
Account balances
Quantity-on-hand

We now discuss these types of semi-additive facts in more detail and see how to
handle these in dimensional modeling and in reports.

(a) Account balances
Account balances are typically semi-additive facts. Consider a customer named
Chuck whose account balance is shown in Figure 5-65. Chuck has an initial
deposit of $1000 on January 1, 2005, withdraws $100 on January 2, deposits
$600 on January 7, withdraws $500 on January 9, withdraws $300 on January
15, and finally deposits $1000 on January 31. His account balance at the end of
January 2005 is $1700.

Date Transaction Amount Balance
January 1st, 2005 $1000
January 2nd , 2005 -100 (Withdrawal) $900
January 7th , 2005 +600 (Deposit) $1500
January 9th , 2005 -500 (Withdrawal) $1000
January 15th , 2005 -300 (Withdrawal) $700
January 31st , 2005 1000 (Deposit) $1700
For January Total $6800 X
Account Balances cannot be added
across Date (Time) Dimensions

Figure 5-65 Account balances are semi-additive


As shown in Figure 5-65 on page 193, adding the monthly balances across the
different days for the month of January results in an incorrect balance figure.
However, if we average the account balance to find out daily average balance
during each day of the month, it would be valid.

How can account balance be calculated across other dimensions?

We mentioned that a semi-additive fact is a fact that is additive across some
dimensions but not others. Consider the star schema shown in Figure 5-66 on
page 195. We see how and why the account balance is additive across the
Customer, Branch, and Account dimensions, but not across the Date dimension.
This is shown in Table 5-22.

Table 5-22 Account balance is semi-additive
Dimension Account balance Why?
additive?

Date NO Explained in example in Figure 5-65

Branch Yes See SQL query in Example 5-7 and result in
Figure 5-67.

Account Yes See SQL query in Example 5-8 and result in
Figure 5-68.

Customer Yes See SQL query in Example 5-9 and result in
Figure 5-69.

With the help of SQL queries, we can see that adding account balance along
other dimensions, except date, can provide a meaningful measure for the total
amount of money the bank is holding at any given point in time.


Semi-Additive
Fact

* Balance can be added across Branch,
Customer and Account over a single day

Figure 5-66 Account balance star schema

We use SQL queries to show that account balances for a single day are additive
across the following dimension:
Account balance is additive across Branch. It can be added at a single day
across the branch dimension. The SQL code to show this is displayed in
Example 5-7.

Example 5-7 Sample SQL showing balance is additive across branch dimension
Select B.BRANCH_NAME, SUM(F.BALANCE)
From
Branch B, Account_Balance_Fact F, Date D
where
B.BRANCH_KEY= F.BRANCH_KEY
and
F.DATE_KEY= D.DATE_KEY
and
D.DATE_DAY=’1’
and
D.DATE_MONTH=’January’
and
D.DATE_YEAR=’2005’
GROUP BY B.BRANCH_NAME

The result of the query in Example 5-7 is shown in Figure 5-67 on page 196.
It shows that the balance is additive at the branch level.


Date: January 1, 2005
BRANCH NAME BALANCE

South West Branch $1000000

North East Branch $4095962

Oriental Bank of Commerce Branch $9090000

Balance is additive across
Branch at a single day.

Figure 5-67 Balance is additive across the branch dimension

Note: It is important to understand that balance is semi-additive across the
branch dimension only for a particular day. If we do not include the where
clause for a day, then the SUM we get is an incorrect figure.

Account balance is additive across Account dimension. Account balance can
be added at a single day across the account dimension. The SQL code to do
this is shown in Example 5-8.

Example 5-8 Sample SQL showing Balance is additive across Account dimension
Select A.ACCOUNT_TYPE, SUM(F.BALANCE)
From
Account A, Account_Balance_Fact F, Date D
where
A.ACCOUNT_KEY= F.ACCOUNT_KEY
and
and
D.DATE_DAY=’1’
and
and
GROUP BY A.ACCOUNT_TYPE

The result of the query shown in Example 5-8 is shown in Figure 5-68 on
page 197. It shows that balance is additive across the Account dimension
(per account type).


BRANCH NAME BALANCE

Savings Account $98300000

Long Term Account $40666962

Joint Account $90909111

Balance is Additive across
Account Type at a Single Day.

Figure 5-68 Balance is additive across the account dimension


Account balance is additive across Customer dimension. It can be added at a
single day across the customer dimension. The SQL code to do this is shown
in Example 5-9.

Example 5-9 Sample SQL showing Balance is additive across Customer dimension
Select C.CUSTOMER_CITY, SUM(F.BALANCE)
From
Customer C, Account_Balance_Fact F, Date D
where
C.CUSTOMER_KEY= F.CUSTOMER_KEY
and
and
D.DATE_DAY=’1’
and
and
GROUP BY C.CUSTOMER_CITY

The result of the query shown in Example 5-9 is shown in Figure 5-69 on
page 198. It shows that balance is additive per customer city.


CUSTOMER CITY BALANCE

New Delhi $9999000

Bombay $4095962

Munirka and Mayur Vihar City $998899889

Balance is Additive across
Customer City at a Single Day.
Figure 5-69 Balance is additive across the customer dimension


Also, if instead of choosing customer city we choose customer name, the sum
of the balance could get summed up as a single value for all people having the
same names. That is, if there were the customers with the name Suniti, all
three balances would have summed up into a single name Suniti. Although
balance is a semi-additive fact across the Customer dimension, it is for the
user to decide whether or not the case for semi-additivity is good or not for a
particular column.

(b) Quantity-on-hand (Inventory)
Quantity-on-hand (inventory or stock remaining) is another typical semi-additive
fact. To understand this, consider a store that maintains inventory for each of its
products at the end of every month. This is shown in Figure 5-70 on page 199.
For a product named P99, the store has an inventory of 9000 at the end of
January, 2005. For the same product P99, the store has an inventory of 5000 at
the end of February, 8000 at the end of March, 1000 at the end of April, and 2000
at the end of May. The store’s final inventory (quantity-on-hand) for product P99
at the end of June, is 1000.


Product Name Month Ending Quantity on Hand
P99 January 9000
P99 February 5000
P99 March 8000
P99 April 1000
P99 May 2000
P99 June 1000
For January Total 2600 X

Quantity-on-hand, or Remaining Stock
Balance, cannot be added across
Date (Time) Dimensions
Figure 5-70 Quantity-on-hand is semi-additive

As shown in Figure 5-70, adding the month end quantity-on-hand stocks across
the different months results in an incorrect balance figure. However, if we
average the quantity on hand to find out the monthly average balance during
each month of the year, it is valid.

How can Quantity-on-hand be calculated across other dimensions?

A semi-additive fact is a fact that is additive across some dimensions, but not
others. Consider the star schema shown in Figure 5-71 on page 200. We can
see that the quantity-on-hand is additive across the product and store
dimensions but not across the Date dimension. This is shown in Table 5-23.

Table 5-23 Quantity-on-hand is semi-additive
Dimension Account balance Why?
additive?

Date NO Explained in Figure 5-65.

Product Yes See SQL query in Example 5-7 and result in
Figure 5-67.

Store Yes See SQL query in Example 5-8 and result in
Figure 5-68.

With the help of SQL queries, we will see that adding quantity-on-hand along
other dimensions such as store can provide a meaningful measure for the total
quantity of products the stores are holding at any given point in time. We will also
see that adding quantity of stock remaining across product (category) or product


(name) gives us an idea of the total stock remaining (across all stores) at a given
point in time.

Semi-Additive Fact

* Quantity-on-hand can be added across
Product and Store for every Month
Figure 5-71 Store inventory on hand star schema

We use SQL queries to show that quantity-on-hand for every month is additive
across the following dimensions:
Store dimension. In Example 5-10, we provide the SQL code we used to
show that Quantity-on-hand can be added for every month across the store
dimension.

Example 5-10 Sample SQL: Quantity-on-hand additive across Store dimension
Select S.STORE_CITY, SUM(F.QUANTITY_ON_HAND)
From
Store S, Quantity_On_Hand_Fact F, Date_Month D, Product P
where
S.STORE_KEY= F.STORE_KEY
and
F.MONTH_KEY= D.MONTH_KEY
and
F.PRODUCT_KEY= P.PRODUCT_KEY
and
and
and
P.PRODUCT_NAME=’P99’ // NOTE: Important to include otherwise we get a sum of
//all unrelated products
GROUP BY S.STORE_CITY


The result of the query from Example 5-10 on page 200 is shown in
Figure 5-72. It shows the quantity-on-hand addition per store city for a
particular product. However, it is important to note that you should include the
product name in the where clause to specify a particular product. If we did not
include the product in the where clause, we would get a report showing the
sum of all products (P1 + P2 + P3 + P4 + P5 + ......... Pn) that are present in
any store city. We certainly do not want that because if we add the sum of all
remaining products (quantity-on-hand) for products such as CDs, pencils,
paper, books, and pens, it would result in a meaningless value.

Product :P99
Month Ending: January
STORE CITY QUANTITY ON HAND

New Delhi 10000009

San Jose 40959600

San Mateo 90900000

*One city can have Quantity on Hand is Additive
many Stores across Store City For Every
Month for a single Product (P99)

Figure 5-72 Quantity-on-hand is additive across the store dimension


Quantity-on-hand is additive across product. It can be added for every month
across the product (category) dimension. When we sum up all products
present inside a category of a store, we have the quantity-on-hand for all
products belonging to a particular category of a particular store. The SQL
code to show this is depicted in Example 5-11.

Example 5-11 Sample SQL: Quantity-on-hand additive across product dimension
Select P.PRODUCT_CATEGORY, SUM(F.QUANTITY_ON_HAND)
From
Quantity_On_Hand_Fact F, Date_Month D, Product P, Store S,
where
F.MONTH_KEY= D.MONTH_KEY
and
F.PRODUCT_KEY= P.PRODUCT_KEY
and


F.STORE_KEY= S.STORE_KEY
and
and
and
S.STORE_NAME=’S1’

GROUP BY P.PRODUCT_CATEGORY

The result of the query shown in Example 5-11 on page 201 is shown in
Figure 5-73. However, there is an important note here. It is important to
include the store name in the where clause to make it a particular store. If we
do not include the store in the where clause, we would get a report showing
the sum of all products belonging to the same category (P1 + P2 + P3 + P4 +
P5 + .........Pn) that are present in any store. The result shows that
quantity-on-hand is additive per product category per store.

Store: S1
PRODUCT CATEGORY QUANTITY ON HAND

Meat 20030445

Cereal 4039032

Milk 949202

* Quantity on hand is Quantity on hand is additive
additive across across Customer City at a
product per store Single Day.

Figure 5-73 Quantity-on-hand is additive across the product dimension

Note: It is important to understand that “quantity on hand” is semi-additive
across the branch dimension only for a particular day. If we do not include the
where clause for a day, then the SUM we get is incorrect.

5.4.3 Composite key design for fact table
A fact table primary key is typically comprised of multiple foreign keys, one from
each dimension table. Such a key is called a composite or concatenated primary
key. It is not a mandatory rule to have all the foreign keys included as the primary
key of the fact table. Sometimes, only a few combinations of foreign keys will be
used.


Also, it is not always true that the combination of all foreign keys of the
dimensions in the fact table will guarantee the uniqueness of the fact table
primary key. In such situations, you may need to include the degenerate
dimension as a component inside the primary key of the fact table. It is
mandatory that such a primary key be unique.

Note: It is important to understand that some or all foreign keys (of
dimensions) present inside the fact table may guarantee uniqueness and may
be used to create the primary key of the fact table.

Composite primary keys and uniqueness

Does the composite primary key design consisting of all dimension foreign keys
guarantee uniqueness? The answer is that uniqueness of the composite primary
key of the fact table is guaranteed by the grain definition and it is not a rule that all
dimension keys will always be unique if the grain definition is not enforced
properly. For example, consider the fact table shown in Figure 5-74 on page 204.
The granularity of the fact table is at the day level. Therefore, we need to sum all
of the transactions to the day level. That would be the definition of the fact.

Consider product dimension, store dimension, date dimension, and fact table as
shown in Figure 5-74 on page 204. Assume that the product dimension has only
two bicycle brands - C1 and C2. The date dimension is at the day level and
contains one row for each day.

Assume that you sell the following number of bicycles on October 22, 2005:
4 Bicycles of Brand C1 in the morning at 8:00 a.m.
5 Bicycles of Brand C1 in the evening at 6:00 p.m., before the store closes.

Since the grain of the fact table in Figure 5-74 on page 204 is one row for each
product sold in each store, we will have one row inserted for the sales of Bicycle
of Brand C1 on October 22, 2005. In this single row, the fact (Sales_Quantity)
would be 9 (4 + 5).

Grain guarantees UNIQUENESS of the composite fact primary key

If we design the primary key of the fact table (see Figure 5-74 on page 204) as
the combination of the foreign keys of all dimensions, then the (Retail_Sales)
primary key would be composite key of Product_ID, Store_ID, and Date_ID. This
primary key is guaranteed to be UNIQUE because of the fact that only one row
will be inserted for each product in each store on a single day. The composite
primary key UNIQUENESS is guaranteed by the GRAIN of the fact table which
says that only ONE row is inserted for each product in each store (no matter how
many times it is sold during the day).


Grain = Single product sold in
each store on a daily basis

Figure 5-74 Retail store schema

Suppose that we want to report sales on an hourly basis. To do so, we design a
star schema as shown in Figure 5-75 on page 205. Assume during a day
(October 14, 2005) we have the following sales:
5 Bicycles of Brand C1 sell at 8:00 a.m.
3 Bicycles of Brand C1 sell at 4:00 p.m.
2 Bicycles of Brand C1 sell at 8:00 p.m.

Since the grain of the fact table in Figure 5-75 on page 205 is one row for each
product sold in each store on an hourly basis, we have 4 rows (instead of five)
inserted for the sales of Bicycles of Brand C1 for October 14, 2005.

Note: Only one single row (Selling_Amount=10) will be inserted, two different
times (8:00 a.m. and 8:30 a.m.) because the grain definition asks us to track
sales on an hourly basis. If several sales occur for the same product in the
same hour, then only one single row would be inserted for the product for that
hour. In other words, the UNIQUENESS of the fact table row (fact composite
primary key) is guaranteed by the GRAIN definition of the fact table.


Grain = Single product sold in
each Store each Hour
Figure 5-75 Sales schema to report hourly sales

If we design the primary key of the fact table in Figure 5-75 as a composite key
consisting of the Product_ID, Date_ID, Time_ID, and Store_ID, the primary key
would be guaranteed to be UNIQUE because the GRAIN of the fact table
assures that only a single row will be inserted for a single product sold on an
hourly basis.

Note: One or more degenerate dimensions may be needed to guarantee the
uniqueness of the fact table row. A detailed discussion is available in 6.5.6,
“Composite key design” on page 301.

5.4.4 Handling event-based fact tables
Event fact tables are tables that record events. For example, event fact tables are
used to record events such as Web page clicks and employee or student
attendance. Events, such as a Web user clicking on a Web page of a Web site,
do not always result in facts. In other words, millions of such Web page click
events do not always result in sales. If we are interested in handling such
event-based scenarios where there are no facts, we use event fact tables which
consist of either pseudo facts or these tables have no facts (factless) at all.

From a conceptual perspective, the event-based fact tables capture the
many-to-many relationships between the dimension tables. To better understand
the concept of event tables, consider the following two examples:
Example one: Hospital and insurance revenue example


Example two: Employee attendance recording

Note: Event-based fact tables may have pseudo facts or no facts (factless
facts) at all. We discuss factless fact tables in Example one: Hospital and
insurance revenue example and pseudo facts in “Example two: Employee
attendance recording” on page 209.

Assume an Internet Service Provider is tracking Web sites and their Web pages,
and all the visitor mouse clicks. We consider each of these examples:

Example one: Hospital and insurance revenue example
Consider an example of a hospital which is very busy with patients for the entire
day. Figure 5-76 on page 207 shows a simple star schema that can be used to
track the daily revenue of the hospital. The grain of the star schema is a single
patient visiting the hospital in a single day. The following are the dimension tables
in the star schema:
Insurance dimension: Contains one row for each insurance the patient has
Date dimension: The data at the daily level
Hospital dimension: A single row for each hospital
Customer dimension: One row for each customer

The insurance dimension consists of a Not applicable row or the Does not have
insurance row. We discussed the concept of Not applicable rows that may be
present inside a dimension table in “Insert a special customer row for the “Not
applicable” scenario” on page 274.

If a patient visits a hospital but does not have insurance, then the row in the
insurance dimension will be the Does not have insurance row. However, if a
patient visits a hospital and has insurance, then the insurance row specifying the
type of insurance is attached to the fact table row.

What this means is that there may be several patients that have insurance, but
they are visible inside the fact table only when they visit the hospital. Once they
visit the hospital, a row appears for these patients inside the fact table
HOSPITAL_DAILY_REVENUE.

So if we have one million customers who have purchased insurance in a month
and only 1 000 of these visit the hospital, then we would not be able to tell which
of the 999 000 customers had insurance, but did not visit the hospital.

From the star schema shown in Figure 5-76 on page 207, we can know about a
customer (patient) having insurance only after the customer visits the hospital
and appears in the fact table. This is primarily because the
HOSPITAL_DAILY_REVENUE fact table records sales-related activity. The


customer appears in this sales table if, and only if, the customer visits the
hospital and pays for the treatment received. If a customer has insurance, but
does not visit any hospital, we cannot tell whether or not they have insurance just
by looking at the schema shown in Figure 5-76. This is also because of the fact
that the Hospital and Insurance are independent dimensions only linked by the
HOSPITAL_DAILY_REVENUE fact table.

Grain is a Patient visiting a Hospital a single day

Figure 5-76 Hospital revenue generating schema

The hospital schema shown in Figure 5-76 can only tell if a customer has
insurance if the customer visits the hospital.

How to find customers who have insurance but who never came to the
hospital.

To find out about customers who have insurance, but who have not visited the
hospital, there are two options:
Insert a single row for all customers in the fact table (see Figure 5-76) on a
daily basis, irrespective of whether or not they visit the hospital. If the
customer does not visit the hospital, then have the fact AMOUNT_CHARGED
show a value of “0”. However, we do not recommend this option of inserting a
row for each customer on a daily basis irrespective of whether the customer
visited the hospital, because the fact table will grow at a very fast rate and
result in performance issues.


Create the event fact table star schema as shown in Figure 5-77 on page 208,
which consists of the following tables:
– Date dimension: Contains the date on which the person is insured.
– Insurance dimension: One row for each type of insurance.
– Customer dimension: A single row for each customer.
– Insurance_Event _Fact table: An event-based factless fact table. It does
not record anything other than the fact that a customer is insured. A
customer getting insured is treated as an Event and so a row is inserted
into the INSURANCE_EVENT_FACT table.

Note: The INSURANCE_EVENT_FACT fact table is a factless fact table. This
factless fact table contains only foreign keys and has no facts. It is a factless
fact table that represents the many-to-many relationship between the various
dimensions such as date, customer, and insurance.

Finding customers who have insurance but never visited a hospital:

To find out about customers who have insurance but never visited any hospital,
we do the following:
1. From Figure 5-76 on page 207, find all customers who visited the hospital.
2. From Figure 5-77, find all customers who have insurance.
3. Find the set difference from step 1 and 2. This would tell you which customers
had insurance but did not visit the hospital.

Grain: 1 row inserted into fact table for each insurance given to
a customer. (This is treated as an insurance giving event.)

Figure 5-77 Insurance event-based star schema


Note: The factless table shown in Figure 5-77 is used in conjunction with an
ordinary sales fact table (see Figure 5-76) to answer the question which
customers had insurance but did not go to the hospital?

Example two: Employee attendance recording
In this section we discuss an event fact table consisting of pseudo facts. The
event-based pseudo fact table is designed for recording the attendance of all
employees working for a company. Important points to consider while designing
the attendance recording system for the company are:
Attendance needs to be recorded for employees in the company.
Each employee belongs to a department.
A department may or may not have employees.
Attendance is recorded on a daily basis. The time of the employees attending
the office is also mandatory.
An employee may or may not have a manager. Some employees do not have
managers because they are their own managers.
The reporting manager of the employee also needs to be tracked.

For the requirements stated above, we designed a star schema as shown in
Figure 5-78 on page 210. The grain of the attendance tracking schema is
attendance of a single employee every day. The star schema consists of the
following dimensions:
Date dimension: Consists of date data at the day level.
Time dimension: Contains the time of day and the description of the event,
such as On Time, Late, Early Shift, Afternoon Shift, Night Shift, or Absent.
Employee dimension: Data about the employee.
Department: Department number to which an employee belongs. This table
also contains a Belongs to No Department row for employees who have no
department.
Reporting Manager: Contains all the managers information. It also contains a
row called No Manager Assigned Yet and another row called Self-Managed.
These special rows are for people who have yet not been assigned managers
or for people who are their own managers.
Project dimension: All the projects in the company. This table also contains a
special row called No Project Assigned Yet. This is for employees who have
not been assigned any project, or are into administrative work.


Client dimension: One row for each company client. This table also contains a
special row called Employee does not work for Client, for employees who
have no clients as of yet.

The star schema shown in Figure 5-78 contains the
EMPLOYEE_DAILY_ATTENDANCE fact table. The grain of this fact table is one
row per employee attendance during each day. One row is inserted into the
EMPLOYEE_DAILY_ATTENDANCE fact table each time an employee swipes
his badge while entering the company. The fact, called ATTENDANCE_COUNT,
gets a value of 1 when the attendance is recorded. For employees who did not
attend during a given day, a row is inserted with ATTENDANCE_COUNT equal
to 0.

Grain: Attendance of an employee on a daily basis

Figure 5-78 Employee attendance recording

To count the total employee attendance for a month, query on the
EMPLOYEE_DAILY_ATTENDANCE fact table and restrict data for an employee
with ATTENDANCE_COUNT=1. We use the COUNT() function to count the
employee attendance. We could also use the SUM() function and get the same
result.

Similarly if we want to find the number of times an employee has been absent in
a year, we query the EMPLOYEE_DAILY_ATTENDANCE fact table, restrict the
data for the employee for a particular year, and set ATTENDANCE_COUNT=0.
We use the COUNT() function instead of SUM because the SUM of all
ATTENDANCE_COUNT=0 will give us 0 where the COUNT of


ATTENDANCE_COUNT=0 will give the number of times an employee was
absent.

Why is the ATTENDANCE_COUNT a pseudo fact?

It is important to understand that ATTENDANCE_COUNT is a pseudo fact.
Instead of using a “1” for employee presence, we could have used any other
number or even a character like “Y”. Similarly for tracking the absence of
employees, we could have set ATTENDANCE_COUNT equal to a character
such as “N” or alternatively have used any number other than “0”.

Assume that for tracking employees present on a daily basis, we set
ATTENDANCE_COUNT=’10’. In this case we cannot SUM the attendance for a
month (30 days x 10) and say that the attendance is 300. However, if we use
COUNT, we get a correct number 30, assuming the employee worked for 30
days.

Similarly, assume that for tracking employee absence on a daily basis, we set
ATTENDANCE_COUNT=’9’, which signifies that the employee is ABSENT. In
this case we cannot SUM the attendance for a month (30 days x 9) and say that
the absence is 270 times. However, if we use COUNT, we get a correct number
30, assuming the employee was absent for 30 days.

(c) Internet Web page click event tracking
Figure 5-79 on page 212 shows a factless fact table star schema to track the
Web sites and each of the pages that has been visited. The grain is a single click
to a Web page. It is important to understand that the fact table named
WEB_PAGE_CLICK_FACT is an event-based table and does not have any facts.
The WEB_PAGE_CLICK_FACT fact table is used to represent a Web page click
event. The fact table captures many-to-many relationships between the
dimension tables and does not contain any facts. To calculate the total number of
Web page clicks per day, we can use the COUNT function on all the combination
of foreign keys.


Grain: A single click on a Web page
Figure 5-79 Web page clicking event star schema

5.5 Physical design considerations
In this section, we discuss the following topics:
DB2 Optimizer and MQTs for aggregate navigation
Indexing for dimension and fact tables

5.5.1 DB2 Optimizer and MQTs for aggregate navigation
DB2 Cube Views™ uses DB2 summary tables to improve the performance of
queries issued to cube models and cubes. A summary table is a special type of a
materialized query table (MQT) that specifically includes summary data.

You can complete expensive calculations and joins for queries ahead of time and
store that data in a summary table. When you run queries that can use the
precomputed data, DB2 UDB will reroute the queries to the summary table, even
if the query does not exactly match the precomputed calculations. By using
simple analytics such as SUM and COUNT, DB2 UDB can dynamically
aggregate the results from the precomputed data, and many different queries
can be satisfied by one summary table. Using summary tables can dramatically


improve query performance for queries that access commonly used data or that
involve aggregated data over one or more dimensions or tables.

Figure 5-80 shows a sample dimensional model based on a snowflake schema
with a Sales facts table, and Time, Market, and Product dimensions. The fact
table has measures and attributes keys, and each dimension has a set of
attributes and is joined to the facts object by a facts-to-dimension join.

Figure 5-80 Sales dimensional model

The hierarchy for each dimension in the dimensional model is shown in
Figure 5-81 on page 214. The boxes connected by the thick dark lines across the
hierarchies represent the data that actually exists in the base tables. Sales data
is stored at the Day level, Store level, and Product level. Data above the base
level in the hierarchy must be aggregated. If you query a base table for sales
data from a particular month, DB2 UDB dynamically adds the daily sales data to
return the monthly sales figures.


Figure 5-81 Hierarchy showing product, market, and time

The sample query in Example 5-12 shows the sales data for each product line, in
each region, by each month in 2004.

Example 5-12 Sample SQL showing sales by product line, region, and month
SELECT LINE_ID, REGION_NAME, MONTH_NUMBER, SUM(SALES)
FROM TIME, STORE, LOCATION, PRODUCT, LINE, SALESFACT
WHERE SALESFACT.STOREID = STORE.STOREID
AND STORE.POSTALCODEID = LOCATION.POSTALCODEID
AND SALESFACT.PRODUCTID = PRODUCT.PRODUCTID
AND PRODUCT.LINEID = LINE.LINEID
AND SALESFACT.TIMEID = TIME.TIMEID
AND YEAR = '2004'
GROUP BY LINEID, MONTH_NUMBER;

The line connecting Line-Region-Month in Figure 5-81 represents the slice that
the query accesses. It is a slice of the dimensional model and includes one level
from each hierarchy. You can define summary tables to satisfy queries at or
above a particular slice. A summary table can be built for the Line-Region-Month
slice that is accessed by the query. Any other queries that access data at or
above that slice including All Time, Year, Quarter, All Markets, All Products, and
Family can be satisfied by the summary table with some additional aggregating.
However, if you query more detailed data below the slice, such as Day or City,
the summary table cannot be used for this more granular query.

In Figure 5-82 on page 215, a dotted line defines the Line-State-Month slice. A
summary table built for the Line-State-Month slice can satisfy any query that
accesses data at or above the slice. All of the data that can be satisfied by a


summary table built for the Line-State-Month slice is included in the set of boxes
inside the “dashed line” area.

Figure 5-82 Summary table slice

The rewriter in the DB2 SQL compiler knows about existing summary tables, and
can automatically rewrite queries to read from the summary table instead of the
base tables. Rewritten queries are typically much faster because the summary
tables are usually much smaller than the base tables and contain preaggregated
data. Queries continue to be written against the base tables. DB2 UDB decides
when to use a summary table for a particular query and will rewrite the query to
access the summary tables instead, as shown in Figure 5-83 on page 216. The
rewritten query accesses a summary table that contains preaggregated data. A
summary table is often significantly smaller, and therefore significantly faster,
than the base tables and returns the same results as the base tables.

You can use the DB2 EXPLAIN facility to see if the query was rerouted, and if
applicable, to which table it was rerouted.


Figure 5-83 Query Rewrite using DB2 Optimizer

The query to see the sales data for each product line, in each region, by each
month in 2004, can be rewritten to use the summary table built for the
Line-Region-Month slice. The original query is shown in Example 5-13.

Example 5-13 Sample SQL to show original query
FROM TIME, STORE, LOCATION, PRODUCT, LINE, SALESFACT
WHERE SALESFACT.STOREID = STORE.STOREID
AND STORE.POSTALCODEID = LOCATION.POSTALCODEID
AND SALESFACT.PRODUCTID = PRODUCT.PRODUCTID
AND PRODUCT.LINEID = LINE.LINEID
AND SALESFACT.TIMEID = TIME.TIMEID
AND YEAR = '2004'
GROUP BY LINEID, MONTH_NUMBER;

The rewritten query is shown in Example 5-14.

Example 5-14 SQL to show rewritten query
FROM SUMMARYTABLE1
WHERE YEAR = '2004'


GROUP BY LINE_ID, REGION_NAME, MONTH_NUMBER;

The rewritten query is much simpler and quicker for DB2 UDB to complete
because the data is preaggregated and many of the table joins are precomputed
so DB2 UDB accesses one small table instead of six tables, including a large fact
table. The savings with summary tables can be tremendous, especially for
schemas that have large fact tables. For example, a fact table with 1 billion rows
might be preaggregated to a summary table with only 1 million rows, and the
calculations involved in this aggregation occur only once instead of each time a
query is issued. A summary table that is 1000 times smaller is much faster than
accessing the large base tables.

In this example, Figure 5-84 shows the summary table for the Line-State-Month
slice. DB2 UDB needs to calculate data for Region from the higher level State
instead of from the lower level Store, so the summary table has fewer rows than
the base tables because there are fewer states than stores. DB2 UDB does not
need to perform any additional calculations to return sales data by Month and
Line because the data is already aggregated at these levels. This query is
satisfied entirely by the data in the summary table that joins the tables used in
the query ahead of time and the joins do not need to be performed at the time the
query is issued. For more complex queries, the performance gains can be
dramatic.

Figure 5-84 Summary table created for line-region-month slice

In some cases, a query might access an attribute that is related to an attribute
that is included in the summary table. The DB2 optimizer can use functional
dependencies and constraints to dynamically join the summary table with the
appropriate dimension table.

When the Optimization Advisor recommends a summary table, all of the
measures in the dimensional model are included. In this example, the


SalesFacts object has only five measures including sales, cost of goods,
advertising, total expense, and profit, which are all included in the summary
table. If you define fifty measures for your dimensional model, all fifty measures
are included in the summary table. The Optimization Advisor does not need to
include all of the related attributes that are defined for a level in the summary
table because DB2 Cube Views defines functional dependencies between the
attributes in a level.

5.5.2 Indexing for dimension and fact tables
In this section we show how to design indexes for one particular star schema.
The output of our activity is data definition language (DDL).

We also show the different access paths the RDBMS optimizer uses for indexed
and non-indexed star schemas.

We have a star schema with three dimensions (date, customer, and product) and
a fact table called SALES. The fact table has three foreign keys. The three foreign
keys together with the degenerate dimension INVOICE_NO (Invoice Number)
forms the primary key of fact table. This is shown in Figure 5-85 on page 219.

The following are the hierarchies involved with the dimensions:
The hierarchy of the PRODUCT dimension is Supplier → Beverage group →
Beverage → Name.
The hierarchy for the CUSTOMER dimension is Region → Country → State
City → Name.
The hierarchy with the DATE dimension is Year → Quarter →
Month_of_year → Day.


Figure 5-85 Star Schema

First of all we design indexes for dimensions tables
Primary key (PK), if not generated by RDBMS
Unique indexes for each dimension hierarchy supporting joins (create them if
they are not generated by RDBMS automatically):

CREATE INDEX CUSTOMER_PK_IDX ON CUSTOMER (CUSTOMER_KEY)
CREATE INDEX PRODUCT_PK_IDX ON PRODUCT(PRODUCT_KEY)
CREATE INDEX DATE_PK_IDX ON DATE (DATE_KEY)
Non-unique indexes for each dimension's hierarchy level:

CREATE INDEX CUSTOMER_NA_IDX ON CUSTOMER (NAME,CUSTOMER_KEY)
CREATE INDEX CUSTOMER_TY_IDX ON CUSTOMER (TYPE,CUSTOMER_KEY)
CREATE INDEX CUSTOMER_CI_IDX ON CUSTOMER (CITY,CUSTOMER_KEY)
CREATE INDEX CUSTOMER_ST_IDX ON CUSTOMER (STATE,CUSTOMER_KEY)
CREATE INDEX CUSTOMER_CO_IDX ON CUSTOMER (COUNTRY,CUSTOMER_KEY)
CREATE INDEX CUSTOMER_RE_IDX ON CUSTOMER (REGION,CUSTOMER_KEY)

CREATE INDEX DATE_D_IDX ON DATE (DAY_AD,DATE_KEY)
CREATE INDEX DATE_M_IDX ON DATE (MONTH_OF_YEAR,DATE_KEY)
CREATE INDEX DATE_Q_IDX ON DATE (QUARTER,DATE_KEY)
CREATE INDEX DATE_Y_IDX ON DATE (YEAR_AD,DATE_KEY)

CREATE INDEX PRODUCT_NA_IDX ON PRODUCT (NAME,PRODUCT_KEY)


CREATE INDEX PRODUCT_BE_IDX ON PRODUCT (BEVERAGE,PRODUCT_KEY)
CREATE INDEX PRODUCT_BG_IDX ON PRODUCT (BEVERAGE_GROUP,PRODUCT_KEY)
CREATE INDEX PRODUCT_SU_IDX ON PRODUCT (SUPPLIER,PRODUCT_KEY)

Note: Including the PK of the dimensional table in the index eliminates
additional fetching of the PK values from the dimension table.

Non-unique Index for the product attribute product_id

CREATE INDEX PRODUCT_PID_IDX ON PRODUCT (PRODUCT_ID,PRODUCT_KEY)
Optionally create indexes for selected dimension hierarchy:

CREATE INDEX PRODUCT_H1_IDX ON PRODUCT (BEVERAGE,NAME,PRODUCT_KEY)

Note: Including the dimension hierarchy in an index eliminates additional
fetching of attributes from dimension tables. This approach is not typically
applicable, because the index tables can be quite large.

Index for fact table foreign keys
CREATE INDEX SALES_FK_PROD_IDX ON SALES (PRODUCT_KEY)
CREATE INDEX SALES_FK_CUST_IDX ON SALES (CUSTOMER_KEY)
CREATE INDEX SALES_FK_DATE_IDX ON SALES (DATE_KEY)

Note: After each change in the physical database you should update statistics
on tables and indexes. Following is the DB2 command for this action:

RUNSTATS ON TABLE TEST.SALES ON ALL COLUMNS WITH DISTRIBUTION ON ALL
COLUMNS AND INDEXES ALL

RUNSTATS ON TABLE TEST.PRODUCT ON ALL COLUMNS WITH DISTRIBUTION ON
ALL COLUMNS AND INDEXES ALL

RUNSTATS ON TABLE TEST.CUSTOMER ON ALL COLUMNS WITH DISTRIBUTION ON
ALL COLUMNS AND INDEXES ALL

RUNSTATS ON TABLE TEST.DATE ON ALL COLUMNS WITH DISTRIBUTION ON ALL
COLUMNS AND INDEXES ALL

Differences between indexed and non-indexed star schemas
In the following section we show the differences between the access paths the
RDBMS optimizer chooses for indexed and non-indexed star schemas for Dicing


and Slicing. Generally if an index is not created, the RDBMS must read all tables
in one continuous scan.

Access plans for OLAP activity
In this section we show the access plans of the query optimizer, and which
indexes are used. We also show the same execution times for OLAP Slice and
Dice query.

For DICE

In Figure 5-86 we show a Dice example for beverage group and region.

Whiskey Whiskey Stan Long

Figure 5-86 Beverage group: Beverage: Product x Region

The figures and labels in Figure 5-86 were obtained by the SQL statement shown
in Example 5-15.

Example 5-15 Select statement for Dice
select P.BEVERAGE_GROUP,P.BEVERAGE,C.REGION,
sum(S.SALES_AMOUNT) as SALES_AMOUNT
fromTEST.SALES S
joinTEST.PRODUCTP
on (S.PRODUCT_KEY = P.PRODUCT_KEY)
joinTEST.CUSTOMER C
on (S.CUSTOMER_KEY = C.CUSTOMER_KEY)
group byP.BEVERAGE_GROUP,
P.BEVERAGE,
C.REGION


Figure 5-87 shows the access plan of the optimizer for the SQL Dice example.
The optimizer decided to read entire tables in one scan, and then used a hash
join rather than fetching rows by index and joining them with the other table rows
in a loop.

RETURN 27 timerons

GRPBY

TBSCAN

SORT

HSJOIN

TBSCAN HSJOIN

TEST.CUSTOMER TBSCAN TBSCAN

TEST.SALES TEST.PRODUCT

Figure 5-87 Access plan for Dice

For Slice
In Figure 5-88, we show a slice example for beverage Wine Oak.

Figure 5-88 Slice for Product dimension

The SQL query for the slice where BEVERAGE_GROUP='Wine Oak', is shown in
Example 5-16 on page 223.


Example 5-16 Select statement for Dice

select S.BEVERAGE, NAME, sum(P.SALES_AMOUNT) as SALES_AMOUNT
from TEST.SALES P
join TEST.PRODUCT S
on (P.PRODUCT_KEY = S.PRODUCT_KEY)
whereS.BEVERAGE in ('Wine Oak')
group by S.BEVERAGE, NAME

In Figure 5-89 on page 224 we show the access plan for the star schema without
indexes, with indexes, and with an index (Product_H1_IDX) on a the following
attributes (BEVERAGE,NAME,PRODUCT_KEY). We get the following results:
Without indexes, the optimizer reads the data with table scans.
With PK and FK indexes, the optimizer use the Indexes to fetch and join
Product and Sales tables. Therefore, reading the entire fact table was
eliminated.
With Index carrying the Product dimension hierarchy, the access path is
simpler because it eliminates a few tasks on the product table. However, this
does not significantly decrease execution time.

The estimated processing time for fact table (1000000 rows) and Product
dimension in our scenario is listed in Table 5-24. Results show that the indexed
star schema for our Slice example performs best.

Table 5-24 Estimated and measured processing times
Index Estimated
[timerons*]

NO 27

PK, FK, and dimension 10

PK, FK dimension, and 9
Product hierarchy

* Timerons are a hybrid value made up of the estimated CPU, elapsed time, I/O,
and buffer pool consumption that DB2 expects the SQL will consume when using
a particular access path.

Note: The estimated time given by the RDBMS tool is a good guide for
designing indexes, but only repeatable measurement of real execution (10
times) gives reliable average values. We used the DB2 Command Center
access plan utility to get estimates and access plans, and DB2batch to
measure execution values.


RETURN
RETURN
RETURN

GRPBY

GRPBY(3)
GRPBY

NLJOIN

TBSCAN(5) NLJOIN

TBSCAN FETCH 0.1

IXSCAN FETCH
SORT SORT IXSCAN TEST SALES

FETCH(12) 0.1 SALES_FK_PROD_IDX PRODUCT_H1_IDX IXSCAN TEST SALES

HSJOIN

IXSCAN TEST.PRODUCT TEST SALES

TEST.PRODUCT SALES_FK_PROD_IDX
TBSCAN TBSCAN

PRODUCT_BE_IDX

TEST SALES

TEST.SALES TEST.PRODUCT
TEST.PRODUCT

Without Index With Index With Index Product_H1_IDT

Figure 5-89 Access plan for slice

5.6 Handling changes
In this section we discuss how to deal with changes. The dimensional model
should be designed to handle changes to the:
Data
Structure
Business requirements

5.6.1 Changes to data
To maintain history in the dimensional model, we have different change handling
strategies. The strategies depend upon whether the dimension changes slowly or
changes very fast. The following list describes different approaches for handling
changes to various types of dimensions:
Slowly changing dimensions: Slowly changing dimensions change slowly
over a period of time. The different approaches to handle slowly changing
dimensions are:
– Type-1: Overwrite the existing value.
– Type-2: Insert a new row.
– Type-3: Add one or more columns in the dimension to store changes.
We discussed slowly changing dimensions in detail in 5.3.5, “Slowly changing


Very fast changing dimensions: Fast changing dimensions are dimensions
whose attributes change very rapidly over a period of time. The approach to
handling fast changing dimension is to split the rapidly changing dimension
into one or more mini-dimensions. We discussed fast changing dimensions in
detail in “Handling fast changing dimensions” on page 163.

5.6.2 Changes to structure
Changes to the structure of a dimensional model may occur because of any of
the following reasons:
Addition of a new dimension to the star schema: This may occur when the
business needs to view existing data across a new dimension are not present
in the model. The new dimension can be easily added to the model if it does
not violate the grain of the fact table. This is one of the primary reasons we
recommend to design the model at the lowest granularity. Then additional
changes to business requirements in the form of new dimensions can be
added easily to the existing structure. Assume that you add a new dimension
to an existing model on 2/2/2005. Whenever you do so, you must be sure to
include an informational row containing information, such as No Data for
prior to Date 2/2/2005, inside this new dimension. All previous rows in the
fact table are then linked to this row for all rows inserted in the fact table
before the new dimension was made available.
Addition of a new attribute to a dimension: A new attribute can be added
to any existing dimension provided the new attribute is true to the grain. If the
attribute takes on a single value in the context of the fact measurements, then
the new attribute can be added to the existing dimensional model. Assume
that you add a new attribute to an existing model on 2/2/2005. When you do
so, you must be sure to include an informational value, such as Not
Applicable prior to Date 2/2/2005, inside this new attribute for all the old
dimension rows for which new dimension attribute is not applicable and has
no value.
Addition of a new fact to the fact table: A new fact can be added to the
existing fact table provided the new fact is true to the grain. If the fact is valid
for the existing grain definition, then it can be added to the existing star
schema. Assume that you add a new fact to an existing fact table on 2/2/2005.
Whenever you do so, you must be sure to include an informational value such
as 0 inside this new fact for all the old fact rows for which new fact is not
applicable and has no value.
Granularity of the dimensional model: The change in the grain definition is
also a cause for the structure of the dimensional model to change. This
occurs primarily when the existing dimensional model was designed at a
lower grain, thereby missing certain dimensions and facts which were
available at a lower grain definition. If the business at a later time requires new


dimensions to be made available at an existing lower detailed grain model,
this is generally possible only if the entire model is redesigned by defining a
more detailed atomic grain.
For more information on this topic, refer to 5.2.2, “Importance of detailed
atomic grain” on page 122.

5.6.3 Changes to business requirements
Dimensional models, if designed at the most atomic detailed grain, can
accommodate a change in business requirements very easily. If the grain is
defined carefully, then the dimensional model can withstand business
requirement changes easily without any change in the front-end applications.
The change is typically transparent to applications.

Typically, any change to business requirements may lead to one or more of the
following:
Addition of a new dimension
Addition of a new dimensional attribute
Addition of a new fact
Change in the granularity of the fact table or the entire dimensional model


6

Chapter 6. Dimensional Model Design
Life Cycle
In this chapter, we discuss the activities involved in building a dimensional
model. This can be a very tedious process because requirements are typically
difficult to define. Many times it is only after seeing a result that you can decide
that it does, or does not, satisfy a requirement. And, the requirements of an
organization change over time. What is valid one day may no longer be valid the
next. Regardless, the requirements identified at this point in the development
cycle are used to build the dimensional model.

But, where do you start? What do you do first? To help in that decision process,
we have developed a dimensional model design life cycle (DMDL) which consists
of the following phases:
Identify business process requirements
Identify the grain
Identify the dimensions
Identify the facts
Verify the model
Physical design considerations
Meta data management

The following sections discuss and describe these phases of the DMDL.


6.1 The structure and phases
The DMDL can help identify the phases and activities that you need to consider
in the design of a dimensional model. It is depicted in Figure 6-1. We have used
this methodology throughout the book as we work with dimensional models.
More specifically, we have used it to help in the example case study documented
in Chapter 7, “Case Study: Dimensional model development” on page 333.

Determine All
Dimensions
Processes

Identify Conformed




Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


We now describe the phases of the DMDL so you will be familiar with it and
better enable its use as you design your next dimensional model:
Identify business process requirements: Involves selection of the business
process for which the dimensional model will be designed. Based on the
selection, the requirements for the business process are gathered. Selecting
a single business process, out of all those that exist in a company, often
requires prioritizing the business processes according to criteria, such as
business process significance, quality of data in the source systems, and the
feasibility and complexity of the business processes.


Identify the grain: Next, we must identify the grain definition for the business
process. If more than one grain definition exists for a single business process,
then we must design separate fact tables. Avoid forcefully fitting multiple grain
definitions into the same fact table. We also need to make sure that we design
the grain at the most atomic level of detail so that we can extend the
dimensional model to meet future business requirements. In other words, we
are able to add new facts and dimensions to the existing model with little or no
change to the front-end applications, or any major rework to the existing
model.
Identify the dimensions: Here we identify the dimensions that are valid for
the grain chosen in the previous step.
Identify the facts: Now we identify the facts that are valid for the grain
definition we chose.
Verify model: Before continuing, we must verify that the dimensional model
can meet the business requirements. Sometimes it may be required to revisit,
and perhaps change, the definition of the grain to assure we can meet the
requirements.
Physical design considerations: Now that the model has been designed,
we can focus on other considerations, such as performance. It may require
tuning by taking actions such as data placement, partitioning, indexing,
partitioning, and creating aggregates.

We describe the DMDL by using a retail sales business process. However, there
may be concepts of the DMDL that are not applicable to the retail sales business
process. For completeness, we cover those concepts in Chapter 5, “Modeling
considerations” on page 103.

6.2 Identify business process requirements
During this phase, we identify the business process for which the dimensional
model will be designed. However, be aware that a business process may require
more than one dimensional model.

In dimensional modeling, the best unit of analysis is the business process in
which the organization has the most interest. A business process is basically a
set of related activities. Business processes are roughly classified by the topics
of interest to the business. To extract a candidate list of high potential business
processes necessitates prioritization of requirements. Examples of business
processes are customers, profit, sales, organizations, and products.

To help in determining the business processes, use a technique that has been
successful for many organizations. Namely, the 5W-1H rule. First determine the

Chapter 6. Dimensional Model Design Life Cycle 229

when, where, who, what, why, and how of your business interests. For example,
to answer the who question, your business interests may be in customer,
employee, manager, supplier, business partner, and/or competitor.

Note: When we refer to a business process, we are not simply referring to a
business department. For example, consider a scenario where the sales and
marketing department access the orders data. We build a single dimensional
model to handle orders data rather than building separate dimensional models
for the sales and marketing departments. Creating dimensional models based
on departments would no doubt result in duplicate data. This duplication, or
data redundancy, can result in many data quality and data consistency issues.

Before beginning, recall the various architectures used for data warehouse
design we discussed in 3.3, “Data warehouse architecture choices” on page 57.
They were:
Business-wide enterprise data warehouse
Independent data warehouse
Dependent data warehouse

Create a dimensional model for a business process directly from OLTP source
systems, as in the case of independent and dependent data warehouse design
architectures. Or create a dimensional model for a business process from the
enterprise data warehouse, as in the case of the business-wide enterprise data
warehouse architecture.

Figure 6-2 on page 231 shows that the source for a dimensional model can be:
An enterprise wide data warehouse
OLTP source systems (in the case of independent or dependent data mart
architectures)
Independent data marts (in this situation, you might be interested in
consolidating the independent data marts into another data mart or data
warehouse)

Note: For more information on data mart consolidation, refer to the IBM
Redbook, Data Mart Consolidation: Getting Control of Your Enterprise
Information, SG24-6653.


EDW

Dimensional
Model

OLTP Source Systems

Independent Data Marts

Data Mart
Consolidation

Figure 6-2 Dimensional model sources

Data warehouse and the dimensional model
When you consider the partitioning of the data in a data warehouse, the most
common criterion is subject area. As you may remember, a data warehouse is
subject-oriented. It is oriented to specific selected subject areas in the
organization, such as customer and product. For a practical implementation of a
data warehouse, we suggest that the unit of measure is the business
process.This is quite different from partitioning in the operational environment.

OLTP systems and the dimensional model
In the operational environment, partitioning is more typically by application or
function because the operational environment has been built around
transaction-oriented applications that perform a specific set of functions. And,
typically, the objective is to perform those functions as quickly as possible. If
there are queries performed in the operational environment, they are more
tactical in nature and are to answer questions concerned with that instant in time.
An example is, “Has the check from Mr. Smith been processed?” Queries in the
data warehouse environment are more strategic in nature and intend to ask
questions concerned with a larger scope. An example of a query is, “What
products are selling well?” or “Where are my weakest sales offices?” To answer
those queries, the data warehouse is structured and oriented to subject areas
such as product or organization. These subject areas are the most common unit
of logical partitioning in the data warehouse.


Figure 6-3 depicts an E/R model (which can be an OLTP or an enterprise data
warehouse) that consists of several business processes.

Business
Process #1

Business Business
Process #3 Process #2

Figure 6-3 E/R model consists of several business processes

6.2.1 Create and Study the enterprise business process list
During this activity, we create a complete enterprise-wide business process list.
A sample business process list is shown in Table 6-1 on page 233, along with a
number of other assessment factors, such as:
Complexity of the source systems of each business process
Data availability of these systems
Data quality of these systems
Strategic business significance of each business process

Table 6-1 on page 233 also shows the value points along with the assessment
factors involved. For example, for the business process Finance, the Complexity
= High(1). The 1 here is the assigned value point. The value points for each
assessment factor are listed in detail in Table 6-2 on page 233.


Table 6-1 Enterprise business process list
Name of business Complexity Data Data quality Strategic
process availability business
significance

Retail sales Low (3) High (3) High (3) High (6)

Finance High (1) High (3) Medium (2) Medium (4)

Servicing Low (3) High (3) Medium (2) High (6)

Marketing Medium (2) Medium (2) Medium (2) Medium (4)

Shipment Low (3) Low (1) High (3) Low (2)

Supply Medium (2) Low (1) Medium (2) Low (2)
management

Purchase orders High (1) Medium (1) Low (1) Medium (4)

Labor Low (3) Low (1) Low (1) High (2)

Table 6-2 shows the value points for each of the assessment factors:
Complexity
Data availability
Data quality
Strategic business significance

Table 6-2 Value points table for assessment factors
Assessment Factor Low Medium High

Complexity 3 2 1

Data availability 1 2 3

Data quality 1 2 3

Strategic business significance 2 4 6

Note: The value points, or weight, given for the assessment factor, Strategic
business significance, is more than the other factors. This is simply because
the company has decided that strategically significant business processes
should be given a higher value than the other assessment factors.

We recommend that you include the assessment factors you think are important
for your business processes and assign them with your evaluation of the


appropriate points or weight. For example purposes, we have included only four
factors in Table 6-2 on page 233.

This exercise will help you quickly prioritize the business processes for which
dimensional models should be built.

6.2.2 Identify business process
In this phase, we prioritize the business processes. The basic idea here is to
identify the most and least feasible processes for building a dimensional model.

Table 6-3 shows a business process list in descending order of priority. The
priority is based on the number of points in the Point sum column. It is a
summary of all the assessment factor points. Table 6-3 is created from the
enterprise-wide business process list shown in Table 6-1 on page 233.

For example, when we look at the finance business, the point sum column has a
value of 10 (1 + 3 + 2 + 4), which is the sum of all assessment factor points.

Table 6-3 Enterprise-wide business process priority listing
Name of Complexity Data Data Strategic Point
business availability quality business sum
process significance

Retail sales Low (3) High (3) High (3) High (6) 15

Finance High (1) High (3) Medium (2) Medium (4) 10

Servicing Low (3) High (3) Medium (2) High (6) 14

Marketing Medium (2) Medium (2) Medium (2) Medium (4) 10

Shipment Low (3) Low (1) High (3) Low (2) 9

Supply Mgmt Medium (2) Low (1) Medium (2) Low (2) 7

Purchase High (1) Medium (1) Low (1) Medium (4) 7
Order

Labor Low (3) Low (1) Low (1) High (2) 7

From Table 6-3, it is observed that the retail sales process gets the highest
priority. Therefore, it is also likely that this will be the business process for which
the first dimensional model (and data mart) will be built.

Therefore, it is the business process we have selected for the example case
study dimensional model in this chapter.


Note: Table 6-3 serves only as a guideline for identifying high priority and
feasible business processes in our example. You will go through a similar
methodology in prioritizing the business processes in your company.

6.2.3 Identify high level entities and measures for conformance
The next step is to determine the high level business entities involved in each
process. We depict this in Table 6-4. The idea is to determine which entities are
common across several business processes. Once identified, we use these
entities as common across all dimensional models (data marts) in the enterprise.
Each business process will then be tied together through these common
(conformed) dimensions.

To create conformed dimensions that are used across the enterprise, the various
businesses must agree on the definitions for these common entities. For
example, as shown in Table 6-4, the business processes, Retail sales, Finance,
Servicing, Marketing, Shipment, Supply management, and Purchase order,
should agree on a common definition of the Product entity, because all these
business processes have the product entity in common. The product entity then
becomes a conformed Product dimension which is shared across all the
business processes.

However, getting agreement on definitions of common entities, such as product
and customer, can be difficult because they typically vary from one business
process to another. And, therefore, changes will likely impact existing
applications.

Table 6-4 Business processes and high level entities
High Level Entities →

Shipment company
Received date
Selling date

Warehouse
Employee

And more
Customer

Supplier
Product

Ledger

Check

Store

Business processes .......

Retail sales X X X X X X

Finance X X

Servicing X X X X

Marketing X X X

Shipment X X X X X X X


High Level Entities →

Shipment company
Received date
Selling date

Warehouse
Employee
Customer

And more
Supplier
Product

Ledger

Check

Store

.......
Business processes

Supply management X X X X X

Purchase order X X X

Human resources X

What are conformed dimensions?
A data warehouse must provide consistent information for queries requesting
similar information. One method to maintain consistency is to create dimension
tables that are shared (and therefore conformed), and used by all applications
and data marts (dimensional models) in the data warehouse. Candidates for
shared or conformed dimensions include customers, time, products, and
geographical dimensions, such as the store dimension.

What are conformed facts?
Fact conformation means that if two facts exist in two separate locations, then
they must have the same name and definition. As examples, revenue and profit
are each facts that must be conformed. By conforming a fact, then all business
processes agree on one common definition for the revenue and profit measures.
Then, revenue and profit, even when taken from separate fact tables, can be
mathematically combined.

Establishing conformity
Developing a set of shared, conformed dimensions is a significant challenge. Any
dimensions that are common across the business processes must represent the
dimension information in the same way. That is, it must be conformed. Each
business process will typically have its own schema that contains a fact table,
several conforming dimension tables, and dimension tables unique to the
specific business function. The same is true for facts.

6.2.4 Identify data sources
In this activity, we identify the data sources involved with the business
processes. Table 6-5 on page 237 shows a sample listing of business processes
and their respective data sources, along with their owner, location, and platform.
Other factors can also be associated to describe and document the data source


in more detail. However, for simplicity of the example, we have chosen the name,
owner, location, and platform of the data source.

Table 6-5 Data sources for business process
Business Data Owner Location Platform
process sources

Retail sales Source 1 Order admin New Jersey DB2

Finance Source 2 Finance admin New York DB2

And more

6.2.5 Select requirements gathering approach
The traditional development cycle focuses on automating the process, making it
faster and more efficient. The dimensional model development cycle focuses on
facilitating the analysis that will change the process to make it more effective.
Efficiency measures how much effort is required to meet a goal. Effectiveness
measures how well a goal is being met against a set of expectations.

As previously mentioned, requirements are typically difficult to define. Often, it is
only after seeing a result that you can decide that it does, or does not, satisfy a
requirement. And, the requirements of an organization change over time. What is
valid one day may no longer be valid the next day. Regardless, we use the
requirements identified at this point in the development cycle to build the
dimensional model.

The question then is how can you build something that cannot be precisely
defined? And, how do you know when you have successfully identified the
requirements? Although there is no definitive test, we propose that if your
requirements address the following questions, then you probably have enough
information to begin modeling.

The questions relate to who, what, when, where and how
Who are the people, groups, and organizations of interest?
What functions need to be analyzed?
Why is the data required?
When does the data need to be recorded?
Where, geographically and organizationally, do relevant processes occur?
How do we measure the performance of the functions being analyzed?
How is performance of the business process measured? What factors
determine the success or failure?


What is the method of information distribution? Is it a data report, paper,
pager, or e-mail (examples)?
What types of information are lacking for analysis and decision making?
What steps are currently taken to fulfill the information gap?
What level of detail would enable data analysis?

There are likely many other methods for deriving business requirements.
However, in general, these methods can be placed in one of two categories:
Source-driven
User-driven

Figure 6-4 depicts an example of this methodology.

User Requirements

What we want

User
Driven
What can be
delivered and
will be useful
Source
Driven

What we have

Operational Data
Figure 6-4 Source-driven and User-driven requirements gathering

Source-driven requirements gathering
Source-driven requirements gathering, as the name implies, is a method based
on defining the requirements by using the source data in production operational
systems. This is done by analyzing an E/R model of source data if one is
available or the actual physical record layouts and selecting data elements
deemed to be of interest.

The major advantage of this approach is that you know from the beginning that
you can supply all the data, because you are already limiting yourself to what is
available. A second benefit is that you can minimize the time required by the
users in the early stages of the project. However, there is no substitute for the
importance and value you get when you get the users involved.


Of course there are also disadvantages to this approach. By minimizing user
involvement, you increase the risk of producing an incorrect set of requirements.
Depending on the volume of source data you have, and the availability of E/R
models for it, this can also be a very time-consuming approach. Perhaps most
important, some of the user’s key requirements may need data that is currently
unavailable. Without the opportunity to identify these requirements, there is no
chance to investigate what is involved in obtaining external data. External data is
data that exists outside the enterprise. Even so, external data can often be of
significant value to the business users.

The result of the source-driven approach is to provide what you have. We think
there are at least two cases where this is appropriate. First, relative to
dimensional modeling, it can be used to develop a fairly comprehensive list of the
major dimensions of interest to the enterprise. If you ultimately plan to have an
enterprise-wide data warehouse, this could minimize the proliferation of duplicate
dimensions across separately developed data marts. Second, analyzing
relationships in the source data can identify areas on which to focus your data
warehouse development efforts.

User-driven requirements gathering
User-driven requirements gathering is a method based on defining the
requirements by investigating the functions the users perform. This is usually
done through a series of meetings and/or interviews with users.

The major advantage to this approach is that the focus is on providing what is
really needed, rather than what is available. In general, this approach has a
smaller scope than the source-driven approach. Therefore, it generally produces
a useful data warehouse or a data mart in a shorter time span.

On the negative side, expectations must be closely managed. The users must
clearly understand that it is possible that some of the data they need can simply
not be made available for a variety of possible reasons. But, try not to limit the
things for which the user asks. Outside-the-box thinking should be promoted
when defining requirements for a data warehouse. This prevents you from
eliminating requirements simply because you think they might not be possible. If
a user is too tightly focused, it is possible to miss useful data that is available in
the production systems.

We believe user-driven requirements gathering is the approach of choice,
especially when developing dependent data marts or populating data marts from
a business-wide enterprise warehouse.


6.2.6 Requirements gathering
During requirements gathering, business users needs are collected and
documented. Requirements gathering focuses on the study of business
processes and information analysis activities with which users are involved. A
user typically needs to evaluate, or analyze, some aspect of the organization’s
business. It is extremely important that requirements gathering focus on the two
key elements of analysis that business users are involved with on a day-to-day
basis:
What is being analyzed?
The evaluation criteria

Requirements gathering, therefore, is extremely oriented toward understanding
the problem domain for which the modeling is done. Typically, requirements at
this stage are documented rather informally or, at least, they are not represented
in detailed schemas.

Assume that we are designing a dimensional model for a retail sales business
process. We identified the retail sales business process from section 6.2.2,
“Identify business process” on page 234. Figure 6-5 shows the E/R model for this
business process.

Figure 6-5 E/R Model for the retail sales business process


Now assume that we perform business requirements analysis for the retail sales
process. We identify that the business needs the answers to the questions
shown in Table 6-6.

Table 6-6 User requirements for the retail sales process
Seq. Business requirement Importance
no.

Q1 What is the average sales quantity this month for each Medium
product in each category?

Q2 Who are the top 10 sales representatives and who are their High
managers? What were their sales for the products they sold
in the first and last fiscal quarters?

Q3 Who are the bottom 20 sales representatives and who are High
their managers?

Q4 How much of each product did U.S. and European customers High
order, by quarter, in 2005?

Q5 What are the top five products sold last month by total High
revenue? By quantity sold? By total cost? Who was the
supplier for each of those products?

Q6 Which products and brands have not sold in the last week? High
The last month?

Q7 Which salespersons had no sales recorded last month for High
each of the products in each of the top five revenue
generating countries?

Q8 What was the sales quantity of each of the top five selling Medium
products on Christmas, Thanksgiving, Easter, Valentines
Day, and the Fourth of July?

Q9 What are the sales comparisons of all products sold on High
weekdays compared to weekends? Also, what was the sales
comparison for all Saturdays and Sundays every month?

Q10 What are the 10 top and bottom selling products each day Time,
and week? Also at what time of the day do these sell? Product
Assuming there are five broad time periods- Early morning
(2AM-6AM), Morning (6AM-12PM), Noon (12PM-4PM),
Evening (4PM-10PM), and Late night (10PM-2AM).

The column named Importance (Low/Medium/High or Critical) signifies the
importance of the questions. Some questions (business needs) may be highly
significant for the business where other questions may be less significant in
comparison.


In addition to the set of business requirements listed in Table 6-6 on page 241, it
is very important to understand how the business wants to preserve the history.
In other words, how does the business want to record the changed data. Some of
the questions (in regard to maintaining history) that you may ask are listed in
Table 6-7.

Table 6-7 Questions relating to maintaining history
Seq. Question How to maintain history
no.

1 What happens if an employee changes Overwrite history or maintain
from Region A to Region B? history

2 What happens if an employee changes Overwrite history or maintain
from Manager A to Manager B? history

3 What happens if a product changes from Overwrite history or maintain
existing Category A to Category B? history

4 What happens when a product is Overwrite history or maintain
discontinued? history

5 More questions Overwrite history or maintain
history

6.2.7 Requirements analysis
During requirements analysis, informal requirements (as gathered in section
6.2.6, “Requirements gathering” on page 240) are further investigated and high
level measures and high level entities (potential future dimensions) are produced.

Table 6-8 shows high level entities and measures identified from the
requirements stated in Table 6-6 on page 241 for the retail sales business
process. Note that the high level entities are potential future dimensions.

Table 6-8 Requirement Analysis - High level entities and measures
Seq. Business requirement High level entities Measures
no.

Q1 What is the average sales quantity this Month, Product Quantity of
month for each product in each units sold
category?

Q2 Who are the top 10 sales representatives Sales Revenue
and who are their managers? What were representative, sales
their sales in the first and last fiscal Manager, Fiscal
quarters for the products they sold? Quarter, Product


Seq. Business requirement High level entities Measures
no.

Q3 Who are the bottom 20 sales Sales Sales
representatives and who are their representative, Revenue
managers? Manager

Q4 How much of each product did U.S. and Product, Quantity of
European customers order, by quarter, in Customers, units sold
2005? Quarter, Year

Q5 What are the top five products sold last Product, Supplier Revenue
month by total revenue? By quantity sales,
sold? By total cost? Who was the Quantity of
supplier for each of those products? units sold,
Total Cost

Q6 Which products and brands have not Product, Brand, Sales
sold in the last week? The last month? Month Revenue

Q7 Which salespersons had no sales Salesperson, Sales
recorded last month for each of the Product, Customer Revenue
products in each of the top five revenue country, Month

Q8 What was the sales quantity of the top Holidays, Products Sales
five selling products on Christmas, revenue
Thanksgiving, Easter, Valentines Day
and the Fourth of July?

Q9 What are the sales comparisons of all Products, Sales
products sold on weekdays compared to Weekdays, revenue
weekends? Also, what was the sales Weekends
comparison for all Saturdays and
Sundays every month?

Q10 What are the 10 top and bottom selling Time, Product Sales
products each day and week? Also, at revenue
what time of the day do these sell?
Assuming there are five broad time
periods- Early morning (2AM-6AM),
Morning (6AM-12PM), Noon
(12PM-4PM), Evening (4PM-10PM) Late
night (10PM-2AM).

We listed each business requirement in Table 6-8 on page 242 and identified
some of the high-level entities and measures. Using this as a starting point, we
now filter down the entities and measures as shown in Table 6-9 on page 244.


Also, any hierarchies associated with each of these high-level entities (which
could be transformed to dimensions later) are documented.

Table 6-9 Final set of high level entities and measures
Entities (potential future dimensions) Hierarchy in entity (potential future
dimension)

Customer Country → Region
(Customer country, such as U.S. or
Europe)

Product Category → Brand
(Brand, category)

Selling Date Fiscal Year → Fiscal Quarter → Fiscal
(Holidays, weekends, weekdays, month, Month
quarter, year)

Supplier of product

Time of selling Hour → Minute

Employee or salesperson or sales
representative

Measures Revenue sales, Quantity of units sold,
(Key Performance Indicators) Total cost

Note: The high level dimensions and measures are useful in determining the
grain, dimensions, and facts during the different phases of the DMDL.

6.2.8 Business process analysis summary
The final output of the Identify business process phase is the requirements
gathering report, which contains the following:
Business process listing
Business process prioritization
High level entities and measures, which are common between various
business processes
Business process identified for which the dimensional model will be built
Data sources listing
Requirement gathering which contains all business process requirements


Requirement gathering analysis
High level entities and measures identified from the requirement analysis

6.3 Identify the grain
In the Identify the grain phase, we focus on the second step of the DMDL as

Determine All
Dimensions
Processes

Identify Conformed




Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


What is the grain?
The following are several characteristics of grain identification:
When identifying the grain, we must specify exactly what a fact table record
means. The grain conveys the level of detail associated with the fact table
measurements. Identifying the grain also means deciding on the level of detail
you want to be made available in the dimensional model. The more detail
there is, the lower the level of granularity. The less detail there is, the higher
the level of granularity.


The level of detail available in a star schema (dimensional model) is referred
to as the grain. Each fact and dimension table is said to have its own grain or
granularity. In other words, each table (either fact or dimension) will have
some level of detail associated with it. The grain of the dimensional model is
the finest level of detail implied by the joining of the fact and dimension tables.
For example, the granularity of a dimensional model consisting of
dimensions, date (year, quarter, month, and day), store (region, district, and
store) and product (category name, brand, and product) is product sold in
store by day.
Both the dimension and fact tables have a grain associated with them. To
understand the grain of a dimension table, we need to understand the
attributes of the dimension table. Every dimension has one or more attributes.
Each attribute associates a parent or child with other attributes. This
parent-child relationship provides different levels of summarization. The
lowest level of summarization or the highest level of detail is referred to as
the grain. The granularity of the dimension affects the design, and can impact
such things as the retrieval of data and data storage.
A grain refers to the level of detail present in each fact table. Each row should
hold the same type of data. For example, each row could contain daily sales
by store by product or daily line items by store.

Examples of grain definitions are:
A line item on a grocery receipt
A single item on an invoice
A single item on a restaurant bill
A line item on a bill received by a hospital
A monthly snapshot of a bank account statement
A weekly snapshot of the number of products in the warehouse inventory
A single airline ticket purchased on a day
A single bus ticket purchased on a day

What is Granularity?
The fact and dimension tables have a granularity associated with them. In
dimensional modeling, the term granularity refers to the level of detail stored in a
table.

For example, a dimension such as date (year, quarter) has a granularity at the
quarter level but does not have information for individual days or months. And, a
dimension such as date (year, quarter, month) table has granularity at the month
level, but does not contain information at the day level.


Note: Differing data granularities can be handled by using multiple fact tables
(daily, monthly, and yearly tables) or by modifying a single table so that a
granularity flag (a column to indicate whether the data is a daily, monthly, or
yearly amount) can be stored along with the data. However, we do not
recommend storing data with different granularities in the same fact table.

6.3.1 Fact table granularity
For our example retail sales business process from section 6.2, “Identify
business process requirements” on page 229, we identify the grain definition as
an individual line item on a grocery store bill. The grain detail is based on the
requirements findings that were analyzed and documented in section 6.2.7,
“Requirements analysis” on page 242.

The grain definition is represented graphically in Figure 6-7 on page 248. It is
important to understand that while gathering business requirements, you should
collect documents, such as invoices, receipts, and order memos. These often
have information which can be used to define the grain. Also such documents
have information which helps identify the dimensions and facts for the
dimensional models.

Note: The grain you choose determines the level of detailed information that
can be made available to the dimensional model. Choosing the right grain is
the most important step for designing the dimensional model.


Customer : Carlos

08/29/2005
1600 Hours

Quantity UP Dsc
Grain= 1 Line 1. Eggs 12 $3 $36
item on a Grocery 2. Dairy Milk 2 $2 $ 4
Bill 3. Chocolate Powder 1 $9 $ 9
4. Soda Lime 12 $ 1.5 $18
5. Bread 2 $4 $ 8

Employee: Amit $75

Figure 6-7 Grain example: A single line item on a grocery bill

Guidelines for choosing the fact table granularity

The grain definition is the base of every dimensional model. It determines the
level of information that is available. Guidelines for choosing the grain definition
are:
During the business requirements gathering phase, try to collect any
documents, such as invoice forms, order forms, and sales receipts. Typically,
you will see that these documents have transactional data associated with
them, such as order number and invoice number.
Documents can often point you to the important elements of the business,
such as customer and the products. They often contain information at the
lowest level that may be required by the business.
Another important point to consider is the date. Understand to what level of
detail a date is associated with a customer, product, or supplier. Is the
information in the source systems available at the day, month, or year level?
When developing the grain, decide whether you would like information to be
stored at a day, month, or year level.


6.3.2 Multiple, separate grains
The primary focus of this activity is to determine if there are multiple grains
associated with the business process for which you are designing your
dimensional model.

There can be more than one grain definition associated with a single business
process. In these cases, we recommend that you design separate fact tables
with separate grains and not forcefully try to put all facts in a single fact table.

Differing data granularities can be handled by using multiple fact tables (daily,
monthly, and yearly tables, as examples). Also, consider the amount of data,
space, and the performance requirements when you decide how to handle
multiple granularities.

Criteria for one or multiple fact tables
To determine whether to use one or multiple fact tables, consider the following
criteria:
One of the most important sets of criteria that helps determine the need for
one or multiple fact tables are the facts. It is important to understand the
dimensionality of the facts to decide whether the facts belong together in one
fact table or separate in fact tables with different grains. For example,
consider Figure 6-7 on page 248 which shows the grain equivalent to a single
line item on a bill. Facts, such as quantity, sales price, and discount per item,
are true to the grain. But facts, such as entire order total or entire order
quantity, would not be true to the grain definition of a single line item on a bill.
We explain the concept of identifying and handling separate grains for the
same business process in more detail in “Handling multiple, separate grains
for a business process” on page 119.
Are multiple OLTP source systems involved? Remember, each source
system is designed with a particular and very specific purpose. If two source
systems were not serving different purposes, it would be better to have one.
Often each source system will cater to a particular requirement of the
business. Generally, if we are dealing with business processes, such as order
management, store inventory, or warehouse inventory, it is likely that
separate source systems are involved, and probably the use of separate fact
tables is appropriate.
It is also important to determine if multiple, unrelated business processes are
involved. Unrelated business processes involve the creation of multiple
separate fact tables. And, it is possible that a single business process may
involve creation of separate fact tables to handle facts that have different
granularity.


If you find that a certain dimension is not true to the grain definition, then this
is a sign that it may belong to a new fact table with its own grain definition.
Consider the timing and sequencing of events. It may lead to separate
processes handling a single event. For example, a company markets its
product. Customers order the products. The accounts receivable department
produces an invoice. The customer pays the invoice. After the purchase, the
customer may return some of the products or send some back for repairs. If
any of the products are out of warranty, this requires new charges, and so on.
Here, several processes are involved in the sequence of single purchase
event. And, each of these processes are likely working with a particular, and
different, point in time. Then, each of these processes would need to be
handled using separate fact tables.

For the example of retail sales, assume that we do not have multiple fact tables,
because the facts (unit sales price (UP), quantity, cost price (Amount), and
revenue earned (Total Due)) defined in Figure 6-7 on page 248 are true to the
grain of the fact table. Therefore, we have one single grain definition, which is an
individual line item on a bill.

We discuss the concept of identifying multiple separate grain definitions (multiple
fact tables) for a single business process in section 5.2.1, “Handling multiple,
separate grains for a business process” on page 119.

Multiple granularities in a single fact table
It is possible to have multiple grains in one fact table. This can be accommodated
by adding a column called the granularity flag. Such a column would indicate
whether the data or row in the fact table is at the daily, monthly, quarterly, or
yearly level. Even though it is possible to store multiple grains in a single fact
table, we do not recommend this approach. We suggest handling multiple grain
definitions by designing separate fact tables and star schemas. Then, if desired,
the separate star schemas may be related by use of conformed dimensions and
conformed facts. For more detail, see section 6.4.3, “Conformed dimensions” on
page 268 and 6.5.2, “Conformed facts” on page 298.

6.3.3 Fact table types
In this activity, we identify the type of fact table involved in the design of the
dimensional model.

There are three types of fact tables:
Transaction: A transaction-based fact table records one row per transaction.
A detailed discussion about the transaction fact table is available in
“Transaction fact table” on page 125.


Periodic: A periodic fact table stores one row for a group of transactions that
happen over a period of time. A detailed discussion about the periodic fact
table is available in “Periodic fact table” on page 126.
Accumulating: An accumulating fact table stores one row for the entire
lifetime of an event. As examples, the lifetime of a credit card application from
the time it is sent, to the time it is accepted, or the lifetime of a job or college
application from the time it is sent, to the time it is accepted or rejected. For a
detailed discussion of this topic, see “Accumulating fact table” on page 127.

We summarize the differences among these types of fact tables in Table 6-10. It
emphasizes that each has a different type of grain. And, there are differences in
ways that inserts and updates occur in each. For example, with transaction and
periodic fact tables, only inserts occur. But with the accumulating fact table, the
row is first inserted. And as a milestone is achieved and additional facts are
made available, it is subsequently updated.

We discuss each of the three fact tables in more detail, with examples, in 5.2.3,
“Designing different grains for different fact table types” on page 124.

Table 6-10 Comparison of fact table types
Feature Transaction Periodic Accumulating

Grain One row per One row per time One row for the entire
transaction. period. lifetime of an event.

Dimension Date dimension at Date dimension at Multiple date
the lowest level of the end-of-period dimensions.
granularity. granularity.

Number of More than periodic Fewer than Highest number of
dimensions fact type. transaction fact dimensions when
type. compared to other fact
table types.

dimensions conformed conformed dimensions.
dimensions. dimensions.

Facts Related to Related to periodic Related to activities
transaction activities. which have a definite
activities. lifetime.

facts conformed facts. conformed facts. dimensions.


Feature Transaction Periodic Accumulating

Database Largest size. At Smaller than the Smallest in size when
size the most detailed Transaction fact compared to the
grain level, tends table because the Transaction and
to grow very fast. grain of the date and Periodic fact tables.
time dimension is
significantly higher.

Performance Performs well and Performs better than Performs well.
can be improved other fact table types
by choosing a because data is
grain above the stored at a less
most detailed. detailed grain.

Insert Yes Yes Yes

Update No No Yes, when a milestone is
reached for a particular
activity.

Delete No No No

Fact table Very fast. Slow in comparison Slow in comparison to
growth to transaction-based the transaction and
fact table. periodic fact table.

Need for High, primarily No or very Low, Medium, because the
aggregate because the data primarily because data is primarily stored
tables is stored at a very the data is already at the day level.
detailed level. stored at a high However, the data in
aggregated level. accumulating fact tables
is lower than the
transaction level.

Note: In the example retail sales business process, the grain is the
transaction fact table type.

6.3.4 Check grain atomicity
In this activity, we review the atomicity (level of detail) of the grain to assure it is at
the most detailed level. This decision should include consideration for anticipated
future needs in order to minimize the potential for a required redesign as
business requirements change.


The importance of having detailed atomic grain
Grain of the dimensional model is extremely important for the dimensional
design. Even if the business requirements need information at the monthly or
quarterly level, it is better to have the information made available at the daily
level. This is because the more detailed (atomic) the dimensions are, the more
detailed information that canj be provided to the business.

For example, consider the date dimension that has only a year attribute. With
this, we cannot get information at the quarter, month, or day level. To maximize
available information, it is important to choose a detailed atomic grain. In this
example, the choice could perhaps be day.

As an example, assume a grain of one product sold in a store in a retail example.
Here we would not be able to associate a customer with a particular product
purchased because there is only one row for a product. If it were purchased a
thousand times by a thousand different customers, we could not know that.

Of course, there are situations where you can always declare higher-level grains
for a business process by using aggregations of the most atomic and detailed
data. However, the problem is that when a higher-level grain is selected, the
number of dimensions are limited and may be less granular. You cannot drill
down into these less granular dimensions to get a lower level of detail.

Note: For more detailed discussion on the importance of having a detailed
atomic grain, see 5.2.2, “Importance of detailed atomic grain” on page 122.

Trade-offs in considering granularity
Granularity provides the opportunity for a trade-off between important issues in
data warehousing. For example, one trade-off can be performance versus the
volume of data (and the related cost of storing that data). Another can be a
trade-off between the ability to access data at a very detailed level versus
performance, and the cost of storing and accessing large volumes of data.
Selecting the appropriate level of granularity significantly affects the volume of
data in the data warehouse. Along with that, selecting the appropriate level of
granularity determines the capability of the data warehouse to satisfy query
requirements.

To help make this clear, refer to the example shown in Figure 6-8 on page 254.
Here we are looking at transaction data for a bank account. On the side of the
high level of detail, we see that 50 is the average number of transactions per
account and the size of the record for a transaction is 150 bytes. As a result, it
would require about 7.5 KB to store the very detailed transaction records to the
end of the month. The side with the low level of detail (with a higher level of
granularity) is shown in the form of a summary by account per month. Here, all


the transactions for an account are summarized in only one record. The
summary record would require a larger record size, perhaps 200 bytes instead of
the 150 bytes of the raw transaction, but the result is a significant savings in
storage space.

Figure 6-8 Granularity of data: The level of detail trade-off

In terms of disk space and volume of data, a higher granularity provides a more
efficient way of storing data than a lower granularity. You also have to consider
the disk space for the index of the data as well. This makes the space savings
even greater. Perhaps a greater concern is with the manipulation of large
volumes of data. This can impact performance, at the cost of more processing
power.

There are always trade-offs to be made in data processing, and this is no
exception. For example, as the granularity becomes higher, the ability to answer
different types of queries (that require data at a more detailed level) diminishes. If
you have a very low level of granularity, you can support any queries using that
data at the cost of increased storage space and diminished performance.

Look again at Figure 6-8. With a low level of granularity, you can answer the
query: How many credit transactions were there for John's demand deposit
account in the San Jose branch last week? With the higher level of granularity,
you cannot answer that question, because the data is summarized by month
rather than by week.

If the granularity does not impact the ability to answer a specific query, the
amount of system resources required for that same query can still differ


considerably. Suppose that you have two tables with different levels of granularity,
such as transaction details and a monthly account summary. To answer a query
about the monthly report for channel utilization by accounts, you could use either
of those two tables without any dependency on the level of granularity. However,
using the detailed transaction table requires a significantly higher volume of disk
activity to scan all the data as well as additional processing power for calculation
of the results. Using the monthly account summary table requires significantly
less resource.

In deciding about the level of granularity, you must always consider the trade-off
between the cost of the volume of data and the ability to answer queries.

6.3.5 High level dimensions and facts from grain
In this activity, we identify high level preliminary dimensions and facts from
whatever can be understood from the grain definition. No detailed analysis is
carried out to identify these preliminary dimensions and facts.

For our retail sales business process, we defined the grain (see 6.3.1, “Fact table
granularity” on page 247) as: An individual line item on a grocery bill. This is

Time Customer Employee

Preliminary Facts are:
1) Unit Sales Price
Grain= 1 Line Item on a Grocery Bill 2) Quantity Sold
3) Total $ Amount
4) Discount

Bill Number Date Supplier

Figure 6-9 Identifying high level dimensions and facts with the grain

Once we define the grain appropriately, we can easily find the preliminary
dimensions and facts.


Note: Preliminary facts are facts that can be easily identified by looking at the
grain definition. For example, facts such as unit price, quantity, and discount
are easily identifiable by looking at the grain. In other words, preliminary facts
are easily visible on the grocery store bill we saw in Figure 6-7. However,
detailed facts such as cost, manufacturing price per line item, and
transportation cost per Item are not preliminary facts that can be identified by
looking at the grain definition. Such facts are hidden and typically never visible
on the grocery store bill. Preliminary facts are not the final set of facts; the
formal detailed fact identification occurs in the Identify the facts phase in 6.5,
“Identify the facts” on page 293.

Using the one line item grain definition shown in Figure 6-9 on page 255, we can
identify all things that are true (logically associated with) to the grain. Table 6-11
shows high level dimensions and facts that are identified from the grain.

Table 6-11 High level dimensions and facts
Dimensions Facts (KPIs)

Customer country Sales revenue

Product (Brand, category) Quantity of units sold

Sale Date Total cost

Supplier of product Discount

Time of sale

Employee, sales person, or sales rep

Bill

Note: These preliminary high level dimensions and facts are helpful when we
formally identify dimensions (see 6.4.1, “Dimensions” on page 259) and facts
(see 6.5.1, “Facts” on page 295) in the later steps of the DMDL. The
dimensions and facts get iteratively refined in each of the phases of the
DMDL.

6.3.6 Final output of the identify the grain phase
The grain definition report is the final output report for this phase. It consists of
one or multiple definitions for the grain of the business process, and the type of
fact table (transaction, periodic, or accumulating) being used. It also includes the
high level preliminary dimensions and facts.


6.4 Identify the dimensions
In this phase, we focus on the third step of the DMDL, which is identify the
dimensions. This is depicted in Figure 6-10.

Determine All
Dimensions
Processes

Identify Conformed



Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


Table 6-12 shows the activities associated with this phase.

Table 6-12 Activities in the identify the dimensions phase
Seq. Activity name Activity description
no.

1 Identify dimensions Identifies the dimensions that are true to the
grain identified in 6.3, “Identify the grain” on
page 245.

2 Identify degenerate Identifies one or more degenerate dimensions.
dimensions

3 Identify conformed Identifies any existing shared dimensions in the
dimensions data warehouse or other star schemas used for
designing the dimensional model.


Seq. Activity name Activity description
no.

4 Identify dimensional Identifies the dimension attributes. It also
attributes and dimensional identifies hierarchies, such as balanced,
hierarchies unbalanced, or ragged, that may exist in the
dimensions. Techniques are suggested to
handle these hierarchies in the design.

5 Identify date and time Identifies the date and time dimensions.
granularity Typically these dimensions have a major impact
on the grain and size of the dimensional model.

6 Identify slowly changing Identifies the slowly changing dimensions in the
dimensions design. Three techniques (Type-1, Type-2, and
Type-3) are described.

7 Identify very fast changing Identifies very fast changing dimensions and
dimensions describes ways of handling them by creating
one or more mini-dimensions.

8 Identify cases for Identifies what dimensions need to be
snowflaking snowflaked.

9 Other dimensional Other challenges relating to dimensions:
challenges to look for are:

→ Identify Multi-valued Looks for multi-valued dimensions and
Dimensions describes ways of handling them, such as by
using bridge tables.

→ Identify Role-Playing Describes ways of looking for dimensions that
Dimensions can be implemented by using role-playing.

→ Identify Heterogeneous Describes ways of identifying heterogeneous
Dimensions products and implementing them.

→ Identify Garbage Describes ways to look for low-cardinality fields
Dimensions and using them to make a garbage dimension.

→ Identify Hot Swappable Describes ways of creating profile-based tables
Dimensions or hot swappable dimensions to improve
performance and secure data.

Note: We have listed activities to look into while designing dimensions. The
purpose is to make you aware of several design techniques you can use.
Knowing they exist can help you make a solid dimensional model.


6.4.1 Dimensions
During this phase, we identify the dimensions that are true to the grain we chose
in 6.3, “Identify the grain” on page 245.

Dimension tables
Dimension tables contain attributes that describe fact records in the fact table.
Some of these attributes provide descriptive information; others are used to
specify how fact table data should be summarized to provide useful information
to the business analyst. Dimension tables contain hierarchies of attributes that
aid in summarization. Dimension tables are relatively small, denormalized lookup
tables which consist of business descriptive columns that can be referenced
when defining restriction criteria for ad hoc business intelligence queries.

Important points to consider about dimension tables are:
Each dimension table has one and only one lowest level element, or lowest
level of detail, called the dimension grain, also referred to as the granularity
of the dimension.
Dimension tables that are referenced, or are likely to be referenced, by
multiple fact tables are conformed dimensions. If conformed dimensions
already exist for any of the dimensions in your model, you are expected to
use the conformed versions. If you are developing new dimensions with
potential for usage across the entire enterprise, you are expected to develop
a design that supports anticipated enterprise needs. If you create a new
dimension in your design which you think could potentially be used by other
dimensional models or business processes, then you should design the new
dimension keeping in mind your business requirements and the future
anticipated needs. Conformed dimensions are discussed in more detail in
“Conformed dimensions” on page 268.
Each non-key element (other than the surrogate key) should appear in only
one dimension table.
All dimension table primary keys should be surrogate keys. An OLTP source
system primary key should not be used as the primary key of the dimension
table. Surrogate keys are discussed in detail in “Primary keys for dimension
tables are surrogate keys” on page 263.
Most dimensional models will have one or more date and time dimensions.
Date and time dimensions are handled separately. The concepts relating to
date and time dimensions are discussed in “Date and time granularity” on
page 279.
If a dimension table includes a code, then in most cases the code description
should be included as well. For example, if region locations are identified by a
region code, and each code represents a region name, both the code and the


region name should be included in the dimension table. It is important to
understand that the quality of a dimensional model is directly proportional to
the quality of the dimensional attributes.
Typically, dimensional models should not have more than 10-15 dimensions.
If you have more dimensions, you might need to find ways to merge
dimension tables into one.
The primary keys (surrogate keys) of the dimension tables should be included
in the fact table as foreign keys.
Typically, each dimension table will have one or more additional Not
Applicable scenario records. This is primarily because of the fact that the
foreign key in a fact table can never be null, since by definition that violates
referential integrity. This concept is explained in more detail in “Insert a
special customer row for the “Not applicable” scenario” on page 274.
The rows in a dimension table establish a one-to-many relationship with the
fact table. For example, there may be a number of sales to a single customer,
or a number of sales of a single product. The dimension table contains
attributes associated with the dimension entry; these attributes are rich and
business-oriented textual details, such as product name or customer name.
Attributes serve as report labels and query constraints.

Note: Each dimension attribute should take on a single value in the context of
each measurement inside the fact table. However, there are situations where
we need to attach a multi-valued dimension table to the fact table. In other
words, there are situations where there may be more than one value of a
dimension for each measurement. Such cases are handled using multi-valued
dimensions, which are explained in detail in 5.3.10, “Multi-valued dimensions”
on page 182.

To summarize, in this activity we formally identify all dimensions that are true to
the grain. By looking at the grain definition, it is relatively easy to define the
dimensions. The dimensions for our retail sales example (for a grocery store) are
depicted graphically (shown in Figure 6-11 on page 261) using the grocery bill.


Bill Number#
Customer Customer: Carlos (Degenerate
Dimension)
08/29/2005 Date
Store Store=S1394 1600 Hours
Time

Grain = 1 Line Quantity UP Dsc Discount
Item on a 1. Eggs 12 $3 $36
Grocery Bill $2 $4
2. Dairy Milk 2
3. Chocolate Powder 1 $9 $9 Unit Price
4. Soda Lime 12 $1.5 $18

Employee Employee: Amit $75 Total Amt

Figure 6-11 Graphical identification of dimensions and facts

List all the dimensions that are thought to be associated with this grain. The
dimensions identified are shown in Table 6-13. For the dimensions, also define
the level of detail (granularity) that to include in the dimension. However, the
descriptive and detailed attributes are defined later in 6.4.4, “Dimensional
attributes and hierarchies” on page 269.

Table 6-13 Dimensions and facts from grain
Seq. Dimensions Dimension granularity
no.

1 Time It is true to the grain, and has a description of time to the
hour and minute level.

2 Customer It is true to the grain, and this dimension describes the
customers.

3 Employee It is true to the grain, and this dimension describes the
employees working in the stores and associated with
the retail sale.


Seq. Dimensions Dimension granularity
no.

4 Supplier It is true to the grain, and this dimension describes the
suppliers of the products.

5 Product It is true to the grain, and this dimension describes the
products, including their brand and category.

6 Date It is true to the grain, and this dimension describes the
different dates on which the products were sold. The
date includes the day, month, quarter, and year
description.

7 Store It is true to the grain, and this dimension describes the
stores which sold the products.

8 Bill Number It is true to the grain, and this dimension describes the
(Pos_Bill_Number) Bill or receipt of the store. This is handled as a
degenerate dimension, described in section 6.4.2,
“Degenerate dimensions” on page 266.

The preliminary dimensional schema that we get at this point is shown in
Figure 6-12 on page 263. The dimension attributes are mentioned in the
dimensional as TBD (To be Determined).

Note: The facts unit price, discount, quantity, and revenue are the preliminary
facts identified from 6.3.5, “High level dimensions and facts from grain” on
page 255. The facts and several other derived measures are identified in 6.5,
“Identify the facts” on page 293.


*Preliminary Facts
(To be iteratively formalized TBD=To be Determined
in the ‘Identify Facts’ Phase)
BILL_NUMBER (DD) = Degenerate Dimension

Figure 6-12 Preliminary retail sales grocery store dimensional model

Note: BILL_NUMBER is a degenerate dimension. It has no attributes
associated with it and goes inside the fact table. We discuss degenerate
dimensions in 6.4.2, “Degenerate dimensions” on page 266.

Primary keys for dimension tables are surrogate keys
It is important that primary keys of dimension tables remain stable. We strongly
recommend that surrogate keys are created and used for primary keys for all
dimension tables. In this section, we discuss what surrogate keys are, and why it
is important to use the surrogate keys as the dimension table primary keys.

What are surrogate keys?
Surrogate keys are keys that are maintained within the data warehouse instead
of the natural keys taken from source data systems. Surrogate keys are known
by many other aliases, such as dummy keys, non-natural keys, artificial keys,
meaningless keys, non-intelligent keys, integer keys, number keys, and technical
integer keys. The surrogate keys join the dimension tables to the fact table.
Surrogate keys serve as an important means of identifying each instance or
entity inside a dimension table.

Reasons for using surrogate keys are:
Data tables in various OLTP source systems may use different keys for the
same entity. It may also be possible that a single key is being used by different


instances of the same entity. This means that different customers might be
represented using the same key across different OLTP systems.
This can be a major problem when trying to consolidate information from
various source systems. Or for companies trying to create/modify data
warehouses after mergers and acquisitions. Existing systems that provide
historical data might have used a different numbering system than a current
OLTP system. Moreover, systems developed independently may not use the
same keys, or they may use keys that conflict with data in the systems of
other divisions. This situation may not cause problems when each
department independently reports summary data, but can when trying to
achieve an enterprise-wide view of the data.
This means that we cannot rely on using the natural primary keys of the
source system as dimension primary keys because there is no guarantee that
the natural primary keys will be unique for each instance. A surrogate key
uniquely identifies each entity in the dimension table, regardless of its natural
source key. This is primarily because a surrogate key generates a simple
integer value for every new entity.
Surrogate keys provide the means to maintain data warehouse information
when dimensions change. To state it more precisely, surrogate keys are
necessary to handle changes in dimension table attributes. We discuss in
more detail in “Slowly changing dimensions” on page 283.
Natural OLTP system keys may change or be reused in the source data
systems. This situation is less likely than others, but some systems have
reuse keys belonging to obsolete data or for data that has been purged.
However, the key may still be in use in historical data in the data warehouse,
and the same key cannot be used to identify different entities.
The design, implementation, and administration of surrogate keys is the
responsibility of the data warehouse team. Surrogate keys are maintained in
the data preparation area during the data transformation process.
One simple way improve performance of queries is to use surrogate keys.
The narrow integer surrogate keys mean a thinner fact table. The thinner the
fact table, the better the performance.
Surrogate keys also help handle exception cases such as the To Be
Determined or Not Applicable scenarios. This is discussed in “Insert a special
customer row for the “Not applicable” scenario” on page 274.
Changes or realignment of the employee identification number should be
carried in a separate column in the table, so information about the employee
can be reviewed or summarized, regardless of the number of times the
employee record appears in the dimension table. For example, changes in the
organization sales force structures may change or alter the keys in the
hierarchy.


This is a common situation. For example, if a sales representative is
transferred from one region to another, the organization may want to track all
sales for the sales representative with the original region for data prior to the
transfer date, and sales data for the sales representative in the new region
after the transfer date.
To represent this organization of data, the salesperson's record must exist in
two places in the Sales_Team dimension table. This is certainly not possible if
the sales representative's company employee identification number is used
as the primary key for the dimension table, because a primary key must be
unique. A surrogate key allows the same sales representative to participate in
different regions in the dimension hierarchy.
In this case, the sales representative is represented twice in the dimension
table with two different surrogate keys. These surrogate keys are used to join
the sales representative's records to the sets of facts appropriate to the
various regions in the hierarchy occupied by the sales representative.
The sales representative's employee number should be carried in a separate
column in the table so information about the sales representative can be
reviewed or summarized regardless of the number of times the sales
representative's record appears in the dimension table. The employee
number also helps us go back to the OLTP system from where the record was
loaded.

Note: Use a separate field in the dimension table to preserve the natural
source system key of the entity being used in the source system. This
helps us to go back to the original source if we need to track from where
(which OLTP Source) the data came into the dimensional model. Also, we
are able to summarize fact table data for a single entity in the fact table
regardless of the number of times the entity’s record appears in the
dimension table.

GUID (Globally Unique Identifiers) as surrogate keys
We should strictly avoid the use of GUIDs as primary keys for the dimension
tables. GUIDs are known to work well in the source OLTP systems, but they are
difficult to use when it comes to data warehouses. This is primarily because of
two reasons:
1. The first reason is storage. GUIDs use a significant amount of space
compared to their integer counterparts. GUIDs take about 16 bytes each,
where an integer takes about 4 bytes.
2. The second reason is that indexes on GUID columns are relatively slower
than indexes on integer keys because GUIDs are four times larger.


Note: Every join between dimension and fact tables in the dimensional model
should be based only on artificial integer surrogate keys. Use of natural OLTP
system primary keys for dimensions should be avoided.

How to identify dimensions from an E/R model
The source of a dimensional model is either the enterprise data warehouse or
the OLTP source systems. It would be correct to say both the data warehouse
and the OLTP source systems are typically based on E/R models; they are in
3NF. We think if you can create a dimensional model from an E/R model, then
you should be able to create dimensional models either from the data warehouse
or directly from the OLTP source systems.

The following are the steps involved in converting an E/R model to a dimensional
model:
1. Identify the business process from the E/R model.
2. Identify many-to-many tables in the E/R model to convert to fact tables.
3. Denormalize remaining tables into flat dimension tables.
4. Identify date and time from the E/R model.

This process of converting an existing E/R model (which could be a data
warehouse or an OLTP source system) is explained in detail in 5.1, “Converting
an E/R model to a dimensional model” on page 104.

6.4.2 Degenerate dimensions
In this section, we focus on identifying degenerate dimensions, which are
dimensions without any attributes. They are not typical dimensions, but often
simply a transaction number that is placed inside the fact table. In order to
understand the concept of degenerate dimensions, we have to understand the
source of the degenerate dimension, which originates in the form of some
transaction numbers inside the OLTP system.

All OLTP source systems generally consist of transaction numbers, such as bill
numbers, courier tracking number, order number, invoice number, application
received acknowledgement number, ticket number, and reference numbers.
These transaction numbers in the OLTP system generally tell about the
transaction as a whole. Consider the retail sales example (for grocery store) and
reanalyze the graphical bill as shown in Figure 6-13 on page 267.


Bill Number#
Customer Customer: Carlos (Degenerate
Dimension)
08/29/2005 Date
Store Store=S1394 1600 Hours
Time

Grain = 1 Line Quantity UP Dsc Discount
Item on a 1. Eggs 12 $3 $36
Grocery Bill $2 $4
2. Dairy Milk 2
3. Chocolate Powder 1 $9 $9 Unit Price
4. Soda Lime 12 $1.5 $18

Employee Employee: Amit $75 Total Amt

Figure 6-13 Retail sales grocery store bill

Try to analyze the Bill Number# information for the retail sales grocery store
example. The Bill Number# is a transaction number that tells us about the
transaction (purchase) with the store. If we take an old Bill Number# 973276 to
the store and ask its manager to find out information relating to it, we may get all
information related to the bill. For example, assume that the manager replies that
the Bill Number# 973276 was generated on August 11, 2005. The items
purchased were Apples, Oranges and Chocolates. The manager also tells us the
quantity, unit price, and discount for each of the items purchased. He also tells us
the total price. In short, the Bill Number# tells us the following information:
Transaction date
Transaction time
Products purchased
Quantity, unit price, and amount for each purchased product

If we now consider the Bill for our retail sales example, the important point to note
is that we have already extracted all information related to the Bill Number# into
other dimensions such as date, time, and product. Information relating to
quantity, unit price, and amount charged is inside the fact table.


A degenerate dimension, such as Bill Number#, is there because we chose the
grain to be an individual line item on a bill. So, the Bill Number# degenerate
dimension is there because the grain we chose represents a single transaction or
transaction line item.

The Bill Number# is still useful because it serves as the grouping key for
grouping together of all the products purchased in a single transaction or for a
single Bill Number#. Although to some, the Bill Number# looks like a dimension
key in the fact table, it is not, because all information relating to the Bill Number#
has been allocated to different dimensions. Hence, the so-called Bill Number#
dimension has no attributes, and we refer it to as a degenerate dimension.

Other important things to consider about degenerate dimensions are:
How to identify a degenerate dimension for a dimensional design.
Should we make a separate dimension for the Bill Number#?
Should we place the Bill Number# inside the fact table? If yes, why?
How to realize that a degenerate dimension is missing.
Under what situation will the Bill Number# no longer be a degenerate
dimension? We discuss one example in “Identify degenerate dimensions” on
page 384. We see that Invoice Number (a type of transaction number) is
handled as a separate dimension and not as a degenerate dimension inside
the fact table.

We discuss interesting topics about degenerate dimensions in more detail in
5.3.1, “Degenerate dimensions” on page 134.

6.4.3 Conformed dimensions
In this activity, we identify any conformed shared dimensions that are available to
use instead of redesigning the dimensions. In this activity, we identify whether a
dimension being used already exists inside the enterprise data warehouse or
dimensional model.

What are conformed dimensions?
A conformed dimension means the same thing to each fact table to which it can
be joined. A more precise definition is that two dimensions are conformed if they
share one, more than one, or all attributes that are drawn from the same domain.
A dimension may be conformed even if it contains only a subset of attributes
from the primary dimension.

Typically, dimension tables that are referenced or are likely to be referenced by
multiple fact tables (multiple dimensional models) are called conformed
dimensions. If conformed dimensions already exist for any of the dimensions in


the data warehouse or dimensional model, you are expected to use the
conformed dimension versions. If you are developing new dimensions with
potential for usage across the entire enterprise warehouse, you are expected to
develop a design that supports anticipated enterprise warehouse needs. In order
to find out the anticipated warehouse needs, you might need to interact with
several business processes to find out how they would define the dimensions.

To quickly summarize, the identify conformed dimensions activity involves the
following two steps:
1. Identifying whether a dimension being used already exists. If the dimension
being used already exists in the enterprise data warehouse, then that
dimension has to be used. If the dimension would exist, then we would not
identify dimensional attributes (see “Dimensional attributes and hierarchies”
on page 269) for that particular dimension. A dimension may be conformed to
the existing dimension even if it contains only a subset of attributes from the
primary dimension.
2. For a nonexistent new dimension, create a new dimension with planned
long-term cross-enterprise usage. When a dimensional model requires a
dimension which does not exist in the enterprise warehouse or any other
dimensional models, then a new dimension must be created. While creating
this new dimension, you must be certain to interact with enterprise business
functions to find out about their future anticipated need of the new dimension.

6.4.4 Dimensional attributes and hierarchies
This activity involves the following:
Identify dimensional attributes for the dimensions we identified in section
“Dimensions” on page 259.
Identify the various hierarchies (such as balanced, unbalanced, and ragged)
associated with each of the dimensions.

Why are good quality dimensional attributes important?
After having identified the dimensions, the next step is to fill the dimensions with
good quality attributes. The dimension tables contain business descriptive
columns that users reference to define their restriction criteria for ad hoc
business queries.

The quality of a good dimensional model is directly proportional to the quality of
attributes present inside these dimension tables. The dimension table attributes
show up as report labels inside the reports for senior management.

So, which attribute should be included in the dimensions? To help answer that
question, here are points to consider:


Non-key columns are generally referred to as attributes. Every dimension
table primary key should be a surrogate key, which is usually not considered
an attribute because it is simply an integer and is not used for analysis.
Use a separate field in the dimension table to preserve the natural source
system key of the entity being used in the source system.
A schema design that contains complete, consistent, and accurate attribute
fields helps enable queries that are intuitive, and reduces the support burden
on the organization responsible for database management and reports.
A well-designed schema includes attributes that reflect the potential areas of
interest and attributes that can be used for aggregations as well as for
selective constraints and report breaks.
If a dimension table includes a code, in most cases, include the code
description as well. As an example, if branch locations are identified by a
branch code, and each code represents a branch name, then include both the
code and the name. Avoid storing cryptic decodes inside a dimensional table
attribute to save space.
Be sure attribute names are unique in the model. If you have duplicate names
for different attributes, use the prime term (entity name) to create a distinction.
For example, if you have multiple attributes called Address Type Code,
rename one Beneficiary Address Type Code, and another might be Premium
Address Type Code.
The dimensional attributes serve as headings of the columns of the report
and should be descriptive and easy to understand. For example, for a given
situation where you want to store a flag such as 0/1 or Y/N, it is better to store
something descriptive, such as Yes/No.
An attribute can be defined to permit missing values in cases where an
attribute does not apply to a specific item or its value is unknown.
An attribute may belong to more than one hierarchy.
Use only the alphabetic characters A-Z and the space character. Do not use
special characters.
While naming attributes, do not use possessive nouns. For example, use
terms such as Recipient Birth Date rather than Recipient's Birth Date.
Do not reflect permitted values in the attribute name. For example, the
attribute name Employee Day/Night Code refers to code values designating
day shift or night shift employees working in the grocery store. The attribute
must be named to reflect the logical purpose and the entire range of values.
For example, Employee Shift Type Code which allows for an expandable set of
valid values.
Do not include very large names for building attributes.


Properly document all dimensional attributes.

Note: It is important to remember that the fact table occupies 85-90% of space
in large dimensional models. Hence, saving space by using decodes in
dimensional attributes does not save much overall space, but these decodes
affect the overall quality and understandability of the dimensional model.

In 6.4.1, “Dimensions” on page 259, we designed the preliminary star schema

*Preliminary Facts

Figure 6-14 Preliminary retail sales grocery store dimensional model

Now fill this preliminary star schema with detailed dimension attributes for each
dimension. Before filling in the attributes of various dimension tables, define the
granularity of each dimension table as shown in Table 6-14.

Note: While deciding on attributes for each of the dimensions, keep in mind
the requirements analysis performed in 6.2.7, “Requirements analysis” on
page 242.

Table 6-14 Granularities for dimension tables
Seq. Dimension table Granularity of the dimension table
no.

1 Employee This dimension stores information at the single
employee level. Also includes manager information.


Seq. Dimension table Granularity of the dimension table
no.

2 Supplier Contains information about supplier’s company name
and address.

3 Date This dimension stores information down to the day level
including month, quarter, and year.

4 Time This dimension stores information about hour, and a
description of the time, such as lunch time, early
morning shift, morning, afternoon shift, and night shift.

5 Customer This dimension stores information about the customer.

6 Product This dimension stores information about the product,
brand, and category name.

7 Store This dimension stores information about the store
dimension.

8 Bill Number This is a degenerate dimension. It has no attributes and
is present as a number (BILL_NUMBER) inside the fact
table.

We explain the dimensions, along with their detailed attributes, below:
Employee dimension: Shown in Figure 6-15 on page 273, the employee
dimension stores information about the employees working at the store. The
grain of the employee dimension table is information about a single
employee.


*Preliminary Facts TBD=To be Determined
(To be iteratively formalized in
the Identify Facts Phase)

Figure 6-15 Employee dimension

The business requirements analysis in Table 6-8 on page 242 shows that
questions Q2 and Q3 require the business to see sales figures for good and bad
performing employees, along with the managers of these employees. The
manager name column in the employee dimension table helps in that analysis.
Supplier dimension: As shown in Figure 6-16 on page 274, the supplier
table consists of attributes that describe the suppliers of the products in the
store.


*Preliminary Facts
in the Identify Facts Phase)
BILL_NUMBER DD) = Degenerate Dimension

Figure 6-16 Supplier dimension

Customer dimension: The customer dimension is shown in Figure 6-17 on
page 275. This dimension consists of customer information.

Insert a special customer row for the “Not applicable” scenario
Typically, in a store environment, many customers are unknown to the store. We
need to include a row in the customer dimension, with its own unique surrogate
key, to identify a Customer unknown to store or Not applicable, and to avoid a
null customer key in the fact table. The concepts of referential integrity are
violated if we put a null in a fact table column declared as a foreign key to a
dimension table. In addition, null keys are the source of great confusion to many
because they cannot join on null keys.

In the customer dimension table, we insert a Customer unknown to store or Not
applicable row inside the customer table and link the fact table row
(measurements) to this row when the customer is unknown.


*Preliminary Facts

Figure 6-17 Customer dimension

Note: You must avoid null keys in the fact table. A good design includes a row
(Not applicable) in the corresponding dimension table to identify that the
dimension is not applicable to the measurement.

Product dimension: The product dimension, shown in Figure 6-18 on
page 276, holds information related to products selling in the stores. It
consists of the following hierarchy: Category → Brand → Product.


(To be iteratively formalized in BILL_NUMBER (DD) = Degenerate Dimension
the Identify Facts Phase) BILL_NUMBER (DD) = Degenerate Dimension
the ‘Identify Facts’ Phase)

Figure 6-18 Product dimension

Store Dimension: The store dimension holds information related to stores.
This table is depicted in Figure 6-19 on page 277.


(To be iteratively formalized in BILL_NUMBER (DD) = Degenerate Dimension
the Identify Facts Phase) BILL_NUMBER (DD) = Degenerate Dimension
Figure 6-19 Store dimension

Date and time dimension: The date and time dimensions are explained in
detail in section “Date and time granularity” on page 279.

Identifying dimension hierarchies
A hierarchy is a cascaded series of many-to-one relationships. A hierarchy
basically consists of different levels, each corresponding to a dimension attribute.

In other words, a hierarchy is a specification of levels that represents
relationships between different attributes within a hierarchy. For example, one
possible hierarchy in the date dimension is Year → Quarter → Month → Day.

There are three major types of hierarchies that you should look for in each of the
dimensions. They are explained in Table 6-15 on page 278.


Table 6-15 Different types of hierarchies
Seq. Name Hierarchy description How is it implemented?
no.

1 Balanced In a balanced hierarchy, all the See “How to implement a
dimension branches have the balanced hierarchy” on
same number of levels. For page 144.
more details, see “Balanced
hierarchy” on page 143.

2 Unbalanced A hierarchy is unbalanced if it See “How to implement an
has dimension branches unbalanced hierarchy” on
containing varying numbers of page 146.
levels. Parent-child dimensions
support unbalanced
hierarchies. For more details,
see “Unbalanced hierarchy” on
page 145.

3 Ragged A ragged dimension contains at See “How to implement a
least one member whose ragged hierarchy in
parent belongs to a hierarchy dimensions” on page 155.
that is more than one level
above the child. Ragged
dimensions, therefore, contain
branches with varying depths.
For more details, see “Ragged

Note: A dimension table may consist of multiple hierarchies. And, a
dimension table may consist of attributes or columns which belong to one,
more, or no hierarchies.

For the retail sales store example, Table 6-16 shows the hierarchies present
inside various dimensions.

Table 6-16 Dimensions and hierarchies
No Name Hierarchy description Type

1 Date Calendar Hierarchy: Calendar Year → Calendar Balanced
Month → Calendar Week

Fiscal Hierarchy: Fiscal Year → Fiscal Quarter →
Fiscal Month → Fiscal Day

2 Time None


No Name Hierarchy description Type

3 Product Category Name → Brand Name → Product Name Balanced

4 Employee None

5 Supplier Supplier Country → Supplier Region → Supplier Balanced
City

6 Customer Customer Country → Customer Region → Balanced
Customer City

7 Store Store Country → Store State → Store Region → Balanced
Store Area

We have a detailed discussion about handling hierarchies in 5.3.4, “Handling
dimension hierarchies” on page 142.

6.4.5 Date and time granularity
It is very important to identify the granularity of the date and time dimensions.
This is primarily because the date and time dimensions help determine the
granularity of the overall dimensional model and the level of information that is
stored in it. Choosing a wrong grain in either the date or time dimension may
result in important dimensions being omitted from the dimensional model. For
example, if we are designing a dimensional model for an order management
system and we choose the grain of the date dimension as quarter, we may miss
many other dimensions (such as time and employee) that could be included in
the model had the date dimension been at the day grain. We explain the date
and time dimensions in detail below:

Date dimension
Every data mart has a date dimension because all dimensional models are
based on a time series of operations. For example, you typically want to measure
performance of the business over a time period. It is also possible that a
dimensional model consists of several date dimensions. Such dimensions are
usually deployed using a concept called role-playing, which is implemented by
views. Role-playing is discussed in detail in section 5.3.9, “Role-playing

For the retail sales grocery store example, the date dimension is shown
Figure 6-20 on page 280. The grain of the date dimension is a single day.


TBD=To be Determined
*Preliminary Facts

Figure 6-20 Date dimension

Note: The time dimension is handled separately from the date dimension. It is
not advised to merge the date and time dimension into one table because a
simple date dimension table which has 365 rows (for one year) would explode
into 365 (Days) x 24 (Hours) x 60 (Minutes) x 60 (Seconds) = 31536000 rows
if we tried storing hours, minutes, and seconds. This is for just one year.
Consider the size if it were merged with time.

Typically, the date dimension does not have an OLTP source system connected
to it. The date dimension can be built independently even before the actual
dimensional design has started. The best way to built a date dimension is to
identify all columns needed and then use SQL language to populate the date
dimension table for columns such as date, day, month, and year. For other
columns, such as holidays, with Christmas or Boxing Day, manual entries may
be required. The same is true for columns such as fiscal date, fiscal month, fiscal
quarter, and fiscal years. You may also use the spreadsheet to create the date
dimension.

The date dimension in Figure 6-20 consists of the normal calendar and fiscal
hierarchies. These hierarchies are shown in Figure 6-21 on page 281.


Fiscal Multiple Calendar
Year Year
Hierarchies

Fiscal
Quarter
Calendar Attributes
Attributes Quarter

Fiscal
Month

Fiscal Calendar
Week Month

Day of
Week

Grain of Date Dimension

Figure 6-21 Calendar and fiscal hierarchy in the date dimension

Handle date as a dimension or a fact?
Instead of using a separate date table, as shown in Figure 6-20 on page 280, we
could have used a date/time column in the fact table. In this way, we could have
removed the date table from our dimensional model. We could have used the
SQL date functions present inside each database to filter out the day, month, and
year as examples. But this design where we replace the date dimension with a
column inside the fact table has issues, for the following reasons:
As shown in Figure 6-20 on page 280, there are several date attributes not
supported by the SQL date functions, including fiscal periods, holidays,
seasons, weekdays, weekends, and national events. This is one of the
primary reasons why we want to have a separate date dimension table. This
enables the business to look at the key performance indicators of their
business across various fiscal and other date-related attributes which are not
possible if we used the SQL date/time column inside the fact table.
From an ease of use point of view, it is much easier to drag the columns from
a date table instead of using complex SQL functions to create the logic for the
reports.

Therefore, we think it is important to have a separate date dimension table
instead of using a simple date/time column field in the fact table.


Insert a special date row for the “Not applicable” scenario
We discussed the Not applicable scenario in “Insert a special customer row for
the “Not applicable” scenario” on page 274 for the customer table. Similarly for
the date dimension, we insert a Not applicable row inside our date dimension
table. In the date dimension table, we include a Date unknown to store or Not
applicable or Corrupt row inside the date table and link the fact table row
(measurements) to this row when the date is unknown, or when the recorded
date is inapplicable, corrupted, or has not yet happened.

Recalling the grain for the retail sales grocery example, which is an individual
line item on a bill, the date dimension is surely going to have a date with each
product sold, and hence each fact table row is pointing to one valid row inside the
date dimension table.

How to handle several dates across International time zones
The topic of handling date and time across International times zones is
discussed in 5.3.3, “Handling date and time across international time zones” on
page 142.

Time dimension
The time dimension for our retail sales store example is shown in Figure 6-22.
Based on the requirements analysis performed in 6.2.7, “Requirements analysis”
on page 242, we designed the time dimension.

*Preliminary Facts
(To be iteratively formalized BILL_NUMBER (DD) = Degenerate Dimension

Figure 6-22 Time dimension

Sample rows for the time dimension are shown in Table 6-17 on page 283.


Table 6-17 Sample rows from the time dimension
Time ID Standard time Description Time indicator

1 0600 hours Early morning AM

2 1650 hours Evening PM

3 2400 hours Late night AM

We recommend that time should be handled separately as a dimension and not
as a fact inside the fact table.

You can handle time in dimensional modeling in two ways:
Time of day as a separate dimension
Time of day as a fact

We discuss more about time handling in 5.3.2, “Handling time as a dimension or
a fact” on page 139.

6.4.6 Slowly changing dimensions
In this phase, we identify the slowly changing dimensions and also specify what
strategy (Type-1, Type-2, or Type-3) we use to handle the change.

What are slowly changing dimensions?
A slowly changing dimension is a dimension whose attributes for a record (row)
change slowly over time.

Assume that David is a customer of an insurance company called INS993, Inc.,
and first lived in Albany, New York. So, the original entry in the customer
dimension table had the record as shown in Table 6-18:

Table 6-18 Insurance customer dimension table
Customer key Social security number Name State

953276 989898988 David New York

David moved to San Jose, California, in August, 2005. How should INS993, Inc.
now modify the customer dimension table to reflect this change? This is the
Slowly Changing Dimension problem.

There are typically three ways to solve this type of problem, and they are
categorized as follows:


Type 1: The new row replaces the original record. No trace of the old record
exists. This is shown in Table 6-19. There is no history maintained for the fact
that David lived in New York.

Table 6-19 Type 1 change in customer dimension table
(Surrogate key)

953276 989898988 David California

Type 2: A new row is added into the customer dimension table. Therefore,
the customer is treated essentially as two people and both the original and
the new row will be present. The new row gets its own primary key (surrogate
key). After David moves from New York to California, we add a new row as
shown in Table 6-20.

(Surrogate key)

953276 989898988 David New York

953277 989898988 David California

Type 3: The original record is modified to reflect the change. Also a new
column is added to record the previous value prior to the change.
To accommodate Type 3 slowly changing dimension, we now have the
following columns:
– Customer key
– Customer name
– Original State
– Current State
– Effective Date
After David moved from New York to California, the original information gets
updated, and we have the following table (assuming the effective date of
change is August 15, 2005):

Customer Social Customer Original Current Effective
key security name state state date
number

953276 989898988 David New York California August 15,
2005


Why do we need to handle changes in data?
We need to specify how to handle changes in data. For example, what happens
if a customer address or name changes? Do we overwrite this change or does
the business need to see each change. Such decisions cannot be made by the
dimensional modeler. It is critical to involve the business users to identify how the
they would like to see the changes that happen in various business entities.

Activities for slowly changing dimensions
In this phase, we need to identify slowly changing dimensions for the
dimensional model designed in Figure 6-23. In addition, we also need to identify
how we would handle the slowly changing dimensions and the strategy we would
use.

*Preliminary Facts
(To be iteratively formalized

Figure 6-23 Identifying slowly changing dimensions

In Table 6-22, we identify the slowly changing dimensions for our retail sales
store star schema, as shown in Figure 6-23.

Table 6-22 Slowly changing dimensions identified for retail sales example
Name Strategy used Structure changed?

Product Type-2 No, only new rows added for each change.

Employee Type-3 Yes, to show current and previous manager.
Current Manager and Previous Manager
columns added.

Supplier Type-1 No, only current rows are updated for each
change. No history maintained.

Time Not applicable


Name Strategy used Structure changed?

Date Not applicable

Store Type-1 No, only current rows are updated for each
change. No history maintained.

Customer Yes, using Fast Changing Dimension
strategy discussed in 6.4.7, “Fast changing

6.4.7 Fast changing dimensions
In this phase, we identify the fast changing dimensions that cannot be handled
using the Type-1, Type-2, or Type-3 approaches used for slowly changing
dimensions. For our retail sales grocery store example, we show how to handle
this fast changing customer dimension.

All other dimensions are handled using the Type-1, Type-2, and Type-3
approach as shown in Table 6-22 on page 285.

What are fast changing dimensions?
Fast changing dimensions are also called rapidly changing dimensions. In 6.4.6,
“Slowly changing dimensions” on page 283, we focused on the typically rather
slow changes to our dimension tables. The next question is what happens when
the rate of change in these slowly changing dimensions speeds up? If a
dimension attribute changes very quickly on a daily, weekly, or monthly basis,
then we are no longer dealing with a slowly changing dimension that can be
handled by using the Type-1, Type-2, or Type-3 approach.

The best approach for handling very fast changing dimensions is to separate the
fast changing attributes into one or more separate dimensions which are called
mini-dimensions. The fact table then has two or more foreign keys—one for the
primary dimension table and another for the one or more mini-dimensions
(consisting of fast changing attributes). The primary dimension table and all its
mini-dimensions are associated with one another every time we insert a row in
the fact table.

Activities for the “Identify fast changing dimensions” phase
To handle fast changing dimensions, here are the activities:
All dimensions are analyzed to find which dimensions change very fast.
Assume that the Customer Dimension is a fast changing dimension.
The Customer dimension is analyzed further to understand the impact on the
size of the dimension table if the change is handled using the Type-2


approach. If the impact on the size of the dimension table would be huge,
then the Type-2 approach is avoided.

Note: The Type-1 approach does not store history, so it is certainly not used
to handle fast changing dimensions. The Type-3 approach is also not used
because it allows us to see new and historical fact data by either the new or
prior attribute values. However, it is inappropriate if you want to track the
impact of numerous intermediate attribute values, which would be the case for
a fast changing dimension.

The next step is to analyze the fast changing Customer dimension in detail to
identify which attributes of this dimension are subject to change fast. Assume
that we identify seven fast changing attributes in the Customer dimension, as
shown in Figure 6-24 on page 288. They are age, income, test score, rating,
credit history score, customer account status, and weight. Such fast changing
attributes are then separated into one new dimension
(Customer_Mini_Dimension) table, whose primary key is attached to the fact
table as a foreign key. This is also shown in Figure 6-24 on page 288.
After having identified the constantly changing attributes and putting them in
the Customer_Mini_Dimension mini-dimension table, the next step is to
convert these identified attributes individually into band ranges. The concept
behind this exercise is to force these attributes to take limited, discreet
values. For example, assume that each of the above seven attributes takes
on 10 different values. Then, the Customer_Mini_Dimension will have 1
million values. Thus, by creating a mini-dimension table consisting of
band-range values, we have avoided the problem where the attributes such
as age, income, test score, rating, credit history score, customer account
status, and weight can no longer change. These attributes cannot change
because they have a fixed set of band-range values (see Table 5-18 on
page 166) instead of having a large number of values.


Fast Changing
Attributes in a
Fast Changing
Dimension
Mini-Dimension

(a) Before (b) After

Figure 6-24 Handling a fast changing dimension

Therefore, we have two dimension tables:
a. One table is the primary original fast changing dimension minus the fast
changing attributes. This is the Customer table as shown in Figure 6-24.
b. The second table consists of the fast changing attributes which are
typically in the mini-dimension. This table is the
Customer_Mini_Dimension as shown in Figure 6-24.

Note: It is possible that a very fast changing dimension table may be split into
one or more mini-dimensions.

The following topics about fast changing dimensions are discussed in more detail
in 5.3.6, “Handling fast changing dimensions” on page 163:
Fast changing dimensions with a real-time example.
Mini-dimensions and band range values.
Fast changing dimensions, resolving issues with band ranges, and
Mini-dimensions.
Snowflaking does not resolve the fast changing dimension problem.

After completing this activity (Identify Fast Changing Dimensions), we get the
schema as shown in Figure 6-25 on page 289.


*Preliminary Facts
BILL_NUMBER (DD)=Degenerate Dimension

Figure 6-25 Retail sales grocery store star schema

6.4.8 Cases for snowflaking
result in the implementation of a snowflake design. A dimension table is said to
be snowflaked when the low-cardinality attributes in the dimension have been
removed to separate normalized tables and these normalized tables are then
joined back into the original dimension table.

environment because it can impact understandability of the dimensional model
and can result in decreased performance because a higher number of tables
need to be joined.

For each dimension selected for the dimensional model, we need to identify
which should be snowflaked. There are no candidates identified for snowflakes in
the grocery store example.

The following topics on snowflaking are discussed in more detail in 5.3.7,
“Identifying dimensions that need to be snowflaked” on page 171:
What is Snowflaking?
When do you do snowflaking?
When to avoid snowflaking?
Under what conditions will snowflaking improve performance?
Disadvantages of snowflaking.


6.4.9 Other dimensional challenges
In this phase, we identify other special types of dimensions, as shown in
Table 6-23.

Table 6-23 Other special dimensions
Seq Type How it is implemented
no.

1 Multi-valued Description
dimension Typically while designing a dimensional model, each
dimension attribute should take on a single value in the
context of each measurement inside the fact table.
However, there are situations where we need to attach a
multi-valued dimension table to the fact table, because
there may be more than one value of a dimension for each
measurement. Such cases are handled using
Multi-valued dimensions.

Implementation
Multi-valued dimensions are implemented using Bridge
tables. For a detailed discussion, refer to 5.3.10,
“Multi-valued dimensions” on page 182.

2 Role-playing Description
dimension A single dimension that is expressed differently in a fact
table using views is called a role-playing dimension. A
date dimension is typically implemented using the
role-playing concept when designing a dimensional
model using an Accumulating snapshot fact table. This is
discussed in more detail in Chapter 5, “Modeling
considerations” on page 103 and “Accumulating fact
table” on page 127.

Implementation
The role-playing dimensions are implemented using
views. This procedure is explained in detail in 5.3.9,
“Role-playing dimensions” on page 179.


no.

3 Heterogeneous Description
dimension The concept of heterogeneous products comes to life
when you design a dimensional model for a company that
sells heterogeneous products with different attributes, to
the same customers. That is, the heterogeneous products
have separate unique attributes and it is therefore not
possible to make a single product table to handle these
heterogeneous products.

Implementation
Heterogeneous dimensions can be implemented in
following ways:
Merge all the attributes into a single product table
and all facts relating to the different heterogeneous
attributes in one fact table.
Create separate dimensions and fact tables for the
different heterogeneous products.
Create a generic design to include a single fact and
single product dimension table with common
attributes from two or more heterogeneous products.
Implementing heterogeneous dimensions is discussed in
more detail in 5.3.12, “Heterogeneous products” on
page 186.

4 Garbage Description
dimension A garbage dimension is a dimension that consists of
low-cardinality columns such as codes, indicators, status,
and flags. The garbage dimension is also referred to as a
junk dimension. Attributes in a garbage dimension are not
related to any hierarchy.

Implementation
The implementation of the garbage dimension involves
separating the low-cardinality attributes and creating a
dimension for such attributes. This implementation
procedure is discussed in detail in 5.3.8, “Identifying
garbage dimensions” on page 176.


no.

5 Hot swappable Description
dimension A dimension that has multiple alternate versions that can
be swapped at query time is called a Hot Swappable
dimension or Profile table. Each of the versions of the hot
swappable dimension can be of a different structure. The
alternate versions of the hot swappable dimensions
access the same fact table but get different output. The
different versions of the primary dimension may be
completely different, including incompatible attribute
names and different hierarchies.

Implementation
The procedure to implement Hot swappable dimensions
is discussed in detail in 5.3.13, “Hot swappable
dimensions or profile tables” on page 188.

For the retail sales grocery store example, we do not have any of the special
dimensions shown in Table 6-23 on page 290. However, we recommend that you
refer to Chapter 5, “Modeling considerations” on page 103 to understand how
these special dimensions are designed and implemented.

The dimensional model designed for the retail sales grocery business process to
the end of the Identify the dimensions phase is shown in Figure 6-26.

*Preliminary Facts
BILL_NUMBER (DD)=Degenerate Dimension

Figure 6-26 Retail sales grocery business process (dimensional model)

The model designed to this point consists of the dimensions and preliminary
facts. These preliminary facts are high level facts identified in section 6.3.5, “High
level dimensions and facts from grain” on page 255. We iteratively identify
additional facts in the next phase, 6.5, “Identify the facts” on page 293.


6.5 Identify the facts
In this phase, we focus on the fourth step of the DMDL, as shown in Figure 6-27.


Document/Study Determine All
Identify Fact Table Identify Facts
Enterprise Business Dimensions
Granularity
Processes

Identify Conformed


Process to Model Requirement Gathering Report Separate Grains Identify Dimensional

Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


To recall, we identified preliminary dimensions and preliminary facts (for the retail
sales grocery business process) in section 6.3.5, “High level dimensions and
facts from grain” on page 255. We used the grain definition, as shown in
Figure 6-28 on page 295, to quickly arrive at high level preliminary dimensions
and facts. In this phase, we iteratively, and in more detail, identify facts that are
true to this grain.

Table 6-24 on page 294 shows the activities that are associated with the Identify
the facts phase.


Table 6-24 Activities in the Identify the facts phase
Activity name Activity description

Identify facts This activity identifies the facts that are true to the grain
identified in 6.3, “Identify the grain” on page 245.

Identify conformed After we identify the facts, we determine whether any of
facts these facts are conformed. If they are, we use those
conformed facts.

Identify fact types In this activity we identify the fact types, such as:
- Additive facts
- Semi-Additive facts
- Non-Additive facts
- Derived facts
- Textual facts
- Pseudo facts
- Factless facts

Year-to-date facts Year-to-date facts are numeric values that consist of an
aggregated total from the start of year to the current date.
Here we verify that such facts are not included in a fact table
with the atomic level line items.

Event fact tables Here we describe how to handle events in the event-based
fact table. We also highlight the pseudo and factless facts
that may be associated with such tables.

Composite key design We discuss general guidelines for designing the primary
composite key of the fact table. We also describe situations
when a degenerate dimension may be included inside the
fact table composite primary key.

Fact table sizing and Guidelines are described for use by the DBA to predict fact
growth table growth.

Figure 6-28 on page 295 shows the high level dimensions and preliminary facts
identified from the grain definition.


Time Customer Employee

1) Unit Sales Price
Grain= 1 Line Item on a Grocery Bill 2) Quantity Sold
3) Total $ Amount
4) Discount

Bill Number Date Supplier

Figure 6-28 Identifying high level dimensions and facts from the grain

6.5.1 Facts
In this activity, we identify all the facts that are true to the grain. These facts
include the following:
We identified the preliminary facts in section “High level dimensions and facts
from grain” on page 255. Preliminary facts are easily identified by looking at
the grain definition or the grocery bill.
Other detailed facts which are easily identified by looking at the grain
definition as shown in Figure 6-28. For example, detailed facts such as cost
per individual product, manufacturing labor cost per product, and
transportation cost per individual product, are not preliminary facts. These
facts can only be identified by a detailed analysis of the source E/R model to
identify all the line item level facts (facts that are true at the line item grain).

For the retail sales grocery store example, we identify the facts that are true to
the grain, as shown in Table 6-25. We found other detailed facts, such as cost per
item, storage cost per item, and labor cost per item, that are available in the
source E/R model.

Table 6-25 Facts identified for the retail sales business process
Facts Fact description

Unit sales price Price of a single product.

Quantity sold Quantity of each individual product that is sold.

Amount Defined as (Unit sales price) X (Quantity sold).

Discount Discount given on a single product.

Cost per item Cost of a single product.


Facts Fact description

Total cost Defined as (Cost per item) X (Quantity sold).

Storage cost Storage cost per product.

Labor cost Labor cost per product.

Note: It is important that all facts are true to the grain. To improve
performance, dimensional modelers may also include year-to-date facts inside
the fact table. However, year-to-date facts are non-additive and may result in
facts being counted (incorrectly) several times when more than a single date is
involved. We discuss this in more detail in 6.5.4, “Year-to-date facts” on
page 300.

How to identify facts or fact tables from an E/R model
We identified the facts as shown in Table 6-25 on page 295. The process of
identifying these facts involved use of the E/R model for the Retail Sales
Business process, as shown in Figure 6-29 on page 297.


Figure 6-29 E/R Model for the retail sales business process

The following are the steps to convert an E/R model to a dimensional model:
1. Identify the business process from the E/R Model.
2. Identify many-to-many tables in E/R model to convert to fact tables.
4. Identify date and time from the E/R Model.

This process of converting an existing E/R model (which can be a Data
Warehouse or an OLTP source system) is explained in detail in 5.1, “Converting
an E/R model to a dimensional model” on page 104.

The final dimensional model, after having identified the facts (shown in
Table 6-25 on page 295), is depicted in Figure 6-30 on page 298.


Figure 6-30 Retail sales dimensional model

6.5.2 Conformed facts
A conformed fact is a shared fact that is designed to be used in the same way
across multiple data marts. So, the shared conformed facts mean the same thing
to different star schemas.

Once the facts have been identified, we must determine whether some of these
facts already exist inside the data warehouse or in some other data marts. If any
do, then ideally we should use these predefined (and hopefully well-tested) facts.

We can assume that for our retail sales business process, there are no
conformed facts.

6.5.3 Fact types
The facts inside the fact table could be of several different types, some of which
are described in Table 6-26 on page 299.


Table 6-26 Fact types
Fact type Description

Additive facts Additive facts are facts that can be added across all of the
dimensions in the fact table, and are the most common type of fact.
Additive facts may also be called fully additive facts. They are
identified here because these facts would be used across several
dimensions for summation purposes.

It is important to understand that since dimensional modeling
involves hierarchies in dimensions, aggregation of information over
different members in the hierarchy is a key element in the
usefulness of the model. Since aggregation is an additive process, it
is good if we have additive facts.

Semi-additive These are facts that can be added across some dimensions but not
facts all. They are also sometimes referred to as partially-additive facts.
For example, facts such as head counts and quantity-on-hand
(inventory) are considered semi-additive.

Non-additive Facts that cannot be added for any of the dimensions. That is, they
facts cannot be logically added between records or fact rows.
Non-additive facts are usually the result of ratios or other
mathematical calculations. The only calculation that can be made
for such a fact is to get a count of the number of rows of such facts.

Table 6-27 shows examples of non-additive facts, and we discuss
the process of handling non-additive facts in 5.4.1, “Non-additive
facts” on page 191.

Derived facts Derived facts are created by performing a mathematical calculation
on a number of other facts, and are sometimes referred to as
calculated facts. Derived facts may or may not be stored inside the
fact table.

Textual facts A textual fact consists of one or more characters (codes). They
should be strictly avoided in the fact table. Textual codes such as
flags and indicators should be stored in dimension tables so they
can be included in queries. Textual facts are non-additive, but could
be used for counting.

Pseudo fact When summed, a pseudo fact gives no valid result. They typically
result when you design event-based fact tables. For more detail,
see 5.4.4, “Handling event-based fact tables” on page 205.

Factless fact A fact table with only foreign keys and no facts is called a factless
fact table. For more detail, see 5.4.4, “Handling event-based fact
tables” on page 205.


Note: Each fact in a fact table should have a default aggregation (or
derivation) rule. Each fact in the fact table should be able to be subjected to
any of the following: Sum (Additive), Min, Max, Non-additive, Semi-additive,
Textual, and Pseudo.

Table 6-27 shows the different facts and their types for the retail sales business
process. They are all true to the grain.

Table 6-27 Facts identified for the retail sales business process
Facts Fact description Type of fact Formula

Unit sales Sales price of a single product Non-additive
price

Quantity sold Quantity of each item sold Additive

Amount Total sales Additive (unit sales price) X
(quantity sold)

Discount Discount on each item Non-additive

Item cost Cost of a single item Non-additive

Total cost Cost of all items Additive (cost per item) X
(quantity sold)

Storage cost Storage cost per item Non-additive

Labor cost Labor cost per item Non-additive

Note: A fact is said to be derived if the fact can be calculated from other facts
that exist in the table, or that have been also derived. You may decide not to
include the derived facts inside the fact table and calculate these derived facts
in a reporting application.

6.5.4 Year-to-date facts
The period beginning at the start of the calendar year to the current date is called
year-to-date. For a calendar year where the starting day of the year is January 1,
the year-to-date definition is a period of time starting from January 1 to the
specified date.

Year-to-Date facts are numeric totals that consist of an aggregated total from the
start of a year to the current date. For example, assume that a fact table stores
sales data for the year 2005. The sales for each month are additive and can be


summed to produce year-to-date totals. If you create a Year-to-Date fact such as
Sales_$$_Year_To_Date, then when you query this fact in August 2005, you
would get the sum of all sales to August 2005.

Dimensional modelers may include aggregated year-to-date facts inside the fact
table to improve performance and also reduce complexities in forming
year-to-date queries. However, to avoid confusion it is typically preferred for such
facts to be calculated in the report application.

Suggested approaches for handling year-to-date facts are as follows:
OLAP-based applications
SQL functions in views or stored procedures

For our retail sales business process dimensional, we do not store any
year-to-date facts.

6.5.5 Event fact tables
Event fact tables are used to record events, such as Web page clicks and
employee or student attendance. Events do not always result in facts. So, if we
are interested in handling event-based scenarios where there are no facts, we
use event fact tables that consists of either pseudo facts or factless facts. We
explain event fact tables in more detail in 5.4.4, “Handling event-based fact
tables” on page 205.

We discuss the topic of event-based fact tables in the Identify the facts phase to
alert the reader to the considerations associated with an event-based fact table.
Some of these considerations are:
Event-based fact tables typically have pseudo facts or no facts at all.
Pseudo facts can be helpful in counting.
The factless fact event table has only foreign keys and no facts. The foreign
keys can be used for counting purposes.

Based on the current requirements, our retail sales business process
dimensional model has no event-based fact table.

6.5.6 Composite key design
A fact table primary key typically is comprised of multiple foreign keys, one from
each dimension table. Such a key is called a composite, or concatenated,
primary key. However, it is not mandatory to have all the foreign keys included in
the primary key.


Also, the combination of all foreign keys of the dimensions in the fact table will
not always guarantee uniqueness. In such situations, you may need to include a
degenerate dimension as a component in the primary key. It is mandatory that a
primary key is unique.

To determine whether a degenerate dimension needs to be included in the fact
table primary key to guarantee uniqueness, consider the dimensional model
depicted in Figure 6-31.

Grain Definition- A Single Line Item on a Grocery Bill (Receipt)

Figure 6-31 Retail sales grocery store business process

The grain definition of this schema is a single line item on a grocery bill. The date
dimension stores data at the day level and the time dimension stores the data at
the hourly level. In order to build the primary key for this fact table, assume that
we create the composite primary key of the fact table consisting of only the
foreign keys of all the dimension. Assume that the following two sales occur:
1. On October 22, 2005, Time: 8:00AM, customer C1 buys product P1 from
store S1, product P1 was supplied by supplier S1, the employee who sold the
product was E1. Customer C1 gets a bill with bill number equal to 787878.
2. On October 22, 2005, Time: 8:30AM, customer C1 buys product P1 from
store S1, product P1 was supplied by supplier S1, the employee who sold the
product was E1. Customer C1 gets a bill with bill number equal to 790051.

The above sales scenarios are represented in the fact table shown in Figure 6-32
on page 303. We observed that the UNIQUENESS of the fact table row is not
guaranteed if we choose the primary key as equal to all foreign keys of all the
dimensions. The uniqueness can only be guaranteed only if we include the
degenerate dimension (BILL_NUMBER). Of course, we correctly assume that for
every new purchase the same customer gets a new bill number.


UNIQUENESS of the Composite Primary Key is not guaranteed

Product Employee Customer Supplier DateID TimeID StoreID Customer BILL_
Key Key Key Key Mini Dim NUMBER
Key (DD)

Row 1 P1 E1 C1 S1 D1 T1 S1 CM1 787878

Row 2 P1 E1 C1 S1 D1 T1 S1 CM1 790051

8:30 AM 8:00 AM
October 22, 2005
Both 8:00 AM and 8:30 AM are represented with
only 1 single time row (T1) because the Time
Dimension granularity is at each hour.

Figure 6-32 Composite fact table design

The composite key for our fact table consists of all the foreign dimension keys
and degenerate dimension (Bill_Number).

Note: Typically the fact table primary key consists of all foreign keys of the
dimensions. However, the uniqueness of the fact table primary key is not
always guaranteed this way. In some scenarios, you may need to include one
or more degenerate dimensions in the fact primary key to guarantee
uniqueness. On the contrary, in some situations, you may observe that the
primary key uniqueness could be guaranteed by including only some of the
many foreign keys (of the dimensions) present inside the fact table.

The guarantee of uniqueness of the primary key of the fact table is determined by
the grain definition for the fact table. However, a composite key that consists of
all dimension foreign keys is not guaranteed to be unique. We discussed one
such case as depicted in Figure 6-32. We discuss more about composite primary
key design in5.4.3, “Composite key design for fact table” on page 202.

6.5.7 Fact table sizing and growth
In this activity we estimate the fact table size and also predict its growth. There
are basically two ways to calculate the growth in fact table data. They are:
Understanding the business: Assume that the retail sales business
generates a gross revenue of $100 million. Also assume that the average
price of a line item is $2. Then there would be ($100 million)/ ($2) = 50 million
line items generated per year, and 50 million rows inserted into the star


schema for the retail sales business process. Recall that the grain for the
retail sales grocery business process was a single line item on a grocery bill.
Technical perspective: The other way to calculate the exact size of fact
table growth is to calculate the size of the foreign keys, degenerate
dimensions, and facts. After we calculate the total size of the fact table
columns, we multiply it by the number of rows that could be possibly be
inserted assuming all permutations of all products that sell in all stores on all
days. We may also similarly calculate the growth of the fact table for a year.
For example, consider the star schema designed for the retail sales business
process. This is shown in Figure 6-33 on page 305. To calculate the
maximum possible size of the fact table, perform the following steps:
a. Calculate the approximate number of rows inside each of the dimensions,
assuming that the dimensions have following rows:
• Time dimension: 4 rows
• Date dimension: 365 rows for 1 year
• Product dimension: 100 rows (100 products in all)
• Store dimension: 2 rows (for 2 stores)
• Customer dimension: 1 million customers
• Customer_Mini_Dimension: 5 rows
• Supplier dimension: 50 suppliers
• Employee dimension: 10 employees
b. Calculate the base level of fact records by multiplying together the number
of rows for each dimension, calculated in the step above. For example:
• 4 x 365 x 100 x 2 x 1 million x 5 x 50 x 10 = 730000000 rows or 730
million rows.
Of course this is a huge number of rows if every product was sold in every
store by every employee to every customer.
c. We calculate the maximum fact table size growth as follows:
• Number of foreign keys = 8
• Number of degenerate dimensions = 1
• Number of facts = 8
Assuming that the fact table takes 4 bytes for an integer column,
Total size of 1 row =(8+1+8) x 4 bytes= 68 bytes.
So, the maximum data growth for this dimensional model is:
730 million rows x 68 bytes (size of 1 row) = 45 GB


Figure 6-33 Retail sales business process dimensional model

This mathematical fact table growth calculation helps the DBA to calculate the
approximate and maximum growth of the fact table. This way the DBA can be
made aware of the possible growth of the dimensional model and consider
potential performance tuning measures.

6.6 Verify the model
The primary focus of this phase is to test the dimensional model to see if it meets
the business requirements. The dimensional model at this point would contain no
data. The testing occurs to see if the existing model can answer all questions
posed during the requirement gathering phase.

6.6.1 User verification against business requirements
Before we complete the dimensional design and pass it on to the ETL team to
begin the ETL design, we must verify the model against the business
requirements analyzed in 6.2.7, “Requirements analysis” on page 242.
Table 6-28 shows the results of the verification.

Table 6-28 Validate model against requirements
Q. Business requirement Meets?
no.

Q1 What is the average sales quantity this month for each Yes
product in each category?


Q. Business requirement Meets?
no.

Q2 Who are the top 10 sales representatives and who are their Yes
managers? What were their sales in the first and last fiscal
quarters for the products they sold?

Q3 Who are the bottom 20 sales representatives and who are Yes
their managers?

Q4 How much of each product did U.S. and European customers Yes
order, by quarter, in 2005?

Q5 What are the top five products sold last month by total Yes
revenue? By quantity sold? By total cost? Who was the
supplier for each of these products?

Q6 Which products and brands have not sold in the last week? Yes
The last month?

Q7 Which salespersons had no sales recorded last month for Yes
each of the products in each of the top five revenue

Q8 What was the sales quantity of each of the top five selling Yes
products on Christmas, Thanksgiving, Easter, Valentine's
Day, and Fourth of July?

Q9 What are the sales comparisons of all products sold on Yes
weekdays compared to weekends? Also, what was the sales
comparison for all Saturdays and Sundays every month?

Q10 What are the top 10 and bottom 10 selling products each day Yes
and week? Also at what time of the day do these sell?
Assuming there are 5 broad time periods - Early morning
(2AM - 6AM), Morning (6AM - 12PM), Noon (12PM - 4PM),
Evening (4PM - 10PM) Late night Shift (10PM - 2AM)

At this point the dimensional modelers should have the report writers build
pseudo logic to answer questions Q1 to Q10 in Table 6-28 on page 305. This
process helps in the verification process from a report feasibility standpoint.
Then, if there are any missing attributes from the dimension tables, they can be
added.

In addition to validating the model against the requirements, also confirm
requirements for handling history. Questions that you may need to consider and
validate are shown in Table 6-29 on page 307. You also need to validate these
history preserving requirements against the dimensional model.


Table 6-29 Questions relating to maintaining history
History requirement Action

Employee changes from Region A to Region B? Overwrite or maintain history

Employee changes Manager Overwrite or maintain history

Product changes from Category A to Category B? Overwrite or maintain history

More questions Overwrite or maintain history

6.7 Physical design considerations
In this phase we focus on physical design considerations, as depicted in the
DMDL in Figure 6-34.

Determine All
Dimensions
Processes

Identify Conformed




Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


The primary focus of this phase is to design the strategy to handle the following:
Aggregation


Aggregate navigation
Indexing
Partitioning

6.7.1 Aggregations
Aggregates provide a restricted list of the key columns of a fact table and an
aggregation - generally a SUM() - of some or all of the numeric facts. Therefore,
a fact table with eight dimensional keys and four numeric facts can be
aggregated to a summary table of three dimensional keys, plus two facts.

In simple terms, aggregation is the process of calculating summary data from
detail base level fact table records. Aggregates are a powerful tool for
increasing query processing speed in dimensional data marts. The aggregation
is primarily performed by using attributes of a dimension which are a part of a
hierarchy.

A dimension table is made of attributes, and consists of hierarchies. A hierarchy
is a cascaded series of many-to-one relationships. Some attributes inside the
dimension table typically belong to one or more hierarchy.

Each attribute that belongs to a hierarchy associates as a parent or child with
other attributes of the hierarchy. This parent-child relationship provides different
levels of summarization. The various levels of summarization provide the
business user the ability to drill up or drill down in the report. Highly aggregated
data is faster to retrieve than detailed atomic level data. And, the fact table
typically occupies a large volume of space when compared to the aggregated
data.

The lowest level of aggregation, or the highest level of detail, is referred as the
grain of the fact table. The granularity of the dimension affects the design of data
storage and retrieval of data. One way to identify candidates for aggregates is to
use automated tools that are available in the market, or write your own
applications to monitor the SQL generated in response to business queries and
identify predictable queries that can be processed using precomputed
aggregates rather than ad hoc SQL.

Costs associated with aggregation
Aggregating detailed atomic fact tables improves query performance. However,
there are costs associated with aggregation, such as:
Storage cost
Cost to build and maintain the ETL process to handle the aggregated tables


Aggregations to avoid
Aggregation should not be considered as a substitute for reducing the size of
large detailed fact tables. If data in the fact table is summarized, detailed
information in the form of dimensions and facts is often lost. If the business
needs detailed data from a summarized fact table, they simply cannot get it.
They would need to look for the details back in the source OLTP system that
provided the aggregated fact table data. Of course if the business has to go back
to the source OLTP systems to get the answers, then the whole purpose of
building a dimensional model should be questioned. We discuss the importance
of having a detailed atomic grain in more detail in 5.2.2, “Importance of detailed
atomic grain” on page 122.

Another important point while creating aggregates is to avoid mixing aggregated
data and detailed data by including year-to-date aggregated facts with the
detailed facts. This is primarily because year-to-date facts are additive, so it
could result in accidental miscalculations. For more on this, see 6.5.4,
“Year-to-date facts” on page 300.

Suggested approaches for aggregation
In this section, we provide guidelines for preparing aggregate tables based on
detailed (highly atomic) base-level fact tables:
1. Identify all dimensions and their hierarchies from the base level atomic
dimensional model. These dimensions and hierarchies are identified from the
base-level atomic dimensional model, depicted in Figure 6-35.

Figure 6-35 Retail sales business process dimensional model

Table 6-30 on page 310 shows all dimensions and all hierarchies
associated with each of these dimensions, including their levels.


Table 6-30 Dimensions and their hierarchies
Name Hierarchy Type

Date Calendar Hierarchy: Calendar Year → Calendar Balanced

Fiscal Hierarchy: Fiscal Year → Fiscal Quarter → Fiscal
Month → Fiscal Day

Time No Hierarchy

Product CategoryName → BrandName → ProductName Balanced

Employee No Hierarchy

Supplier Supplier Country → Supplier Region → Supplier City Balanced

Customer Customer Country → Customer Region → Customer Balanced
City

Store Store Country → Store State → Store Region → Store Balanced
Area

Customer_ No Hierarchy
Mini_
Dimension

2. Identify all possible combinations of these hierarchy attributes which are used
together by business for reporting.
In this step we identify all attributes from the hierarchies (see Table 6-30) to
determine which of these are used frequently together. This is extremely
critical, particularly if there are a huge number of dimensions with several
hierarchies having several attributes in them.
Assume from studying the business requirements, we find that the attributes
(relating to Date, Product, and Store dimensions) are used together. These
attributes are shown in Table 6-31.

Table 6-31 Candidate dimensions and their hierarchies for aggregation
Name Hierarchy Type

Date Calendar Hierarchy: Calendar Year → Calendar Balanced

Product CategoryName → BrandName Balanced

Store Store Country → Store State → Store Region Balanced


3. Calculate the number of values of each attributes selected for aggregation, in
Table 6-31 on page 310.
It is important to consider the number of values for attributes that are
candidates for aggregation. For example, suppose that you have 1 million
values for the lowest level product name (assuming that the store sells 1
million different types of products). We do not include the product name as a
candidate to be aggregated because we are dealing with a huge number of
rows. We use the brand and category attributes of the product hierarchy
because the brand has about 10 000 values and category has about 1 500
values. Moreover, the business is more interested in understanding the sales
of brands and categories of products, than individual products.
The number of values each attribute has is indicative of whether the attribute
is a candidate to be aggregated. For example, if we find that a low level
member in the hierarchy has been included and has a huge number of
members (values), then we may drop that particular attribute and choose a
higher level attribute which would ideally have fewer values.
Table 6-32 shows attributes with examples of the number of possible values.

Table 6-32 Dimension attributes with values
Name Hierarchy Type

Date Calendar Hierarchy: Calendar Year (12 Years) → Balanced
Calendar Month (12 Month Names) → Calendar Week
(52 Weeks)

Product CategoryName (1,500 Categories) → BrandName Balanced
(10,000 Brands)

Store Store Country (50 Countries) → Store State (50 Balanced
States) → Store Region (6000 Regions)

4. Validate the final set of attribute candidates and build the aggregated
dimensional model.
In this step we validate the final dimensional attributes identified in Table 6-32.
We may also decide to drop one or more attributes if we think that the
attribute has a large number of values. The final set of the aggregated
dimensional model is shown in Figure 6-36 on page 312. Depending upon the
need, one or more such aggregated models can be created.


Base Level Aggregate Fact table

Grain: A single line item on Grocery Receipt

View created from Date Dimension

View created from Product Dimension View created from Store Dimension

Aggregated Fact table

Grain: Product by Brand by Week by Store Region

Figure 6-36 Aggregated dimensional design

6.7.2 Aggregate navigation
Aggregate navigation is considered an advanced concept in data warehousing,
although the concept is actually simple.

What is aggregate navigation?
Aggregate navigation is software that intercepts SQL requests and transforms
them to use the best available aggregates. Aggregate navigation is software that
intercepts an SQL request (say SQL1) and transforms it to a new SQL statement
(assume SQL2) to be used against a particular aggregate. Aggregate navigation
is a technique that involves redirecting SQL queries to appropriate precomputed
aggregates. The concept behind this technique is that the SQL queries are
intercepted and rewritten to take the best advantage of aggregate tables
available inside the warehouse. As shown in Figure 6-37 on page 313,
Aggregate navigator
is a software application that sits between the user and data warehouse.


Detailed Day level Star
Aggregate Navigator
(Accepts and
SQL Transforms SQL1 to
Aggregate
SQL2), plus redirects
request the SQL2 to Aware
(SQL1) appropriate Agg. (SQL2)
Table)
Week Level Star

SQL1 Statement SQL2 Statement
requesting data at requesting Year data
Year Level from Month Level Star
Month Level Star

Figure 6-37 Aggregate navigator

The Aggregate navigator accepts the SQL1 statement and analyzes it. For
example, as shown in Figure 6-37, the aggregator accepts the SQL1 statement
and sees that the user requests data at the year level. Now there are 3 star
schemas in Figure 6-37 from which the user can get the data for year level. The
aggregate navigator chooses the month level star schema. Choosing the month
level star schema instead of the day or week level star schema improves the
performance of the query. This is because fewer rows in month need to be
summarized to get to the year level data. Had we chosen the day or week, more
numbers of rows would have to be summarized to get to the same result.

Aggregate navigation helps to optimize the queries by choosing the most
appropriate aggregated schema to get to the desired results. Several database
vendors have chosen to implement aggregate navigation software in their
databases. In the next section, we discuss optimization with IBM DB2.

DB2 Optimizer and DB2 Cube Views
DB2 Cube Views V8.2 works together with partner BI tools to accelerate OLAP
queries. Queries from many data sources, supported by IBM WebSphere
Information Integrator, can be accelerated by DB2 Cube Views. DB2 Cube Views
works by using cube meta data to design specialized summary tables containing
critical dimensions and levels — or slices — of the cube. The DB2 optimizer
rewrites incoming queries, and transparently routes eligible queries to the
appropriate summary tables for significantly faster query performance. The Cube
Views-created summary tables, also called Materialized Query Tables (MQTs),
can accelerate all SQL-based queries to the data warehouse, not just those


using a particular tool or interface. The DB2 optimizer rewrites and redirects the
incoming queries as shown in Figure 6-38. For a more detailed discussion how
the DB2 SQL optimizer navigates between the various base and summary
tables, refer to 5.5.1, “DB2 Optimizer and MQTs for aggregate navigation” on
page 212.

Optimizer transparently
rewrites incoming queries 1. Base data?
Fetch & Sum
n rows

Select
Sum (Sales)... 2. Partial Aggregate?
Where Product in ( 'Cola','Root Fetch & Sum
Beer')
Group by Product

DB2 Optimizer 3. Full Aggregate?
Simple Fetch

Figure 6-38 DB2 Optimizer

6.7.3 Indexing
Indexing of database tables can improve query performance, for example in
situations where a number of tables are joined or a subsets of records are
retrieved.

In general, indexes keep information about where the record (row) with a
particular key or attribute is physically located in a table. If a table is indexed, the
database engine does not need to scan the entire table to look for a row with the
specific key or attribute. Instead of scanning the entire table, the database reads
the index entries and directly retrieves the rows from the table. It is typically
better to scan indexes than the entire table.

Two types of indexes
In this section we provide a brief description of index types. They are:
Unique: Each row is uniquely identified by the value of the index.
Non-unique: The value of the index can identify one or more rows in the
table.


In addition, indexes can differ based on their physical organization:
B-Tree: Leaf nodes contain the value of the index and a pointer to the
physical row. Use this type of index for attributes with high cardinality, such as
PK or FK.
Bitmap: Each unique value has a corresponding bitmap where each bit
represents each row in the table, and value 1 in the bitmap means the
corresponding row has this value. Use this type of index for attributes with low
cardinality. An example is a field storing a status such as active or inactive.

Clustering
B-tree indexes can also be clustered. Here physical rows are sorted in order of
the index, which can dramatically decrease the number of I/O operations when
tables are read in order of index.

Clustered indexes are most efficient if no modification to tables are performed
after the clustered index is created. If new rows are added or old rows are
updated and/or deleted, the efficiency of a clustered index decreases because
the rows are no longer in the physical order of the index. To correct this, you
should recreate clustered indexes from time to time. Although, this can be a
rather expensive operation if the tables become very large. In some RDBMS,
indexes can be partitioned in a similar way that tables are partitioned. That is,
some part of the indexes can be put in separate logical spaces. This can also
improve query response time since some parts of index (index partitions) can be
completely eliminated from the scan.

Index maintenance
In addition to the advantages that indexes typically bring, there also costs.
Indexes require additional space and also increase processing time when
records are inserted, deleted, or updated. For example, when a row is updated,
the RDBMS must also delete the old index entries corresponding to old values
and add new index entries for new values of updated columns in the row.

Note: Activities that may help in maintaining indexes are:
Use the run statistics utility for all tables and all indexes.
Reorganize clustered indexes.

Indexes for star schema
To summarize, the star schema model consists of dimension and fact tables.
A fact table contains facts and foreign keys (FK) of dimensions. The fact table
foreign key (FK) points to corresponding primary keys (PK) in the dimension
tables. Dimension tables consist of a primary key and attributes that describe the
measures of the fact table.


Fact tables may have one composite primary key to identify a unique set of
measures and foreign keys (FK). If the set of foreign keys is not unique, we can
introduce one or more degenerate dimensions, which, along with the FKs, will
make it a unique key.

Note: The RDBMS optimizer decides when to use and when to not use
indexes, using table and index statistics to make the decision.

Indexes for dimensions
Indexing dimension tables is an iterative activity. Dimension tables may have
many indexes, and those listed below should be considered as mandatory:
One unique B-tree index for the primary key (PK) identifying the row in
dimension table.
For each attribute corresponding to a dimension hierarchy level one, a
non-unique index should be used - unless it is snowflaked.
One or more additional non-unique indexes on attributes, to be used as filters
in queries.
Non-unique B-tree index for the foreign key of a snowflaked dimension table.

Indexes for a fact table
First consider if we need to ensure uniqueness within the fact table. If yes, then
create a unique B-tree index from foreign keys, and degenerate dimensions if
applicable.

Then create a non-unique non-clustering B-tree index for each foreign key and
degenerate dimensions.

Do not index numeric facts attributes. If you have need to filter facts by some
numeric, you may introduce a new dimension such as a fact value band. This
table may store band range of values, such as100-9900 and 9901-99992. You
can also address this need with an OLAP reporting application. However, we
show a technique that can be used to index facts. It is discussed in more detail in
“Selective indexes” on page 318

Note: The type of indexes you choose depends on attribute cardinality and on
the capabilities of underlying indexes. We suggest using B-tree for high
cardinality attributes and Bitmap for low or medium cardinality attributes. For
example, if attribute values are highly duplicated, such as gender, status, or
color attributes, consider using bitmap indexes. Bitmap indexes, if well
designed, can dramatically improve query response times. Multidimensional
clustered (MDC) or block indexes (see “MDC indexes” on page 318) are
suitable for indexing fact table FKs and degenerate dimensions.


Vendor specific indexing techniques
Several vendors offer special proprietary types of indexes, designed for
optimizing queries, for example, the multidimensional clustered (MDC) indexes
in IBM DB2 or generalized key (GK) indexes in IBM Informix Extended Parallel
Server (XPS).

Generally these indexes are defined either on one, or more than one, table (GK
indexes) or on clusters of values from various attributes (MDC indexes). The
primary goal of such advanced indexing is to eliminate unnecessary joins and
table scans, and to reduce query response time. Generalized key indexes
provide the following three types of advanced indexes that can improve OLAP
query performance:
Join indexes
Virtual indexes
Selective indexes

Join index
This capability allows you to create an index on a fact table that contains
attributes of the dimension table. Using this index can eliminate the need to join
the tables.

For example, if we have a SALES fact table, with item and sales amount, and the
dimension table PRODUCT, describing items that have attribute BEVERAGE,
then you can create an index on the fact table containing BEVERAGE of each
item. As an example, you can do that with the following command:

CREATE GK INDEX type_of_sale ON SALES
(SELECT PRODUCT.BEVERAGE FROM SALES, PRODUCT
WHERE SALES.PRODUCT_KEY = PRODUCT.PRODUCT_KEY)

Then when querying for sales amounts of certain product types, the join of the
SALES table and PRODUCT dimension can be eliminated. For example,
consider the following query:

SELECT SUM(SALES_AMOUNT) FROM SALES, PRODUCT
WHERE SALES.PRODUCT_KEY = PRODUCT.PRODUCT_KEY AND PRODUCT.BEVERAGE =
"SODA"

Virtual indexes
This capability allows you to create an index not only on one or more columns,
but also on expressions based on those columns. This enables fast queries on
computed values, and can save disk space as well.

For example, if you have a table with columns UNITS and UNIT_COST, you can
create an index on the COST_OF_SALE as follows:


CREATE GK INDEX I_cost_of_sale on SALES (SELECT
SALES.UNITS*SALES.UNIT_COST FROM SALES)

This index then speeds up queries such as the following:

SELECT * FROM SALES WHERE UNITS*UNIT_COST >60

Note: Virtual indexes are useful when a fact table contains derived facts.

Selective indexes
This capability enables you to index only part of a large table, saving disk space
and speeding up the queries.

For example, you can create an index such as the following:

CREATE GK INDEX since1995 ON SALES (SELECT AMOUNT FROM SALES JOIN
DATE_DIM ON SALES.DAY_DIM_KEY=DATE_DIM.DAY_DIM_KEY WHERE DATE_DIM.YEAR
>=1995)

Then you can create a fast query such as this:

SELECT * FROM SALES WHERE amount >500 AND YEAR >=1995

MDC indexes
IBM DB2 V8.x enables you to create advanced clustered indexes, also known as
multidimensional clustered (MDC) indexes, which are optimized for OLAP
queries. MDC indexes are based on blocks (extents), while traditional indexes
are based on rows. Therefore in an MDC index, the leaf entries point to blocks of
pages which have rows with the same value of the columns. Traditional index leaf
entries point to a physical row on a page. MDC indexes are clustered as rows
and grouped by values of columns in the indexes. At least one block is created for
each unique value of a dimension.

A cell is a unique combination of values across all dimensions. Data in the MDC
tables is clustered via cells.

When creating MDC tables, a blocked index is created for each dimension and
one covering all dimensions. Blocked indexes optimize query response time
because the optimizer can quickly identify blocks of pages with required values in
particular columns. Further, the maintenance of MDC tables is simplified as no
data reorganization is needed. If there is no cell for new or updated rows or if all
blocks for this cell are full, another block is allocated and a new entry for this
block is added into block indexes.


Note: When designing MDC tables, the cost of disk space should be
considered. Although block indexes demand less disk space, the minimal disk
space needed for data in MDC tables is the cartesian product of all
cardinalities of all dimensions multiplied by block (extent) size.

You can find a more detailed example of indexing in 5.5.2, “Indexing for
dimension and fact tables” on page 218, which includes the following topics:
Indexes for the star schema (includes indexing for dimensions and facts)
Differences between indexed and non-indexed star schemas
Access plans for querying an indexed star schema with SQL
Access plans for querying a non-indexed star schema with SQL

6.7.4 Partitioning
Partitioning a table divides the table by row, by column, or both. If a table is
divided by column, it is said to be vertically partitioned. If by row, it is said to be
horizontally partitioned. Partitioning large fact tables improves performance
because each partition is more manageable and provides better performance
characteristics. Typically, you partition based on the transaction date dimension
in a dimensional model. For example, if a huge fact table has billions of rows, it
would be ideal for one month of data to be assigned its own partition.

Partitioning the data in the data warehouse enables the accomplishment of
several critical goals. For example, it can:
Provide flexible access to data
Provide easy and efficient data management services
Ensure scalability of the data warehouse
Enable elements of the data warehouse to be portable. That is, certain
elements of the data warehouse can be shared with other physical data
warehouses, or archived on other storage media.

Guidelines for partitioning star schema data are explained below:
A large fact table with a large number of rows should be partitioned based on
the date dimension.
Depending upon the growth of the fact table, an individual partition may be
created for each unique months of data or each unique quarter of data.
Large dimension tables, such as customer table of a government agency,
having millions of customers may also be partitioned.


We usually partition large volumes of current detail data by dividing it into smaller
segments. Doing that helps make the data easier to:
Restructure
Index
Sequentially scan
Reorganize
Recover
Monitor

In addition, other advantages of partitioning are:
Improved response times because the SQL query only accesses the partition
of data needed to answer the query.
Partitions can be more easily maintained compared to one large table.
Involves no extra cost, as most partitioning capability is typically included as
part of the RDBMS.

Every database management system (DBMS) has its own specific way of
implementing physical partitioning, and they all can be quite different. For
example, it is an important consideration whether or not the DBMS also supports
partition indexing. Instead of DBMS or system level partitioning, you can
consider partitioning by application. This would provide flexibility in defining data
over time, and portability in moving to the other data warehouses. Notice that the
issue of partitioning is closely related to multidimensional modeling, data
granularity modeling, and the capabilities of a particular DBMS to support data
warehousing.

6.8 Meta data management
Figure 6-39 on page 321 shows that the meta data management block spans the
entire DMDL. That is, every phase of dimensional modeling produces some level
of meta data.



Granularity
Processes

Identify Conformed



Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


In the traditional development cycle, a model sees only sparse use after
completion. This is typically when changes need to be made, or when other
projects require the data. In the data warehouse, however, the model is used on
a continuous basis. The users of the data warehouse constantly reference the
model to determine the data they want to use in their data analysis. The rate of
change of the data structure in a data warehouse is much greater than that of
operational data structures. Therefore, the technical users of the data warehouse
(administrators, modelers, and designers, as examples) will also use the model
on a regular basis.

This is where the meta data comes in. Far from just a pretty picture, the model
must be a complete representation of the data being stored.

To properly understand the model, and be able to confirm that it meets
requirements, you must have access to the meta data that describes the
dimensional model in business terms that are easily understood. Therefore,
non-technical meta data should also be documented in addition to the technical
meta data.


At the dimensional model level, a list should be provided of what is available in
the data warehouse. This list should contain the models, dimensions, facts, and
measures available as these will all be used as initial entry points for data
analysis.

For each model, provide a name, definition, and purpose. The name simply gives
something to focus on when searching. Usually, it is the same as the fact. The
definition identifies what is modeled, and the purpose describes what the model
is used for. The meta data for the model should also contain a list of dimensions,
facts, and measures associated with it, as well as the name of a contact person
so that users can get additional information when there are questions about the
model.

A name, definition, and aliases must be provided for all dimensions, dimension
attributes, facts, and measures. Aliases are necessary because it is often difficult
to come to agreement on a common name for any widely used object. For
dimensions and facts, a contact person should be provided.

Meta data for a dimension should also include hierarchy, change rules, load
frequency, and the attributes, facts, and measures associated with the
dimension. The hierarchy defines the relationships between attributes of the
dimension that identify the different levels that exist within it. For example, in the
seller dimension we have the sales region, outlet type (corporate or retail), outlet,
and salesperson, as a hierarchy. This documents the roll-up structure of the
dimension. Change rules identify how changes to attributes within a dimension
are dealt with. In some instances, these rules can be different for individual
attributes. Record change rules with the attributes when this is the case. The
load frequency allows the user to understand whether or not data will be
available when needed.

The attributes of a dimension are used to identify which facts to analyze. For
attributes to be used effectively, meta data about them should include the data
type, domain, and derivation rules. At this point, a general indication of the data
type (such as character, date, and numeric) is sufficient. Exact data type
definitions can be developed during design. The domain of an attribute defines
the set of valid values. For attributes that contain derived values, the rules for
determining the value must be documented.

Meta data about a fact should include the load frequency, the derivation rules
and dimensions associated with the fact, and the grain of time or date for the
fact. Although it is possible to derive the grain of time for a fact through its
relationship to the time dimension, it is worthwhile explicitly stating it here. It is
essential for proper analysis that this grain be understood.


6.8.1 Identifying the meta data
The following meta data is collected during the different phases of the DMDL:
Identify business process: The output of this phase results in the
requirements gathering report. This report primarily consists of the business
requirements for the selected business for which you will design the
dimensional model. In addition to this, it also consists of various business
processes, owners, source systems involved, data quality issues, common
terms used across business processes and other business-related meta data.
Identify the grain: The output of this phase results in the grain definition
report, which consists of one or multiple definitions of the grain for the
business process for which the dimensional model is being designed. Also
the type of fact table (transaction, periodic, or accumulating) being used is
mentioned. The grain definition report also includes high level preliminary
Identify the dimensions: The meta data documented for this phase contains
the information as shown Table 6-33:

Table 6-33 Identify the dimensions phase meta data
Dimension meta data Description

Name of dimension Name of the dimension table.

Business definition Business definition of the dimension.

Alias Specifies the other known name by which the business
users know the dimension.

Hierarchy Defines the hierarchies present inside the dimension, such
as balanced, unbalanced, or ragged.

Change rules Specify how to handle slowly changing dimension (type-1,
type-2, or type-3) or fast changing dimension.

Load frequency The frequency of load for this dimension, such as daily,
weekly, or monthly.

Load statistics Consists of meta data such as:
→ Last load date: N/A
→ Number of rows loaded: N/A

Usage statistics Consists of meta data such as:
→ Average Number of Queries/Day: N/A
→ Average Rows Returned/Query: N/A
→ Average Query Runtime: N/A
→ Maximum Number of Queries/Day: N/A
→ Maximum Rows Returned/Query: N/A
→ Maximum Query Runtime: N/A



Archive rules Specifies whether or not data is archived.

Archive statistics Consists of meta data such as:
→ Last Archive Date: N/A
→ Date Archived to: N/A

Purge rules Specifies any purge rules. For example, customers who
have not purchased any goods from the store in the past 48
months will be purged on a monthly basis.

Purge statistics Consists of meta data such as:
→ Last Purge Date: N/A
→ Date Purged to: N/A

Data quality Specifies data quality checks. For example, when a new
customer is added, a search determines if the customer
already does business with another location. In rare cases
separate branches of a customer are recorded as separate
customers because this check fails.

Data accuracy Specifies the data accuracy. For example, incorrect
association of locations of a common customer occur in
less than .5% of the customer data.

Key The key to the dimension table is a surrogate key.

Key generation method This meta data specifies the process used to generate a
new surrogate key for a new dimension row. For example,
when a customer is copied from the operational system,
the translation table (a staging area persistent table) is
checked to determine if the customer already exists in the
dimensional model. If not, a new key is generated and the
key along with the customer ID and location ID are added
to the translation table. If the customer and location already
exist, the key from the translation table is used to determine
which customer in the dimensional model to update.



Source This includes the following meta data:

→ Name of the source system table: <Table Name>

→ Conversion rules: This specifies how the insert/update
to the dimension table occurs. For example, rows in each
customer table are copied on a daily basis. For existing
customers, the name is updated. For new customers, once
a location is determined, the key is generated and a row
inserted. Before the update/insert takes place, a check is
performed for a duplicate customer name. If a duplicate is
detected, a sequence number is appended to the name.
This check is repeated until the name and sequence
number combination is determined to be unique. Once
uniqueness has been confirmed, the update/insert takes
place.

→ Selection logic: Only new or changed rows are
selected.

Conformed dimension This specifies if the dimension is a conformed dimension.

Role-playing dimension This specifies if the dimension is being implemented using
the role-playing concept.



Attributes The meta data for all the dimension attributes includes the
(All columns of a following:
dimension)
→ Name of the attribute: <Attribute Name>

→ Definition of the attribute: Attribute definition

→ Alias of the attribute: <Attribute Name>

→ Change rules for the attribute: For example, when an
attribute changes, then use Type-1, Type-2, or Type-3
strategy to handle the change.

→ Data Type for the attribute: Data type, such as Integer
or Character.

→ Domain values for the attribute: Domain range such as
1-99.

→ Derivation rules for the attribute: For example, a system
generated key of the highest used customer key +1 is
assigned when creating a new customer and location entry.

→ Source: Specifies the source for this attribute. For
example, for a surrogate key, the source could be a system
generated value.

Facts This specifies the facts that can be used with this
dimension.
Note: Semi-additive facts are additive across only some
dimensions.

Subsidiary dimension This specifies any subsidiary dimension associated with
this dimension.

Contact person This specifies the contact person from the business side
responsible for maintaining the dimension.

The meta data information shown in Table 6-33 on page 323 needs to be
captured for all dimensions present in the dimensional model.
Identify the facts: The meta data documented for this phase contains the
information as shown Table 6-34.

Table 6-34 Identify the facts phase meta data
Fact table meta data Description

Name of fact table The name of the fact table.



Business Definition The business definition of the fact table.

Alias The alias specifies another name by which the fact table
is known.

Grain This specifies the grain of the fact table.

Load frequency The frequency of load for this fact table such as daily,
weekly or monthly.

Load statistics Consists of meta data such as:

Usage statistics Consists of meta data such as:

Archive rules Specifies whether or not the data is archived. For
example, data will be archived after 36 months on a
monthly basis.

Archive statistics Consists of meta data such as:

Purge rules Specifies any purge rules. For example, data will be
purged after 48 months on a monthly basis.

Purge statistics Consists of meta data such as:

Data quality Specifies quality checks for the data. For example,
assume that we are designing a fact table for an
inventory process. Inventory levels may fluctuate
throughout the day as more stock is received into
inventory from production and stock is shipped out to
retail stores and customers. The measures for this fact
are collected once per day and thus reflect the state of
inventory at that point in time, which is the end of the
working day.



Data accuracy This specifies the accuracy of fact table data. For
example, assume that we are designing a fact table for
the inventory process. We may conclude that the
measures of this fact are 97.5% accurate at the point in
time they represent. This may be based on the results
of physical inventories matched to recorded inventory
levels. No inference can be made from these measures
as to values at points in time not recorded.

Grain of the date Specifies the grain of the date dimension. For example,
dimension the date dimension may be at the day level.

Grain of the time Specifies the grain of the time dimension. For example,
dimension the time dimension may be at the hourly level.

Key The key of the fact table typically consists of
concatenation of all foreign keys of all dimensions. In
some cases, the degenerate dimension may also be
concatenated to guarantee the uniqueness of the
primary composite key.

Key generation method The key generation method specifies how all the foreign
keys are concatenated to create a primary key for the
fact table. Sometimes a degenerate dimension may be
needed to guarantee its uniqueness. This is shown in

Source The meta data for the source includes the following:

→ Name of the Source: <Source Name>

→ Conversion rules for the source: Rules regarding the
conversion. For example, each row in each inventory
table is copied into the inventory fact on a daily basis.

→ Selection Logic: The selection logic behind selecting
the rows.

Facts This specifies the facts involved in the fact table. They
could be:
→ Additive
→ Non-additive
→ Semi-additive
→ Pseudo
→ Derived
→ Factless fact
→ Textual



Conformed fact This specifies whether or not there are any conformed
facts in the fact table.

Dimensions This specifies the dimensions that can validly use these
facts. Note: Some facts are semi-additive and can only
be used across certain dimensions.

Contact person This specifies the contact person from the business side
responsible for maintaining the fact table.

Verify the model: This phase involves documenting meta data related to the
testing (and its result) done on the dimensional model. Any new requests
made or changes to the requirements are documented.
Physical Design Considerations: The meta data documented for this phase
contains the information as shown in Table 6-35.

Table 6-35 Physical design considerations meta data
Physical design Description
consideration meta data

Aggregation The aggregation meta data includes the following:

→ Number of aggregate tables

→ Dimension tables involved in creating aggregates
→ Dimension hierarchies involved in creating
aggregation

→ Fact table and facts involved in creating aggregates.

Other information relating to the aggregate tables
includes the following:
→ Load frequency
→ Load statistics
→ Usage statistics
→ Archive rules
→ Archive statistics
→ Purge rules
→ Purge statistics
→ Data quality
→ Data accuracy

Indexing This specifies the indexing strategy used for dimension
and fact tables.


6.9 Summary
In 6.2.2, “Identify business process” on page 234, we created an enterprise
business process list for which there are business needs to build a data mart or
dimensional model.

After identifying all the business processes, we assessed each of them for a
number factor, such as:
Complexity of the source systems of the business process
Data availability of these systems
Data quality of these systems
Strategic business significance of the business process

After having assessed the various business processes, we developed the final
prioritization list shown in Table 6-36.

Table 6-36 Enterprise-wide business process priority listing
Name Complexity Availability Quality Significanc Points
e

Retail sales Low (3) High (3) High (3) High (6) 15

Finance High (1) High (3) Medium (2) Medium (4) 10

Servicing Low (3) High (3) Medium (2) High (6) 14

Marketing Medium (2) Medium (2) Medium (2) Medium (4) 10

Shipment Low (3) Low (1) High (3) Low (2) 9

Supply Medium (2) Low (1) Medium (2) Low (2) 7
management

Purchase High (1) Medium (1) Low (1) Medium (4) 7
order

Labor Low (3) Low (1) Low (1) High (2) 7

Table 6-36 helped us to prioritize the business processes for which we can
design the dimensional models. In this chapter, we chose the top priority retail
sales business process and designed a data mart using the DMDL. After we
finish with the design of one data mart, we can then move to the next priority data
mart from Table 6-36.

The DMDL helps us segment the larger task of building data marts for the entire
organization by choosing a step by step approach of handling each business


process, one at a time. This approach enables you to start small, and complete
the project in planned phases.

In addition, you remember that data warehousing itself is a process. That is,
typically you are never really finished. You will undoubtedly continue to add
people, processes, and products to your enterprise. Or, specifications about
those people, processes, and products will change over time. As examples,
people get married, have children, and move. This requires modifying, for
example, dimensional data that describes those people.

So, as with any data warehousing project, building dimensional models is a
continuous and ongoing project. Having a defined DMDL, and using suggestions
in this redbook, can help you to build a structured, controlled, planned, and
cost-effective approach to building several dimensional models for your data
warehousing environment which are integrated by using conformed dimensions
and conformed facts.


7

Chapter 7. Case Study: Dimensional
model development
In this chapter we describe in detail how to methodically apply the Dimensional
Model Design Life Cycle (DMDL) when designing a dimensional model.

The goal of this case study is to show the usage of the DMDL technique when
applied to specific business requirements. For that, we will follow the DMDL
technique step by step, describing our assumptions and explaining the decisions
taken during the process.

The design of the dimensional model was developed using the IBM Rational®
Data Architect product.


7.1 The project
In the following sections we describe a fictitious project as a case study. We
define the scope and objectives, and relate them to the requirements of a
fictitious client. This is a good test case to see that a good dimensional model
can be achieved by following the DMDL.

7.1.1 The background
In this fictitious case study, the client has been suffering during many years of not
having a data warehouse solution. With many systems spread through many
countries, and many databases created for ad hoc and local transaction
management purposes, the client was not able to effectively consolidate the data
in a manner that satisfied the reporting requirements. In order to solve this
problem, the client decided to build a data warehouse. After some time, and a
few iterations, the client finished implementing a data warehouse. It had all the
subject areas that, from the business point of view, represented significant
strategic value.

The new data warehouse, depicted in Figure 7-1 on page 335, provided a
consolidated view of the company data that they were looking for, and satisfied
most of their stated requirements for standard reporting. Initially, the reporting
needs were more of a static nature. In other words, once reports were written,
they did not need to be rewritten or changed. However, as the number of users of
the data warehouse grew, the reporting requirements became more complex in
nature. The client soon realized that the reports started changing dynamically.

These changes were primarily because the business realized the potential of
data available in the data warehouse and the client wanted to analyze
information in a number of new and different ways. In other words, the value of
analyzing the data was realized and thus emerged the requirement for more ad
hoc reporting. This change in reporting business requirements resulted in more
dependence on the reports, and an increased requirement for programmers to
develop them. The increase in the number of reports and the increase in the
volume of data accessed by them, started to drastically impact the performance
of the reports. Upon analysis, the client determined this was primarily because
the data warehouse data was in 3NF, and based on an E/R model.


RDBMS Systems

Business Users

Report/query Analysts

ERP Aplication
Servers

EDW
E/R Data Model

CRM system

Figure 7-1 Client Data Warehouse Architecture

In addition, because of the complexity of the E/R model, it was extremely difficult
for the business users to query the data warehouse on an ad hoc basis.

To solve the performance problem, the client created a new set of denormalized
tables in the data warehouse. The purpose was to help the users get a faster
response time when accessing their reports. In addition, it was also expected
that the users would be able to do ad hoc reporting against this new layer
branching out from the warehouse. This process of denormalizing several tables
was done on a report-by-report basis. Soon, it was observed that the client
ended up with a formal data warehouse, plus a huge number of very highly
denormalized tables. These denormalized tables were supposed to provide easy
and fast access to the data. However, in reality, it turned out that the data across
the tables was not consistent and the tables were still complex, in cases even
more complex, to interrelate, link, and maintain, than that in the data warehouse.
This made the new data warehouse very inefficient to be used as a data source
for reporting purposes.

The client realized the problems of creating denormalized tables for each report
and decided to adopt another strategy. Following the same iterative approach
that the client had already successfully applied for the development and
implementation of the data warehouse, the client decided to start a new project
to implement and maintain new data structures in the form of dimensional

Chapter 7. Case Study: Dimensional model development 335

databases (also called data marts) that would allow reporting and data analysis
tasks in the way the business was demanding. The dimensional models would
not only help the clients with ad hoc analysis but also would be easier to read
and understand. In addition to this, the dimensional models would give much
better performance when compared to the E/R modeled data warehouse tables.
The client was confident that the necessary data was not only available but also
quality checked in the data warehouse. This important exercise of quality
checking the data was done by the data warehouse development team. So now
the client was ready to implement dimensional models from the data warehouse.

To help the team learn the proper design skills needed for dimensional design,
the client decided to approach the problem in a more efficient way. That meant
sending the project manager, a business analyst, a data modeler, the Enterprise
IT architect, and a couple of developers to a training center for education on
dimensional modeling. In addition to learning dimensional modeling techniques,
they became familiar with a methodology called DMDL.

To review DMDL and dimensional modeling technique, review Chapter 6,
“Dimensional Model Design Life Cycle” on page 227. There we see that to design
a dimensional model, we should first identify a business process candidate. The
next step is then to identify the corresponding dimensional models that would be
developed and implemented. And that is the methodology we used in this case
study. But before we do that, we will try to understand more about the company’s
business.

7.2 The company
The Redbook Vineyard was founded in 1843 in Spain, in the region of La Rioja. It
started producing and selling its own wine, first locally and then spread through
most of Europe. The company specialized in Rioja wine and for many years the
sales of this wine continued to represent its primary source of income.

In 1987, Redbook Vineyard acquired a distributor of alcoholic and non-alcoholic
drinks and moved the corporate office to Madrid, Spain. Since then, it has also
incorporated other non-drinkable items, such as T-Shirts.

The company is number ten in sales of spirits, and number three as a wine
producer in Europe, having a market share of 4% and 23%, respectively.

7.2.1 Business activities
The Redbook Vineyard is an old company owned by one family. Since 1843, they
have been producing their own wine mainly made with mazuela grapes
harvested from different wine locations across the La Rioja region. The wine is


produced at their winery in La Uvilla. The production of wine includes different
activities, from buying grapes, though the wine processing, to the distribution of
the wine in bottles to customers.

In 1987 the Redbook Vineyard decided to broaden their activity portfolio and
started to distribute major brands of wine and spirits to America and Australia.
Since then they have established sales branch offices in most of the major
European cities.

The Redbook Vineyard has built up very good business relationships with large
retail store chains over the years, and its market share in spirits and wine is
about 10% of the entire European market.

The Redbook Vineyard is highly customer-oriented. They collect all
customer-related information in their Customer Relationship Management (CRM)
system at each branch location.

In order to manage its business across different countries, the Redbook Vineyard
uses modern IT systems. For their core businesses, they use an Enterprise
Resource system (ERP) and a CRM system. In order to support top
management for decision-based activities, the Redbook Vineyard has just built a
centralized Enterprise Data Warehouse in Spain.

7.2.2 Product lines
As shown in Figure 7-2 on page 338, the company markets several lines of
products. However, in reality, the beverages are the products that keep the
business going for Redbook Vineyard.


Catalog

Group classes
Beverages Non-Beverages

Wine Non- Spirits Beer
Beverage groups
Alcoholic

Dark Alcohol ..... Merchandising
Beverages Soda T-Shirt
Red-1 Bev-1 items

Oak
Lemonade Alcohol
White-1 Lighter
Bev-2

Light .....
Red-1

Products (beverages and merchandising items)

Figure 7-2 Product lines for Redbook Vineyard

The product catalog of the Redbook Vineyard company is divided into two main
categories:
Beverages
Non-beverages

We discuss each of these product categories in more detail below.
Beverages
Beverages are initially any sort of potable drink. During the last few years,
despite the appearance of new types of drinks, such as light drinksand energy
drinks, the product structure has not changed. It is however very likely that this
structure will need to change very soon. As seen in Figure 7-2, the Redbook
Vineyard presently has the following beverage lines: wine, non-alcoholic, spirits,
and beer.

We discuss each of these beverage types in more detail below.

Wines
The line of wine covers any sort of wine that has no more than 25% alcohol. That
includes, for example, sherry and oporto wine, which has a graduation
significantly higher than normal wine.


Non-Alcoholic
The non-alcoholic beverages presently include most types of soda with 0%
alcohol contents. It is foreseen that Redbook Vineyard will market stimulant
drinks, also called energy drinks, in the near future. In addition to this, milk and
yogurt drinks will also be experimentally introduced and temporarily distributed to
a few chosen resellers. Until now dairy drinks were not introduced because of the
extra overhead involved with these short shelf-life drinks.

Beer
Any sort of beer with under 10% alcohol is included in this line. Beers with an
alcohol content above 10% are uncommon. Beers with low alcohol content, also
called light beers, are also marketed under this product line. However, depending
on the legal clarifications, the client is evaluating the possibility of moving the light
beers with an alcohol volume of less than 1% into the non-alcoholic product line.

Spirits
Any other marketed alcoholic drink not covered in the beverage lines previously
described will be included in this group. Spirits are typically beverages with the
highest alcohol content.

Non-Beverages
This is the other product line, apart from beverages. It offers merchandising
products, such as T-shirts, glasses, and lighters. The company started these
products in order to support marketing the beverage product line. However,
objects such as certain jars, or art decó lighters, are very successful and the
company is considering marketing them separately.

7.2.3 IT Architecture
The Redbook Vineyard IT infrastructure is distributed across the different branch
offices. Every branch office in every country has its own ERP system. This is
because each country has different legal requirements. Activities related to the
customers are covered by the CRM system. The topology of the architecture of a
typical branch office is shown in Figure 7-3 on page 340.


Central location - Spain

A branch office
CRM
CRM
Production

Inventory ERP
ERP
Transport EDI
Company
Sales

Central
Procurment

Accounting Data flow
DWH
DWH
Marketing

Figure 7-3 Typical branch office IT architecture

The data warehouse component, as shown in the Figure 7-3, is the only IT
component that is centralized. The data warehouse component provides
services to the Spanish Central location (headquarters) and to the remote users
from all other European branch offices via corporate virtual private network
(VPN). The data warehouse is refreshed overnight.

The main consumable data source for business users is the company’s data
warehouse system of record (SOR), also called the Enterprise Data Warehouse.
The data in this system is cleansed and reliable. It contains data that originated
in the different operational systems used at the Redbook Vineyard.

7.2.4 High level requirements for the project
In the following section, we list the requirements identified for the design and
implementation of the business intelligence solution that the Redbook Vineyard is
demanding.

User requirements
The data mart must have a high business strategic value. This is a critical
requirement.


Ideally the system availability should be 24 x 7. This is not a critical
requirement because to have the system available during weekdays and
office hours would actually suffice.
The system should be implemented in the shortest period of time possible.
This is a critical requirement.
The data must be consistent, reliable, and complete (the truth, the whole
truth, and nothing but the truth). This is the most critical requirement.

IT requirements
Some of the IT requirements are:
There should be one decision support system of choice covering all source
systems with business strategic value within the company.
Complex advanced analysis should be possible against consolidated data
from several source systems.
A generic data warehouse structure should be made available for retail and
manufacturing processes to store and retrieve data/information.
The MIS and primary source systems should be separated. This will help
increase performance for the OLTP systems.
The data warehouse must include detailed history data to avoid overloading
OLTP systems when restoring archived tapes for inquiry purposes.
The data must be standardized by integrating available data standards into
the data modeling tool.
There should be a minimum basic meta data management system easily
accessible, with information about the data warehouse components.

OLAP client requirements
One central OLAP system should be available.
A user-friendly GUI should be available to do analytical work with minimal
assistance. The GUI must permit drag and drop operations to build the OLAP
queries without having to type field and table names.

7.2.5 Business intelligence - data warehouse project architecture
In Figure 7-4 on page 342, we show the proposed architecture for the data
warehouse, showing the major components. The shading in the figure
differentiates between the present architecture and the components needed for
completing the data warehouse solution.


Data Warehouse Environment
Source Systems
Staging Enterprise Data Warehouse
Full history 3NF
Cental Locations Area No user access
and
Branch Offices

Analytic Area
MultiDimensional Model
Summary Area
Full history
External Full History
User access
Data Marts (Internal Data Marts)

Phase 1
OLAP Server
Phase 2 Query/Reporting
OLAP Analysis

Figure 7-4 High level project data warehouse architecture

The architectural components of the data warehouse, shown in Figure 7-4, are
as follows:
Source systems: ERP central systems and CRM systems at each branch
office.
Data warehouse environment: Consists of the following sub-components:
– Staging area. This area contains all the data consolidated, preformatted
and transformed before it is loaded the into System Of Records (SOR)
area or the Enterprise data warehouse.
– Enterprise data warehouse. This system of records area stores all
detailed information, including all the company history. It is a highly
normalized database (designed in 3NF), covering the subject areas with
the most strategic business value.
– Analytical area. This includes detailed atomic dimensional models (data
marts). The data stored in these dimensional models is highly detailed
and atomic.
– Summary area. The summary area includes the aggregate tables based
on the detailed analytical dimensional models.
Reporting: The reporting component includes basic static reports (via SQL
queries). This will be replaced by a new OLAP/reporting solution.
OLAP: This component includes advanced analytical function with ad hoc
capabilities and what-if analysis.


External data marts: This component includes the data marts for mobile
Customer Representatives.

Note: The dimensional model or the multidimensional database is in the
analytical area of the data warehouse architecture (see Figure 7-4).

7.2.6 Enterprise data warehouse E/R diagram
Figure 7-5 shows an example of an E/R diagram. The present enterprise data
warehouse could have one that looks similar. The objective here is not to show a
complete E/R diagram that is meaningful. It is simply to demonstrate that the size
and number of interrelationships can result in quite a complex diagram.

The primary subject areas of the enterprise data warehouse model are:
Arrangements (AR): consisting of Orders, Invoices, and Orders Lines
Transactions (TXN)
Finance
CRM

Figure 7-5 Physical E/R model


7.2.7 Company structure
Figure 7-6 shows the company organization structure. The IT and the human
resources departments are not included in this case study. The structure is the
same as the one the company had before the business expansion of 1987. The
company is now considering a reorganization of departments and this chart will
certainly change. But that should not influence the results of this case study.

The Red Vineyard Company

General Manager
Assistant Director

Finance Sales Procurement Manufacturing

Business
Information
Marketing Inventory Production Processes
Technology

Human Sales Procurement Distribution
Resources

Accounting Order
Handling

Payment Invoicing
Accounting

Figure 7-6 Company structure

7.2.8 General business process description
In this section we describe three important aspects of the Redbook Vineyard
company. They are the:
Products
Customers
Business processes

We now focus on each of these in more detail below.

Products
We discussed the company products in 7.2.2, “Product lines” on page 337. In our
case study, we focus only on the Beverages line of products as shown in
Figure 7-7 on page 345, since beverages are the primary source of income for
the company.


Beverages
Beverages

Dark
Dark Oaks
Oaks Light
Light Alcohol
Alcohol Alcohol
Alcohol
Soda
Soda Lemonade
Lemonade .....
.....
Red-1
Red-1 White-1
White-1 Red-1
Red-1 Bev-1
Bev-1 Bev-2
Bev-2

Dark
Dark Alcohol
Alcohol
Soda
Soda
Red-1 ....
.... ....
.... Bev-1
Bev-1 ...
...
Red-1 1 Ltr
1 Ltr
1 Ltr 1 Ltr
1 Ltr
1 Ltr
Bottle Bottle
Bottle
Bottle
Soda
Soda
Dark ...
...
Dark 2 Ltr
2 Ltr
Red-1
Red-1
1 Ltr
1 Ltr
Tetrapak
Tetrapak Soda
Soda
Valuepack
Valuepack
12 x 1 Ltr
12 x 1 Ltr

Figure 7-7 Beverage Items

The products we show in Figure 7-7 were valid at the time of creating the
diagram. Important points to consider about these beverage items are:
Products could be ordered and shipped, and therefore they are considered
active (status = active).
There are beverages that are no longer produced (status = discontinued).
There could also be beverages that are new and are about to be released.
Although not yet released, new beverages can be ordered (status = new).
Another characteristic of the products is their presentation. A beverage is not
always in bottles. Other forms of presentation include tetrabriks, soft plastic
bags, and barrels.
The characteristics of the products normally remain unchanged during their
lifetimes. However, the structure itself is subject to change every two or three
years.

Customers
We now discuss the Redbook Vineyard Company customers who have been the
key to the company’s success over the years. An example of the customer
structure is shown in Figure 7-8 on page 346.


Customers
Customers

Cust-1 Inc.
Cust-1 Inc. Cust-2 Inc.
Cust-2 Inc. ....
....

Cust-2 Cust-2 Cust-2
Cust-2
Cust-2 Cust-2
Cust-1 Spain
Cust-1 Spain Cust-1 India
Cust-1 India Products Chechia Europe
Europe
Products Chechia
(Procurement)
(Procurement)

Purchase
Purchase
Orders
Orders

Figure 7-8 Customer structure example

Important points to note about the Redbook Vineyard customers are:
The typical customer is a medium size company that resells the beverages to
smaller companies and to end-consumers.
Customers are grouped by classes: "WHOLESALE", "STORE CHAIN", "GAS
STATION CHAIN", "RESTAURANT", or "SMALL SHOP".
The customer gets a credit limit and a discount volume depending upon
certain factors, such as seniority of the customers, the average order volume
per year, and the paying discipline of the customer. The credit limit can be any
amount starting from 10 Euros, and the discount can be between 1% and
40%.
Customers usually have a preferred way of payment, such as money
transfers, or credit cards. The company assigns payment terms to customers.
The default is 30 days for existing customers and pay upon delivery for new
customers. However, this can vary between 5 and 90 days.
Some customers are small companies, that have just have one department or
address from which all the orders come. Other customers are multinationals
with presence in several countries. In some of the countries they have only
one organization, in others, they have several organizations with divisions
with different names, but each of them are actually the same company.
Customer attributes, such as name or address, usually remain stable and do
not change across the life of the customer in the company database. There
are, however attributes, such as credit limit (for example, 10 000 dollars),
payment terms (for example, 30 days after invoice), payment method (for
example, bank transfer), and discount (for example, 10%) that are
recalculated and reassigned monthly. For example, during specific phases,
such as two months before, and during Christmas, these attributes are
recalculated on a weekly basis.


Business processes
In this section we discuss the business processes that are shown in Figure 7-6
on page 344. The business processes were accurately and thoroughly
documented at the beginning of the data warehouse project. Most of the data for
these business processes was made available in the enterprise data warehouse
that the Redbook Vineyard company built. However there are a few exceptions.
These exceptions are mentioned below.
In this section, we discuss the following business processes of the Redbook
Vineyard company:
a. Inventory handling (also called Inventory management)
b. Procurement
c. Distribution
d. Production
e. Accounting
i. Payment accounting (also called Credit Controlling)
f. Sales
i. Sales Order Handling
ii. Sales Invoicing
iii. Sales Marketing/CRM

a. Inventory management
The main function of this process is to manage the inventory. This process is
responsible for the following:
Receiving the wine when it is shipped from the bottling plant. In addition to
this, inventory management is also responsible for receiving other beverages
that are shipped by external suppliers.
The placement of received products in their corresponding storage areas.
The inventory department maintains inventory levels. Also on a weekly basis,
the inventory department communicates to the procurement department
about inventory levels of products. This communication helps the
procurement team to reorder products depending on the need.
Inventory is also responsible for packing and shipping the ordered products.
Although in exceptions where certain items are expensive and need special
handling, single bottles of certain spirits are separately packaged and sent.
Normally each product is packaged in boxes or containers of several units
each. Typically packages will have six, 12, 24, and 100 bottles each.
Additionally, for better handling of the goods, the packages can be stored on
pallets which may hold up to 100 packages or boxes.


Note: The shipping charges are dependent on the order volume and total
weight and not the number of items ordered. However there are a few items,
such as very expensive and luxurious champagne bottles, whose packaging
and special handling demands that we track the shipping cost per order line
rather than per order. Therefore, shipping charges are invoiced at the order
line level and not at the order level. Also, products ordered in one order can be
shipped from different warehouses.

The Redbook Vineyard customers order the goods throughout the year, and,
therefore, the Redbook Vineyard has its own warehouses in each European
country. When customers order the product, the Redbook Vineyard fulfills the
request from the warehouse in the country where the order was placed. If the
product is not available, then it is shipped from the central warehouse or any
other close available warehouse. If no warehouse has the ordered goods, then
Redbook Vineyard orders the goods from the corresponding supplier responsible
for maintaining the warehouse.

b. Procurement
The main function of this process is to manage the purchase orders sent to the
suppliers and manufacturing plants. This process, however, has no special
business strategic relevance, since the main business income is from the wine
produced by the owning company. The data warehouse contains data about this
process but at a very summarized level.

This procurement business process is responsible for:
The coordination with the inventory process for stock level maintenance. We
discussed earlier that inventory communicates to procurement about the
inventory levels for various products.
Procurement is responsible for the creation, submission, and tracking of the
purchase orders. In addition to this, the process also checks and closes the
order after receiving the ordered goods.

c. Distribution
This is a very straightforward and simple process. This process generates high
infrastructure costs. It is not part of the core business because the company has
outsourced it to an external company called Unbroken Glass Hauler Inc. (UGH).
UGH uses the IBM Informix database manager for its operations. Electronic Data
Interchange (EDI) is used between the two companies and it has been set up in
a way that enables periodic transfer of the minimum data necessary to
coordinate and report the product transport operations. The method of goods
delivery typically depends on the customer, who will have a preferred method
such as by truck, train, or plane.


d. Production
Presently the company produces only wine. This process is very complex
because of innovative techniques implemented for the production of wine, and
no operational data is fed yet into the data warehouse. It is presently supported
by an ERP application, and if data is needed for reporting or analysis, then the
ERP application is usually accessed directly without many restrictions. This
production process mainly covers:
Planting and collection of wine grapes.
Reception, processing of the grapes, and storage in the wine cellars.
Bottling and shipping of the wine.

e. Accounting
This process covers all the key financial and accounting aspects of the company.
However, for similar reasons to the ones pointed out for the distribution process,
the company is presently evaluating the possibility of outsourcing this process.
Although the main data is fully available in the data warehouse, there are a
number of tables that need to be recreated every day and access to them from
the data warehouse is periodically restricted. This process manages, among
others, the following:
Maintenance of the General Ledger.
Control of the completeness, consistency, and accuracy of all financial
statements from all the branches and offices worldwide
Financial reporting to the local authorities

The Accounting process has another sub-process which is called the payment
accounting process. This process is discussed in more detail below.
e-i. Payment accounting (also called Credit Controlling)
This is a special process working very closely to the sales business process,
especially regarding the customer credit analysis and reclassifying.
This process is responsible for:
– Ensuring that the customer pays on time and does not owe more than
their allowed credit limit.
– Analyzing customer debts. The process classifies the customer credit limit
according to factors such as purchase volume, payment discipline, and
credit history. This input is further provided to the sales process for
classifying the customer bonus and discount percentages.

f. Sales
Apart from the production process, which we reviewed earlier, the sales process
is one of the most complex. This process heavily and directly impacts the


generation of revenue and is considered to be highly critical for the business.
The sales business process encompasses the following:
The design, execution, and analysis of marketing campaigns
Analysis of the market for introduction of new products or elimination of
non-profitable products
Responsibility for providing all customer assistance and guidance by a Sales
Representative
The reception, handling, and invoicing of the Customer Orders

The sales process is divided into three separate sub-processes. They are sales
order handling, invoicing, and marketing. We discuss each of these processes
below.
f-i. Sales Order Handling: This sub-process covers the order handling of the
sales orders from receiving the order to shipping. It deals with the following:
– Receiving an order and processing it
– Coordination with inventory for shipping purposes
– Coordination with procurement for ordering purposes
– Overall tracking of the sales order
– Any related customer interaction
f-ii. Sales Invoicing: The invoicing is less complex than the order handling
process. But from a reporting standpoint, it is the most important one. This is
because the main financial and corporate reports are made against the
invoiced amounts. This process manages:
– Printing and sending the invoice data to the customer
– Quantitative analysis (volumes, regions, customers, and products)
– Financial analysis (total sales, discounts, regions, customers, and
products)
– Correction or cancellation of invoices for returned or damaged material
f-iii. Sales CRM/Marketing: This sub-process has a objective to analyze to
decide on the introduction of new profitable beverages. In addition, this
process also tracks the decrease in the acceptance and market demand of
existing beverages. Based on this input, a decision is made whether or not to
continue these beverages. The marketing process responsibilities include:
– Analysis of market by channel and product.
– Ad hoc screening of the market along with the competitors.
– Tracking and analysis of marketing activity (type of actions, regions, and
customers).


– The analysis and segmentation of customer groups to allow a better
targeting during marketing campaigns.
– The execution of marketing campaigns around new or existing products to
acquire new customers or increase the volume of sales to existing ones.
– Preparing, performing, and analyzing customer contacts by Sales
Representatives.

A part of the data needed for the execution of marketing campaigns is delivered
(purchased) by external companies specialized in market analysis. This data is
not presently part of the data warehouse, and is usually obtained in the form of
XML data structures, which, depending on the amount of data, are sent by e-mail
or by DVD.

7.2.9 Developing the dimensional models
The Redbook Vineyard has decided to design dependent data marts for several
of its business processes. Each of these dependent data marts would get the
data from the existing data warehouse. We have understood well the critical
business needs for which the Redbook Vineyard wants to move forward with this
initiative.

But the next question is, where do you start? What do you do first? To help you in
that decision process, we have designed a Dimensional Model Design Life Cycle
(DMDL). For a detailed description of the DMDL, see Chapter 6, “Dimensional
Model Design Life Cycle” on page 227.

In the remaining sections, we complete the case study by developing a
dimensional data model for the company. We use the DMDL as a guide to
demonstrate how it can help to develop a robust and complete dimensional
model.

In the next section, we start with the first phase which is Identify the
Requirements.

7.3 Identify the requirements
The first phase of the DMDL starts with identifying the requirements. This phase
is heavily dependent on analysis of the business processes inside the
organization. This is shown in Figure 7-9 on page 352.



Granularity
Processes

Identify Conformed



Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management

Figure 7-9 Business Process Analysis

The main goals of this phase are to:
Create an enterprise-wide business process list
Identify the business process (includes prioritization)
Identify high level entities and measures for conformance
Identify various data sources involved
Identify the requirement gathering approach you would follow
Gather requirements
Analyze requirements

7.3.1 Business process list
This is a very straightforward task, assuming that the documentation available
from the first phase of the project is still up to date. To recall, the Redbook
Vineyard company implemented a data warehouse for its company a few months
back. We will not include here the detailed business process descriptions
because they are very extensive. We identified important business processes in
detail in 7.2.8, “General business process description” on page 344. They are:


Inventory handling (also called Inventory Management)
Procurement
Distribution
Production
Accounting
– Payment accounting (also called Credit Controlling)
Sales
– Sales Order Handling
– Sales Invoicing
– Sales Marketing/CRM

Business process assessment
In this activity we identify the factors that the company considers most important
to evaluate each business process. We also prioritize them and assign weights.
These assessment factors are shown in Table 7-1.

Table 7-1 Point table for assessment factors
Assessment Low Medium High
factor

Complexity 6 4 2

Data quality and 3 6 9
availability

Strategic business 2 4 6

System availability 1 2 3

Note: We have mentioned only a few assessment factors in Table 7-1.
However, we encourage you to assess your business processes across
various factors while deciding to prioritize a process.

The next step is to relate those factors to each business process and rate each
combination accordingly. You can see the results in Table 7-2.

Table 7-2 Business process assessment
Name of Complexity Data quality System Strategic
business and availability business
process availability significance

Distribution Low Medium Low Low

Sales CRM/ Medium Medium Medium Medium
Marketing


Name of Complexity Data quality System Strategic
business and availability business

Inventory Medium High High Medium

Procurement Medium Medium High Low

Sales Orders Medium High High Medium

Accounting Medium High Medium Low

Production High Low High High

Sales Low High High High
Invoicing

Note: We have rated both the data quality and availability, and the
system availability factors as High for all the processes whose data is
already present in the data warehouse system of record. For all the
outsourced processes, or process candidates for outsourcing, we have rated
the strategic business significance as Low.

Marketing data and system availability has been rated as Medium because
the data partially comes from external sources.


7.3.2 Identify business process
In this activity, we prioritize the business processes. This means we identify the
most and least feasible process for building a dimensional model. In order to do
this, we use the Point table for assessment factors as shown in Table 7-1 on
page 353, and we turn the priority indicators (Low, Medium, and High) into
numbers. Doing so, we get results depicted in Table 7-3.

Table 7-3 Business process prioritization
Business Complexity Data quality System Strategic Final points
and availability business

Distribution 6 6 1 2 15

Sales CRM/ 4 6 2 4 16
Marketing

Inventory 4 9 3 4 20

Procurement 4 6 3 2 15

Sales Orders 4 9 3 4 20

Accounting 4 9 2 2 17

Production 2 3 3 6 14

Sales 6 9 3 4 22
Invoicing

Observing Table 7-3 we see that Sales Invoicing has the most points. In other
words, this means that Sales Invoicing is one of the important and feasible
business processes. We will then use the Sales Invoicing business process to
design our dimensional model.

Note: Sales, as a process, is divided into three processes. They are Sales
Invoicing, Sales Orders, and Sales Marketing/CRM. However, after assessing
the various business processes, we have found that the Sales Invoicing
process has the most points.

More on the sales process
The corresponding process detailed description is taken from the data
warehouse documentation, which is relatively fresh and easily accessible. The
sales process consists of the following processes:


One: Sales CRM/Marketing
The Sales CRM/Marketing process involves the following activities and
procedures:
A Sales Representative contacts a customer or is contacted by a customer.
The contact can be established by telephone, e-mail, fax, instant messaging,
or in person.
The Sales Representative assists the customers and guides them through the
purchase decision process. The most important priority is the satisfaction of
the customer. Next is to generate an immediate sales order.

Two: Sales Order handling
The Sales Order handling process involves the following activities and
procedures:
If the sale evolves positively, the Sales Representative enters the order by
means of a mobile computer. Otherwise, the order can be sent in later, by one
of the following ways:
– Online, using the ordering application available through the Redbook
Vineyard Web page
– By sending a printed copy of the order produced via the Redbook Vineyard
Web application
– By sending the order using the postal service, fax, or e-mail

Note: The customer can access the order form through the Web
application, print it, and send it later rather than immediately submit it. Both
existing and new customers can also enter an order via the company Web
page without having made any previous contact with any Sales
Representative.

The Customer Order is processed by the sales office closest to the customer
address.
Ordered products are consolidated, within each order, according to their
product identifiers, so that one product id only appears in one line and no
more than in one line of the same order. In other words, if a customer orders
the same product twice in the same order, that same product is represented
in one line with the total quantity.
The sales department sends an order acknowledgement to the customer,
including the Redbook Vineyard order number and an estimated delivery
date.


Three: Sales invoicing
Sales invoicing involves the following activities and procedures:
An order can be partially shipped, in which case it can also be partially
invoiced. The following describes partial order handling:
– Partial order delivery with partial invoice:
This can happen when an item is not in stock, and has to be reordered. In
this case, only the items available in stock are shipped. The other activities
in partial shipments are:
• Notify the procurement department of items not available.
• Items in stock have to be shipped immediately.
• The inventory transaction and the invoice amount are registered in the
General Ledger.
– No partial order delivery:
In this case the order must be entirely shipped at once.
• Notify the procurement department of items not available.
• Order is shipped only when all the items are available.
• The final price is calculated and adjusted.
• The invoice is printed and sent to the customer.
• The inventory transaction and the invoice amount are registered in the
General Ledger.
– Partial order delivery without partial invoice:
The instructions are the same as the case described above for “Partial
order delivery with partial invoice” except that the invoice can only be
produced after the order has been fully shipped.
– Closing the order
After the order has been fully shipped and invoiced, the order is formally
closed in the system.
– Returned and damaged material:
A printed invoice cannot be changed. In case of damaged or returned
material, updates or corrections are performed by issuing a credit note
(which is the same as an invoice but with a negative amount).

Note: An order item can be shipped from different warehouses. Different
shipments of the same product are allowed, if having ordered more than
one unit, the product can only be fully delivered when shipping it from
different warehouses. Every shipment corresponds to one line in the
corresponding invoice. An invoice line does not consolidate different
shipments of the same product.


Figure 7-10 shows an example of an invoice form.

Invoice Number: 10078902 01.10.2005

Customer
Order Number: 123678 Order date: 21.09.2005

Customer: Discount Cust-1, Limited
Delivery date: 01.10.2005
Sales Rep: SalesPerson-1

Price/ Qty Disc Item Acquired
Line # Item name Unit Sold % Price Bonus

1 Oak White-1 0.75l 12.00 50 10% 540.00 10

2 Alcohol Bev-1 0.5l 16.00 10 5% 152.00 10
3 Bottle Water 5.00 100 0% 500.00 50
Pack 6 x 0.33l
4 Soda 10.00 20 0% 200.00 100
Value Pack 2 x 0.75
Currency: EU 1392.0
Total: 0

Figure 7-10 Example invoice

E/R model for order handling and invoicing
Figure 7-11 on page 359 depicts the tables in the E/R model used by the Sales
Order handling and Sales Invoicing processes. A dimensional model will be
designed for the Sales Invoicing business process.


Figure 7-11 E/R model for Sales Process

7.3.3 High level entities for conformance
In this activity we identify the entities and measures that need to be consistently
defined and shared across the different business processes. We were able to
select the entities from the data warehouse. A sample of those selected are


Table 7-4 High level entities
High Level Entities --->

Orders + Invoice
Organization

Warehouse
Employee

Shipment
Beverage

Currency

More
Time
Business Processes

Sales Invoicing X X X X X X X X

Accounting X X X X X X

Inventory X X X X X

Sales CRM/Marketing X X X X X

Distribution X X X X X X X

Procurement X X X X X X

Production X X X X

More

For simplicity, we have limited the sample set of entities in Table 7-4 to only those
initially needed for the Sales order handling process, rather than all the ones
needed for the complete process.

The idea behind this exercise is to find out what entities are common across
several business processes. Once we are able to find out these common entities,
we want to make sure that all data marts use common or conformed entities in all
dimensional models. Each of the business processes are tied together through
these common entities.

In order to create conformed dimensions that are used across the enterprise, the
various businesses must come to agreement on defining common basic
definitions for common entities.

For example, as shown in Table 7-4, the business processes such as Sales,
Accounting, Inventory, CRM/Marketing, Distribution, Procurement, and
Production should agree on a common definition of the Beverage (Entity). This is
because all these business processes have the beverage entity in common. This
beverage entity may later become a conformed beverage (product) dimension
which will be shared across all the business processes. However, getting
agreement on common entities such as Product (beverage) and Customer may
be difficult because these entities are subjective and their definition will vary from
one business process to another. On many entities it will be easy to reach


agreement for defining a common definition. For example, entities such as
stores, dates, and geographies, because these entities and their definitions are
are fairly straightforward.

7.3.4 Identification of data source systems
In this activity we identify and document the data sources that support the
business processes described in 7.3.2, “Identify business process” on page 355
and 7.3.3, “High level entities for conformance” on page 359. The data sources
are documented in Table 7-5.

Other information contained in Table 7-5 can be derived from the Redbook
Vineyard company structure and its business process descriptions as described
in 7.2.7, “Company structure” on page 344 and 7.2.8, “General business process
description” on page 344.

Table 7-5 Data sources for business process
Business Data Data owner Location Platform Data
process sources update

Sales order Enterprise Sales Madrid DB2 Daily
handling data
warehouse

Sales Enterprise Sales Madrid DB2 Daily
Invoicing data
warehouse

Accounting Enterprise Finance Madrid DB2 Daily
data
warehouse

Inventory Enterprise Procurement Madrid DB2 Daily
data
warehouse

Sales CRM/ Enterprise Sales Madrid DB2/ Daily/
Marketing data XML file upon
warehouse request
and external
source
systems

Distribution Outsourced Procurement/ Madrid Informix/ Periodic
ERP Sales DB2 Real
time


Business Data Data owner Location Platform Data
process sources update

Procurement Enterprise Procurement Madrid DB2 Daily
data
warehouse

Production ERP Production Madrid DB2 Real
time

Note: The data owner is not the system owner. For example, the Sales order
data available in the Data Warehouse is owned by the Sales order
department, but the data warehouse system of record is owned by the IT
department.

7.3.5 Select requirements gathering approach
We discussed the source-driven and user driven requirement gathering
approaches in detail in 6.2.5, “Select requirements gathering approach” on
page 237. For this exercise, we have decided that the requirements gathering
approach will be user-driven rather than source-driven. This is because there
was no request to include all the source data, so we decided to involve the users
in the decision since they best understand the business needs.

Note: The business users typically have clear ideas about what source data is
required. So be careful if asked to integrate all data from all existing data
sources. Though easy to ask, it may result in a much higher cost and larger
effort than expected. Therefore, an accurate analysis of the data sources will
minimize the potential for these issues.

7.3.6 Gather the requirements
In this activity we gather business requirements for the Sales Invoicing business
process. We perform a series of interviews with top and middle management,
and users involved in Sales Invoicing-related reporting activities. The objective is
to inquire, collect, and document the different requirements for reporting and
online data analysis purposes.

In 6.2.6, “Requirements gathering” on page 240, there was a list of aspects to
consider while interviewing the user. Here are more specific considerations to
use when gathering requirements:


Ask for existing reports and for information about queries or programs used to
get the required data.
Ask how often each report or query is run.
For each requirement, ask about:
– Specific data needed
– Selection condition or filtering criteria to apply
– Consolidation or grouping needed

Note: Questions formulated as follows: “Show me the total sales and total
discounts per month” and “Show customers sorted by total sales for a given
country” should be encouraged. This enables them to be quickly converted
into SQL pseudo statements, such as “Select A where B grouped by C order
by D” for query development.

In Table 7-6, we show a sample list of the requirements collected during the user
interviews. For intermediate filtering, we asked the user to prioritize the
requirements by assigning one of these three categories: very critical, necessary,
and nice-to-have.

Nice-to-have requirements are excluded from this list. You might decide a
different categorization, and cover all the requirements in the first design. The
questions have been numbered using two different prefixes: Q for very specific
requirements, and G for general requirements.

Table 7-6 User requirement table
Q# Sales Invoicing business process requirements

Q01 What is the gross profit of product sales grouped by beverage and customer
category during a given month?

Q02 Debt analysis by customer: Print active and outstanding debts at a given date,
by customer and grouped in ranges or bands of days. The user needs to have
a report to list the sum of payments due in the next seven days per customer
or per customer country. Also the user needs to report the total invoiced
amount with payment due longer than 30 days.

Q03 What is total amount of outstanding debts on a selected date by beverage and
by category? And also by customer and invoice number?

Q04 What is the number of units sold and net sales amount per product for
non-beverage items, per quarter?

Q05 What are the 10 most profitable beverages by:
1. Largest total sales per year to date and region.
2. Largest gross margin per year to date and region.



Q06 Who are my 10 most profitable customers per region by largest total sales
volume per year?

Q07 How is my net profit developing since the beginning of year? (Display profit
during current year at the end of month, calculating monthly and accumulated
profit per month.)

Q08 How many bottles (containers) and liters were delivered per time period in
months compared to the previous three years during the same period?

Q09 List the total sale discount per beverage or product group or customer or
customer category per time period compared to previous three years during
the same period.

Q10 Organization credit history. For a selected month interval, list the outstanding
payments at the end of each month grouped by customer and sorted by
outstanding amount.

Q11 Report showing net revenue, gross sales, gross margin, and cost of sales per
beverage and month, quarter and year.

Q12 Analysis of delayed invoicing due to partial shipments: Total amount grouped
per warehouse, beverage, or beverage group for a selected period where the
invoice date is a number of days greater than the first shipping date.
Business reason: If the customer does not want partial invoicing, the order
will be invoiced once it is fully shipped. Partial order deliveries will wait to be
consolidated in a full order invoice. Orders may take long time to complete,
from the time of the first partial shipment. It is necessary to analyze the impact
of the indirect credits that the customer gets in this situation, and the increase
in the sales performance if we improve the shipping process. The pertaining
data needs to be analyzed only once per month.

Q13 What is the sales order average value and sales order maximum value per
month per Sales Representative?

Q14 What customers have placed orders greater than a given value during the last
year?

Q15 What are the companies that last year exceeded their credit by more than
20%?

Q16 List gross sales totals and total discounts by Sales Representative for the past
year sorted by gross sales amount in descending order.

Q17 List the total discount amount, percentages, and gross margin per Sales
Representative and customer sorted by gross margin in ascending order
during last year.



Q18 List the number of orders per month containing one or more non-beverage
items since the first year available.

Q19 Gross sales per Sales Representative, sales office, month, and currency
selected by year.

Q20 Gross sales amount per each method of payment and country during last
year.

Q21 Three top week days per month according to average sales by Web orders.
The listing must include total gross sales and number of orders.
Business reason: For planning system maintenance, is there a day (or days)
of the week in which the number and orders increase or decrease? (sales
office = WEB)

Q22 Number of orders and average sales order volume per transmission type (fax,
telephone, Web) per month and year.

Q23 Revenue, gross margin, and cost of goods during last year per supplier sorted
by revenue in descending order.

Q24 For tracking urgent customer orders specially ordered to the supplier, a report
is needed listing: Supplier id, Supplier name, Contact name, Telephone
number, Discount%, Web page and Web access code from the supplier.

Q25 Monthly report for accounts containing: Month, Customer id, Net sales, Duty,
Cost of item handling, Cost of goods.

G01 All sales and cost figures for beverage reports must be available in both
reporting currency and invoice currency.

G02 Beverages need to be summarized or added by product group.

G03 Allow tracking of average delivered alcohol liters, maximum and minimum
alcohol content values per organization, Sales Representative and/or
beverage group
Business reasons: Restructuring of the existing marketing product groups,
and supporting resellers with the legal requirements regarding not selling
beverages to the non-adult population.

G04 Gross sales, discount and bonus totals per year per corporation.

G05 Changes to beverage and sales organization structure and all corresponding
attributes must be tracked across the time (history needed).

G06 Changes to customer attributes need to be fully maintained and stored.


Note: Points to review about the requirements
Gross profit = Profit calculated as gross sales income less the cost of
sales.
Gross sales = Total invoiced amount minus total discount including
returned and damaged items (in the requirements, Sales means Gross
sales by default, unless otherwise indicated).
Gross margin = Percentage difference between the cost of sales and the
gross sales.
Customer active debt = Amount invoiced to a customer and not yet due for
payment.
Customer outstanding debt = Amount invoiced and due for payment.
Net sales = Gross sales minus returned and damaged items.
Net profit = Net sales - Cost of sales - Duty amount - Cost of item handling
(Item shipping charges).
Cost of sales = Cost of the goods that have been sold, tax excluded.

7.3.7 Analyze the requirements
In this activity we analyze the Sales Invoicing business process requirements
gathered in 7.3.6, “Gather the requirements” on page 362. In this section we do
the following:
Analyze each business question.
Identify high level entities and also possible hierarchies associated with each
of these entities.
Identify high level measures.

The high level entities and measures identified here give us an idea about the
kind of data with which we are dealing. It is highly possible that the entities
identified here could become dimensions or facts in the future phases of the
Dimensional Model Design Life Cycle.

Table 7-7 on page 367 shows the analyzed business requirements and the
identified entities and measures.


Table 7-7 Identified entities and measures requirement table
Q# Sales Invoicing business process High level Measures
requirements entities

Q01 What is the gross profit of product sales Beverage, Gross Profit
grouped by beverage and customer Organization, Amount
category during a given month? Time (Month),
Organization
Category

Q02 Debt analysis by customer: Print active Organization, Outstanding
and outstanding debts at a given date, by Time (Day), Debt (Net
customer and grouped in ranges or bands Country Sales Amount)
of days. The user needs to have a report to
list the sum of payments due in the next
seven days per customer or per customer
country. Also the user needs to report the
total invoiced amount with payment due
longer than 30 days.

Q03 What is total amount of outstanding debts Beverage Outstanding
on a selected date by beverage and by group, Debt (Net
category? And also by customer and Organization, Sales Amount)
invoice number? Invoice, Time
(Day)

Q04 What is the number of units sold and net Beverage, Net Sales
sales amount per product for Product group, Amount,
non-beverage items, per quarter? Time (Quarter) Quantity Sold

Q05 What are the 10 most profitable beverages Beverages, Gross Sales,
by: Time (Year to Gross
1. Largest total sales per year to date Date), Region Margin%
and region.
2. Largest gross margin per year to date
and region.

Q06 Who are my 10 most profitable customers Organization, Gross Sales
per region by largest total sales volume per Time (Year to Amount
year? Date), region

Q07 How is my net profit developing since the Time (Year to Net Profit
beginning of year? (Display profit during Date), Time Amount
current year at the end of month, (Month)
calculating monthly and accumulated profit
per month)



Q08 How many bottles (containers) and liters Time (Month) Quantity Sold,
were delivered per time period in months Liters
compared to the previous three years Delivered
during the same period?

Q09 List the total sale discount per beverage or Beverage, Discount
product group or customer or customer Product Amount
category per time period compared to Group,
previous three years during the same Organization,
period. Organization
Category

Q10 Organization credit history. For a selected Time (month), Net Sales
month interval, list the outstanding Invoices, Amount
payments at the end of each month Organization
grouped by customer and sorted by
outstanding amount.

Q11 Report showing Net revenue, Goss sales, Product Net Revenue
Gross margin, and Cost of sales per Group, Time Amount, Gross
beverage and month, quarter, and year. (Month), Time Sales Amount,
(Quarter), Gross
Time (Year) Margin%, Cost
of Goods

Q12 Analysis of delayed invoicing due to partial Warehouse, Invoice
shipments: Total amount grouped per Product group, Delayed
warehouse, beverage, or beverage group Shipment, Amount
for a selected period where the invoice Invoices
date is a number of days greater than the
first shipping date.
Business reason: If the customer does
not want partial invoicing, the order will be
invoiced once it is fully shipped. Partial
order deliveries will wait to be consolidated
in a full order invoice. Orders may take a
long time to complete, from the time of the
first partial shipment. It is necessary to
analyze the impact of the indirect credits
that the customer gets in this situation, and
the increase in the sales performance if we
improve the shipping process. The
pertaining data needs to be analyzed only
once per month.



Q13 What is the sales order average value and Time (Month), Gross Sales
sales order maximum value per month per Employee Amount
Sales Representatives?

Q14 What customers have placed orders Organization, Gross Sales
greater than a given value during the last Time (Year) Amount
year?

Q15 What are the companies that last year Organization, Outstanding
exceeded their credit by more than 20%? Time (Year) Debt

Q16 List gross sales totals and total discounts Employee, Gross Sales
by Sales Representative for past year Time (Year) Amount,
sorted by gross sales amount in Discount
descending order. Amount

Q17 List the total discount amount, Employee, Discount
percentages and gross margin per Sales Organization, Amount,
Representative and customer sorted by Time (Year) Discount
gross margin in ascending order during Percentage,
last year. Gross
Margin%

Q18 List the number of orders per month Time (month), Number of
containing one or more non-beverage Product Group orders
items since the first year available.

Q19 Gross sales per Sales Representative, Sales Gross Sales
sales office, month, and currency selected Representativ Amount
by year e, Sales Office,
Time (Month),
Time (Year)

Q20 Gross Sales amount per each method of Payment Gross Sales
payment and country during last year. Method, Amount
Country

Q21 Three top weekdays per month according Time Gross Sales,
to average sales by Web orders. The (Weekday), Number of
listing must include total gross sales and Sale Office, Orders
number of orders.
Business reason: For planning system
maintenance, is there a day (or days) of
the week in which the number and orders
increase or decrease? (Sales office =
Web)



Q22 Number of orders and average sales order Order Number of
volume per transmission type (fax, transmission orders
telephone, Web) per month and year. type, Time
(year), time
(month)

Q23 Revenue, gross margin, and cost of goods Organization Gross
during last year per supplier sorted by (supplier), Margin%, Cost
revenue in descending order. Time (Year) of Goods,
Revenue

Q24 For tracking urgent Customer Orders Organization NA
specially ordered to the supplier, a report is (Supplier)
needed listing: Supplier Id, Supplier name,
Contact name, Telephone number,
Discount%, Web Page and Web access
code from the supplier.

Q25 Monthly report for Accounts containing: Organization Net sales
Month, Customer id, Net Sales, Duty, Cost (customer), amount, Duty
of Item handling, Cost of Goods. Time (Month), Amount, Cost
of Item
Handling, Cost
of goods

G01 All sales and cost figures for beverage Currency, All money
reports must be available in both reporting Product Group amount
currency and invoice currency. measures
requested in
the other
requirements

G02 Beverages need to be summarized or Beverage - All measures
added by product group. Product group applicable to
beverages

G03 Allow tracking of average delivered alcohol Organization, Alcohol
liters, maximum and minimum alcohol employee, Content%
content values per Organization, Sales product group Spirit liters.
Representative and/or Beverage group
Business reasons: Restructuring of the
existing marketing product groups, and
supporting resellers with the legal
requirements regarding not selling
beverages to the non-adult population.



G04 Gross Sales, Discount and Bonus totals Organization Gross Sales
per year per corporation. (customer Amount,
corporation) Discount
Amount,
Bonus Amount

G05 Changes to Beverage and Sales Beverages, N.A.
Organization structure and all Product group
corresponding attributes must be tracked
across the time (history needed).

G06 Changes to customer attributes need to be Organization N.A.
fully maintained and stored.

High level entities and measures identified
Table 7-8 shows the high level entities (potential future dimensions), and any
associated hierarchies with the measures identified in Table 7-7 on page 367.

Table 7-8 High level dimension identified
Entities Hierarchy

Beverage Product Group → Beverage

Organization Region → Country → City → Organization
(Customer/Supplier) Organization Category → Organization
Corporation → Organization → Division

Time Year → - Quarter → Month → Day → Time
WeekDay → Day → Time

Employee Region → Country → City → Sales Office → Sales Employee

Warehouse Region → Country → City → Warehouse

Currency Country → Currency

Shipment -

Orders Order Transmission Type - Order

Invoices -

Table 7-9 on page 372 shows the high level measures identified from the
requirements analyzed in Table 7-7 on page 367.


Table 7-9 High level measures
Measures

Alcohol Content%
Cost of goods
Discount Amount
Discount Percentage
Gross Profit Amount
Gross Sales Amount
Gross Margin%
Gross Sales Amount
Invoice Delayed Amount
Liters Delivered
Net Profit Amount
Net Sales Amount
Net Revenue Amount
Number of orders
Outstanding Debt
Quantity Sold
Revenue Amount
Spirit liters
YTD Gross Sales
YTD Gross Margin
Duty Amount
Accrued bonus
Cost of item handling

7.3.8 Business process analysis summary
The final output of the Identify business process phase is the requirements
gathering report. This requirements gathering report contains the following:
Business process listing
Business process prioritization
High level entities and measures common between the business processes
Business process identified for which dimension model will be built
Data sources listing
Requirement gathering for the business process
Requirement gathering analysis
High level entities and measures identified from the requirement analysis


7.4 Identify the grain
In this phase we focus on the second step of the dimensional model design life
cycle, which is Identify the grain, as shown in Figure 7-12.

Determine All
Dimensions
Processes

Identify Conformed




Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management

Figure 7-12 Identify the grain

In 7.3, “Identify the requirements” on page 351, we identified the Sales Invoicing
business process for which we will design the dimensional model. In this phase
we will design the grain definition for the requirements we gathered for the Sales
Invoicing process in 7.3.6, “Gather the requirements” on page 362.

The primary goals of this phase are to:
Define the fact table granularity.
Identify existence of multiple grains within one business process.
Identify type of fact table (Transaction, Periodic, or Accumulating).
Study the feasibility to be able to provide lower grain for fact table than is
requested.
Identify higher level dimensions and measures from the grain definition.


Produce the grain definition report which contains all the results and final
output of this phase.

7.4.1 Identify fact table granularity
We have previously defined the sales process as the action of invoicing an order.
The sales process was discussed in detail in 7.3.2, “Identify business process”
on page 355. Therefore, in this case the natural candidate to look at when
searching for the right granularity is the invoice form. An invoice consists of a
header and one or many line items. A sample invoice form is shown in
Figure 7-13. The invoice form does not show all associated attributes with the
invoice. In this form we find the lowest grain for the fact table.

The grain is identified as a single invoice line in the invoice form (highlighted in
the example in Figure 7-13).

Invoice Number: 10078902 01.10.2005

Customer
Order Number: 123678 Order date: 21.09.2005

Customer: Discount Cust-1, Limited
Delivery date: 01.10.2005
Sales Rep: SalesPerson-1

Price/ Qty Disc Item Acquired
Line # Item name Unit Sold % Price Bonus

1 Oak White-1 0.75l 12.00 50 10% 540.00 10

2 Alcohol Bev-1 0.5l 16.00 10 5% 152.00 10
3 Bottle Water 5.00 100 0% 500.00 50
Pack 6 x 0.33l
4 Soda 10.00 20 0% 200.00 100
Value Pack 2 x 0.75
Currency: EU 1392.0
Total: 0

Figure 7-13 The grain of the sales fact table

Note: The grain you choose determines the level of detailed information that
can be made available to the dimensional model.


7.4.2 Identify multiple separate grains
The primary focus of this activity is to determine if there are multiple grains
associated with the Sales Invoicing business process for which we are designing
a dimensional model.

There will be times when more than one grain definition is associated with a
single business process. In such a scenario, we recommend to design separate
fact tables with separate grains rather than to forcefully try to put facts that belong
to separate grains in a single fact table.

In short, this activity will allow us to decide if there are multiple grains associated
with the sales process, and if we need to implement one or multiple fact tables.

In order to decide whether to go for one or multiple fact tables, we suggested
criteria in 6.3.2, “Multiple, separate grains” on page 249. Here are those criteria:
One of the most important criteria that helps us determine the need for one or
multiple fact tables are the facts. It is very important to understand the
dimensionality of the facts to make the decision of whether the facts belong
together in one fact table or in separate fact tables with different grains. For
example, let us consider the sales business process.
We have found two facts Outstanding Debt and Invoice Delayed Amount that
belong to the sales process as a whole, but are not true at the grain definition
at the line item level we have chosen. We will therefore exclude them now and
leave them for a subsequent iteration.

Multiple granularity: Facts such as Number of Orders, Invoice Delayed
Amount, and Outstanding Debt will be implemented in separate fact tables.

Other criteria to consider is whether or not multiple OLTP source systems are
involved. Generally if we are dealing with business processes such as order
management, store inventory, or warehouse inventory, it is very likely that
separate source systems are involved and hence the use of separate fact
tables may be appropriate. However, in our Sales Invoicing business process,
we have only one source of data and that is the enterprise data warehouse.
It is also important to find out if multiple business processes are involved.
Multiple business processes involve the creation of multiple separate fact
tables. And, it is possible that a single business process may involve creation
of separate fact tables to handle facts that belong to different granularity. For
our dimensional model, we have chosen only one business process and that
is Sales Invoicing.


Note: It is important to understand that had we chosen the sales process as a
whole, we would have to design separate dimensional models for various
subprocesses inside the sales process. In 7.2.8, “General business process
description” on page 344, we discussed the various subprocesses inside the
sales process, which were Sales Invoicing, Sales Orders, and Sales
CRM/Marketing.

Another important criteria to consider are the dimensions. If you find that a
certain dimension is not true to the grain definition, then this is a sign that it
may belong to a new fact table with its own grain definition.

Based on the assessment factors discussed above, we decide that there would
be one fact table to handle the Sales Invoicing data. And, facts such as Number
of Orders, Invoice Delayed Amount, and Outstanding Debt will be implemented
in separate fact tables.

Are multiple grain definitions identified?
Yes, there are multiple grain definitions identified for the Sales Invoicing process.
And, they are:
A single line item on the invoice form: In this chapter, we will be designing the
dimensional model for this grain definition only. The other grain definitions
defined below will be handled separately in different fact tables.
Outstanding debt for each customer.
Single delayed invoice for a customer on an order.

Note: It is extremely important not to merge facts or dimensions which are
true at different grains into one fact table. Instead, a new dimensional model
should be created with a new grain.


7.4.3 Identify fact table types
In this section we identify the types of fact tables. There are basically three types
of fact tables, and they are:
Transaction fact table: A table which records one row per transaction. A
detailed discussion about the transaction fact table is available in “Transaction
fact table” on page 125.
Periodic fact table: Stores one row for a group of transactions. In other words,
a periodic fact table stores a single row for a number of transactions over a
period of time. A detailed discussion about the periodic fact table is available
in “Periodic fact table” on page 126.
Accumulating fact table: Stores one row for the entire lifetime of an event. For
example, the lifetime of a credit card application being sent, until the time it is
accepted by the company. Another example is the lifetime of a job or college
application being sent, until the time it is accepted or rejected by the job
posting company or college. A detailed discussion about the accumulating
fact table is available in “Accumulating fact table” on page 127.

We also recommend that you review 6.3.3, “Fact table types” on page 250 for
more details.

We now know that the grain of our sales fact table is the invoice line. Invoices are
generated once and they never get updated again. If changes are made, they
have to be handled through credit notes.

Invoice lines are considered transactions, and one invoice line item will represent
one row in the fact table. In other words, each single item on an invoice will form
a part of the fact table.

Fact table type: For the Sales Invoicing business process, we have identified
three separate grains. However, for this case study, we will be working on the
first grain definition which is a single item on an invoice. The fact table for this
grain is a transaction fact table.

Regarding the Sales Invoicing process, we have already chosen the lowest grain
possible, which is the invoice line.

Both the dimension table and fact table have a grain associated with them. To
understand the grain of a dimension table, we need to understand the attributes
of the dimension table. Every dimension has one or more attributes. Each
attribute associates a parent or child with other attributes. This parent-child
relationship provides different levels of summarization. The lowest level of


summarization or the highest level of detail is referred as the grain. The
granularity of the dimension affects the design such as the retrieval of data and
data storage.

Grain adjustment: For our sales invoice grain definition, we have chosen the
most detailed atomic grain as a single line item on the invoice. Lowering this
grain definition is neither necessary nor possible.

7.4.5 Identify high level dimensions and facts
In this activity, we identify high level preliminary dimensions and facts from
whatever can be understood from the grain definition. No detailed analysis is
carried out to identify these preliminary dimensions and facts.

For our Sales Invoicing business process, we defined the grain (see 7.4.1,
“Identify fact table granularity” on page 374) as a single line item on the invoice.
An example of this is shown in Figure 7-14.

Customer
Time Customer Representative

Grain: 1) Price per item
2) Quantity
1 line item on an invoice
3) Discount%

Invoice Date Currency
Number

Figure 7-14 Identifying high level dimensions and facts from the grain

From the invoice form shown on Figure 7-14, we have identified several facts and
dimensions. Those preliminary facts and dimensions are listed in Table 7-10.

Table 7-10 Preliminary facts and dimensions
Name Output

Preliminary facts Quantity, Price per Item, Discount%

Preliminary dimensions Invoice, Customer, Time, Beverage,
Customer Representative, Currency


These preliminary high level dimensions and facts are helpful when we formally
identify dimensions (see 7.5, “Identify the dimensions” on page 380) and facts
(see 7.6, “Identify the facts” on page 409). The dimensions and facts get
iteratively refined in each of the formal phases of the DMDL.

Note: Preliminary facts are facts that can be easily identified by looking at the
grain definition. For example, facts such as Price per Item, Quantity, and
Discount% are easily identifiable by looking at the grain. However, detailed
facts such as cost, manufacturing cost per line item, and transportation cost
per item, are not preliminary facts that can identified by looking at the grain
definition. Preliminary facts are not the final set of facts. Formal detailed fact
identification occurs in the Identify the facts phase in 7.6, “Identify the facts” on
page 409. The same is true for preliminary dimensions.

7.4.6 Grain definition summary
The output of the Identify the grain phase is the Grain Definition Report, which
contains information such as that shown in Table 7-11.

Table 7-11 Main output of the Identify the grain phase
Name Output

Fact table granularity Invoice line

Are there multiple grains Yes, there are multiple grain definitions identified for
within one business process? the Sales Invoicing process. They are:
A single line item on the invoice: In this chapter,
we will be designing the dimensional model for
this grain definition only. The other grain
definitions will be handled separately in different
fact tables. However, they are out of the scope of
this chapter.
Outstanding debt for each customer.
Single delayed invoice for a customer on an
order.

Type of fact table Transactional (Grain is a single line item on an
invoice.)

Check grain atomicity No
(Necessary to lower it?)

Preliminary facts Quantity, Price per Item, and Discount%

Preliminary dimensions Invoice, Customer, Time, Beverage, Customer
Representative, and Currency


7.5 Identify the dimensions
In this section, we focus on the Identify the dimensions phase of the dimensional
model design life cycle, depicted in Figure 7-15.

Determine All
Dimensions
Processes

Identify Conformed




Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


Table 7-12 shows the activities that are associated with the Identify the
dimensions phase.
Table 7-12 Activities in the Identify the dimensions phase
Seq Activity name Activity description
no.

1 Identify dimensions Identifies the dimensions that are true to the
identified grain.

2 Identify degenerate Identifies one or more degenerate dimensions.
dimension


Seq Activity name Activity description
no.

3 Identify conformed Identifies any existing shared dimensions present in
dimensions the data warehouse or other star schemas that may
be used for designing the dimensional model.

4 Identify dimensional Identifies the various dimension attributes for the
attributes and dimension. It also identifies the balanced,
dimensional unbalanced, or ragged hierarchies that may exist in
hierarchies the dimensions. Techniques are suggested to handle
these hierarchies in the design.

5 Identify date and time This activity identifies the date and time dimensions
granularity in the dimensional design. Typically it is these
dimensions that have a major impact on the overall
grain and size of the dimensional model.

6 Identify slowly Identifies the various slowly changing dimensions in
changing dimensions the design. Also three major techniques (Type-1,
Type-2, and Type-3) are described for handling
slowly changing dimensions.

7 Identify very fast Identifies very fast changing dimensions and
changing dimensions describes ways of handling such dimensions by
creating one or more mini-dimensions.

8 Identify cases for Identifies which dimensions need to be snowflaked.
snowflaking

9 Other dimensional Description of other challenges:
challenges are:

→ Identify Looks for multi-valued dimensions and describes
Multi-valued ways of handling such dimensions in the design by
Dimensions using bridge tables.

→ Identify Describes ways of looking for dimensions that can
Role-playing be implemented using the role-playing concept.
Dimensions

→ Identify Describes ways of identifying heterogeneous
Heterogeneous products and implementing them in the design.
Dimensions

→ Identify Garbage Describes ways to look for low-cardinality fields and
Dimensions use them for making a garbage dimension.

→ Identify Hot Describes ways of creating profile-based tables or
Swappable hot swappable dimensions to improve performance
Dimensions and secure data.


Note: We have listed activities to consider when designing dimensions. The
purpose is to make you aware of several design techniques available to use,
depending on the particular situation.

7.5.1 Identify dimensions
During this phase we identify the dimensions that are true to the grain selected in
7.4, “Identify the grain” on page 373. We were able to identify preliminary
dimensions and facts just by looking at the grain definition in 7.4.5, “Identify high
level dimensions and facts” on page 378. However, there are typically several
dimensions that are true to the grain but cannot be identified by looking at the
grain definition. The same is true for facts. The goal of this phase is to formally
identify all dimension tables that are true to the grain definition.

The dimensions identified for the grain definition (a single line item on an invoice)
are shown in Table 7-13.

Table 7-13 Dimensions list
Identified Granularity of the dimension Source table
dimension

Invoice Identifies each invoice. INVOICE
(Typically invoice numbers are treated as
a degenerate dimension. We discussed
this in 6.4.2, “Degenerate dimensions”
on page 266. However, the Invoice
number or invoice id is not treated as a
degenerate dimension. We discuss this
in 7.5.3, “Identify degenerate
dimensions” on page 384.)

Customer Contains information about the CUSTOMER -
customer. A customer for the Redbook CUSTOMER_DETAIL
Vineyard is an organization.

Sales date Contains all dates on which the products DAY
were sold. The date should be stored at
the day granularity.

Beverage Contains all beverages. PRODUCT,
(Product) PRODUCT_DETAIL

Sales Contains the information about the Sales SALES_REP
Representative Representatives.

Currency Contains the currency. CURRENCY


Identified Granularity of the dimension Source table
dimension

Warehouse Consists of information for all WAREHOUSE
warehouses that supply the products.

The preliminary dimensional schema is shown in Figure 7-16. The dimension
attributes are notated in the dimensional as TBD (To be Determined).

*Preliminary Facts

Figure 7-16 Preliminary Sales Invoicing star schema


Note: How to identify dimensions from an E/R model

As previously discussed, the source of a dimensional model can be either the
enterprise data warehouse or the OLTP source systems. Both the data
warehouse and the OLTP source systems may be based on an E/R model.
That is, both may be in third normal form. So, once you learn to create a
dimensional model from an E/R model, then you can create dimensional
models either from the data warehouse or directly from the OLTP source
systems.

The following are the steps involved in converting an E/R model into a
dimensional model:

1: Identify the business process from the E/R model.
2: Identify the many-to-many tables in E/R model to convert to fact tables.
3: Denormalize the remaining tables into flat dimension tables.
4: Identify the date and time from E/R model.

For more detail about this process, see 5.1, “Converting an E/R model to a
dimensional model” on page 104.

7.5.2 Check for existing conformed dimensions
This step in our case study is relatively simple. This is the first dimensional model
implemented in the Redbook Vineyard company, and therefore no other
processes exist which could have shared conformed dimension that we might
use for our dimensional model. For more details, see 6.4.3, “Conformed

7.5.3 Identify degenerate dimensions
After we have identified dimensions, the next step is to identify degenerate
dimensions (dimensions without attributes). The term degenerate dimension is
not actually a dimension, but is often a form of a transaction number, from the
OLTP source system, that has been placed the fact table. Degenerate
dimensions are discussed in more detail in:
“Identify degenerate dimensions” on page 384
“Degenerate dimensions” on page 134

Coming back to our Sales Invoicing case study, we identified the dimensions in
Table 7-13 on page 382. We discussed earlier that a transaction number, such as
invoice number, is a degenerate dimension which is stored in the fact table as a


number. However, in our case study, we observe that the invoice number is not a
number, but has several attributes associated with it. Attributes are:
Order Number associated with the invoice
First delivery date
Order opening date
Invoice date (note that invoice date is different from sales date)
And more

We discuss more such attributes associated with the Invoice number next.

7.5.4 Identify dimensional attributes and hierarchies
In this activity we identify:
Dimensional attributes for the dimensions identified in “Identify dimensions”
on page 382
Hierarchies, such as balanced, unbalanced, and ragged, associated with
each of the dimensions

The purpose of this activity is to identify and define the attributes needed for each
of the dimensions. These attributes are derived from the requirements collected
in 7.3.6, “Gather the requirements” on page 362. More details are available in
6.4.4, “Dimensional attributes and hierarchies” on page 269. In addition to the
attributes, we also identify all possible associated hierarchies with each of these
dimensions.

In 7.5.1, “Identify dimensions” on page 382, we designed the preliminary star
schema shown in Figure 7-16 on page 383.

Now let us fill this preliminary star schema with detailed dimensional attributes for
each dimension. But first, let us define the granularity of each dimension table as

Table 7-14 Dimensions list with granularity
Identified Granularity of the dimension
dimension

Invoice This dimension identifies each invoice and contains information
such as order number and order date associated with a particular
invoice number.
(Typically invoice numbers are treated as a degenerate dimension.
We discussed this in 6.4.2, “Degenerate dimensions” on page 266.
However, in this case study, the Invoice number, or invoice id, is not
treated as a degenerate dimension. We discuss this in 7.5.3,
“Identify degenerate dimensions” on page 384.)


Identified Granularity of the dimension
dimension

Customer The customer dimension contains information about the customer.
A customer for the Redbook Vineyard is an organization.

Sales date The sales date dimension contains all dates on which the products
were sold. The date should be stored at the day granularity.

Beverage The beverage dimension contains all beverages.
(Product)

Sales This dimension contains the information about the Sales
Representative Representatives.

Currency The currency dimension contains the currency used.

Warehouse The warehouse dimension consists of information for all
warehouses that supply the products.

The dimensions, along with their detailed attributes, are described in the
following list:
Invoice dimension: Contains information about the invoices for Customer
Orders that have shipped. In other words, the granularity of the invoice
dimension table is at single invoice header or invoice number. The invoice
dimension is shown in Table 7-15.

Table 7-15 Invoice dimension’s attributes
Attribute Description

Invoice Key It is a surrogate key which is a system-generated integer number.

Invoice Id Identifier of the invoice.

Customer The order reference number used by the customer to order the
Order Id respective products. It will normally be a meaningless number such
as KJS8374L.

First Delivery Deliver date. When there are partial deliveries, it is the date when the
Date first shipment takes place.

End Buyer Id The company or person to whom the beverages are resold. This
might be a company tax identifier or the name of a person.

Invoice Date Date when the invoice is printed and sent to the customer.

Order Date Date the order was opened.

Order Key Surrogate key uniquely identifying an order dimension instance.



Payment Due Date when the payment of the invoice is due. This date can be the
Date same as when the last shipment of the order was or can be later.

Payment Date when the invoice was paid by the customer.
Effective Date

Payment The type of payment method, for example, credit card, cash, bank
Method transfer, and check.

Customer dimension

The customer dimension contains data about the organizations, normally
resellers, to whom the company sells. It also describes the hierarchical structure
of companies or corporations. In other words, the organizations which buy from
the Redbook Vineyard company are its customers. The customer dimension is

Table 7-16 Customer dimension’s attributes

Customer A surrogate key, which is a system-generated integer number. This
Key number uniquely identifies each customer.

Name The name of the organization or the customer. For example,
Composed Gupta and Gupta, Inc., or Ballard Halogen Lights, Inc.

Acronym The acronym or short name by which the company is known, for
example, CGGS for Composed Gupta and Gupta, Inc., or BHL for
Ballard Halogen Lights, Inc.

Allows It specifies if partial shipment is allowed in case the order cannot be
Partial fully shipped at once. If the flag is true and one or more of the ordered
Shipment products is in stock, then the order will be partially shipped.

City The city where the customer has its main address, for example, New
Delhi or Prague.

Comments Any comment pertinent and of value to the reporting about the
customer that cannot go in any of the existing fields.

Company The identifier of the company used for tax purposes. This identifier is
Tax Id unique for every customer and can be used for identifying the customer
across its history in the customer dimension table.

Country The country of the customer. For example, United Kingdom or
Australia.



Delivery The preferred delivery method for the organization.
Method

Financial The name of the internal financial group within the customer.
Group

Marketing Marketing customer category.
Category

Parent It maps to the surrogate key of the parent organization to which the
Organization customer belongs. This is used to store information about the
Key
hierarchical structures of corporations. For example, CorpHQ France
mapping to the parent CorpHQ, which is the top parent company
worldwide, or Buy and Drink Overseas Department, belonging to Buy
and Drink, Inc.

Region The region, as defined by the company marketing organization, where
the customer country is located, for example, EMEA for Morocco, or
SouthAmerica for Argentina,

State The state, province, department, or canton, where the customer has
their main address. For example, California in USA, Provence in
France, or Valais in Switzerland.

Type The type of the customer. For example, small, medium, or large.

Sales Date dimension: The date on which the product was ordered. This
table could be progressively populated as new orders arrive, or we could also
make a one time insertion with the all the date-month-year combinations for a
particular time period. The granularity of the sales date dimension is a single
day.

There are more details about the importance of choosing a correct granularity for
the date dimension in 6.4.5, “Date and time granularity” on page 279. The sales
date dimension attributes are shown in Table 7-17.

Table 7-17 Sales date dimension attributes

Sales Date Key Surrogate key uniquely identifying a Sales Date dimension
instance.

Day Date indicating an exact day, such as December 1, 2010.

Week Day Day of the week, such as Monday or Tuesday.



Month Name of the month, such as January or February.

Quarter The quarter of the year: 1, 2, 3, or 4.

Year It indicates a year, such as 1994 or 2008.

For the case study on the Sales Invoicing business process, we discussed the
sales date dimension in 7.5.6, “Date and time dimension and granularity” on
page 399.
Product dimension: Describes the products or beverages sold by the
company. It currently only contains the drinkable products also commonly
known as beverages. The product indicates what drink is contained in what
type and size of container. For example, a soda in a one liter glass bottle, or in
a 33cl. aluminum can. The company also sells other non-beverage products,
such as T-Shirts and glasses, but they reside in a different sales fact table and
are not covered in this process.

The granularity initially requested in the case study was at the beverage level.
For example, just specifying a brand of lemonade or a brand of soda.
However, we should always try to go to the lowest level of available granularity
to enable handling more demanding requirements that might exist in the
future. The product granularity is a single product, which is defined as a
particular beverage with a particular volume and a particular presentation
form. The product dimension table attributes are defined in Table 7-18.

Table 7-18 Product dimension attributes

Product Key Surrogate key, which is a system-generated integer.

Name The name of the product.

Acronym A short name, or initials, for a beverage, such as ATS for Amit Tasty
Soda.

Beverage Identifies the beverage, such as Amit Tasty Soda or RedBook
Lemonade.

Beverage The group of drinks to which a beverage belongs, for example, Soft
Group Drinks.

Description An adequate description of the product with important characteristics
that do not have a corresponding field in the product table.



Financial Which financial group includes this product.
Group

Presentation The type of container used, such as a bottle or value-pack.

Product Id The identifier of the product, commonly known to the users, or as
generated in the source system. Normally a number or code such as
SOD-40-L, not necessarily with any particular meaning.

Status The status of the product, such as active or inactive.

Supplier The name of the supplier of the products, such as Soda Supplier-1.

Units Per The number of units shipped in simple packages.
Package

Units Per The number of units of product that can be shipped per pallet or crate.
Pallete

Vat Code The VAT or Tax code applied to the product, such as 5%, or 20%.

Sales Representative Dimension

The Sales Representative dimension contains data about the company sales
force. The Redbook Vineyard management is keenly interested in tracking the
performance of their Sales Representatives. The granularity of the Sales
Representative dimension is a single Sales Representative. The Sales
Representative dimension attributes are shown in Table 7-19.

Table 7-19 Sales Representative dimension attributes

Sales Surrogate key which is a system-generated integer. This number
Representative uniquely identifies each employee.
Key

Name Employee name.

Comments Any additional comments required about the Sales
Representative.

Manager Name Name of the manager of the sales employee.

Employee Id A number or code identifying the employee.

Sales Office Sales office where the customer representative works.



City Sales office city location.

Country Name of the country where the sales office is located.

Warehouse dimension: This dimension contains information about the
warehouse from which the ordered products are shipped. The granularity of
the warehouse is a single warehouse at a location. Of course, each single
warehouse is actually structured in areas, rooms, and shelves. To better track
which product came from which shelf of the warehouse, we could keep the
grain of the warehouse dimension to shelf-location within a warehouse.
However, this atomic level detail of information was not available in the
original ERP system from where the data warehouse was populated. The
warehouse dimension attributes are described in Table 7-20.

Table 7-20 Warehouse dimension’s attributes

Warehouse Surrogate key which is a system-generated integer. This number
Key uniquely identifies each warehouse.

Name Name of the warehouse.

Acronym A short name for the warehouse.

City The name of the city where the warehouse is located.

Country The country where the warehouse is located.

Description Text describing additional features and characteristics, or special
remarks.

Currency dimension: This dimension contains a list with the currencies used
in local orders and invoices. It also identifies which is the currency used for
reporting and consolidation purposes. The granularity of the currency
dimension is a single currency. The currency dimension table attributes are
described in Table 7-21.

Table 7-21 Currency dimension’s attributes

Currency Key Surrogate key uniquely identifying a currency dimension instance.

Currency The official currency name, such as Sterling Pound and Swiss
Name Franc.



Description The description of the currency.

Is Corporate This flag indicates the currency used for consolidating financial
Currency? statements corporate and worldwide. Only one currency can be the
corporate currency, but it can change across the Sales Date.

ISO Code The ISO code by which the currency is known. For example, USD
for US dollars.

The dimension schema we get after identifying the dimensions attributes is

*Preliminary Facts
(Iteratively formalized in
the ‘Identify Facts’ Phase)

Figure 7-17 Star schema after identifying dimension attributes

In the next activity, we further identify the hierarchies that each dimension has
and also identify what attributes of the hierarchies.


7.5.5 Identifying the hierarchies in the dimensions
basically consists of different levels, with each level corresponding to a
dimension attribute.


There are three major types of hierarchies, and they are described in Table 7-22.

S.No Hierarchy Hierarchy description How is this hierarchy
name implemented?

1 Balanced A balanced hierarchy is a See “How to implement a
hierarchy in which all the balanced hierarchy” on
dimension branches have the page 144.
same number of levels. For
more details, see “Balanced

2 Unbalanced A hierarchy is unbalanced if it See “How to implement
has dimension branches an unbalanced hierarchy”
containing varying numbers of on page 146.
levels. Parent-child dimensions
support unbalanced
hierarchies. For more details,
see “Unbalanced hierarchy” on
page 145.

least one member whose ragged hierarchy in
parent belongs to a hierarchy dimensions” on page 155.
that is more than one level
above the child. Ragged
dimensions, therefore contain
branches with varying depths.
For more details, see “Ragged

Note: A dimension table may consist of multiple hierarchies, as well as
attributes or columns that belong to one, more, or no hierarchies.


For the Sales Invoicing example, Table 7-23 shows the various hierarchies
present in the dimensions.

Table 7-23 Dimensions and hierarchies
Seq. Dimension Hierarchy description Type of
no. name hierarchy

1 Sales Date Year → Quarter → Month → Week Day Balanced

2 Invoice None

3 Product Financial Group → Beverage Group → Balanced
Beverage

4 Sales Sales Rep Country → Sales Rep Balanced
Representative Region → Sales Rep Office

5 Currency None

6 Customer Customer Country → Customer Balanced
Region → Customer City

7 Customer Customer has an unbalanced hierarchy. Unbalanced

8 Warehouse Warehouse Country → Warehouse Balanced
Region → Warehouse City

A detailed discussion about handling hierarchies is given in 5.3.4, “Handling
dimension hierarchies” on page 142.

Note: Geographical objects, such as cities, countries, and regions, should
always be considered as part of a Geography hierarchy.

A hierarchy is unbalanced if it has dimension branches containing varying
numbers of levels. Parent-child dimensions support unbalanced hierarchies. For
example, the parent-child relationship present in the Customer dimension (see
Figure 7-18 on page 395) supports an unbalanced hierarchy as shown in


Figure 7-18 Recursive pointer to show parent-child relationship

It is important to understand that representing an arbitrary, unbalanced hierarchy
is an inherently difficult task in a relational environment. A common example of
an unbalanced hierarchy is one that represents the parent-child companies. A
company at a given level may have several smaller companies below it, while
others at the same level may have none or a few. If you wish to create a report
which computes the sales totals for all companies at a given level (see levels in
Figure 7-19 on page 396), you will find this question more efficiently answered
with an OLAP (cube-based) reporting system than with a relational database.

However, if your hierarchies are symmetrical, then arguably either type of
technology (OLAP Cubes or Relational database) is equally capable of providing
the answer. For example, the date hierarchy is completely symmetrical. Each
year has the same number of quarters, each quarter has the same number of
months, and each month has same number of weeks in it.


Level 1
Customer Company 1

Customer Company 2 Customer Company 3 Level 2

Customer Company 4 Customer Company 5 Customer Company 6 Level 3

Customer Company 7 Level 4

Figure 7-19 Unbalanced hierarchy represented by the Customer table

The customer dimension is an unbalanced hierarchy, as shown in Figure 7-19.
It has levels with a consistent parent-child relationship, but that have logically
inconsistent levels. The hierarchy branches also can have inconsistent depths.
An unbalanced hierarchy can represent a parent-child company.

Let us assume that we want a report which shows the total sales revenue at the
node Customer Company 2. What this means is that we want the report to show
the total sales made to Customer Company 2, Customer Company 4, Customer
Company 5, Customer Company 7, and Customer Company 8. We cannot use
SQL to answer this question because the GROUP BY function in SQL cannot be
used to follow the recursive tree structure downward to summarize an additive
fact such as sales revenue. Therefore, because of this problem, we cannot
connect a recursive dimension (Figure 7-18 on page 395) to any fact table.

So, how do we solve the query using an SQL Group-by clause? And, how can
we summarize the sales revenue at any node?

The answer is to use a Bridge table between the Customer and SALES table as
shown in Figure 7-20 on page 397. The aim of the bridge table is to help us
traverse the unbalanced hierarchy.

The customer table will have a hierarchy built in itself. For example, Customer A
is a corporation that owns Customer B and C, and Customer D is a division of B.
In our dimensional model, the Fact Gross Sales Amount for customer A will not
include any of the sales amounts relative to B, C, or D. The same applies to
every single child in the hierarchical tree.


For traversing the Customer
Hierarchy, join the following:

(1) Descending down the Hierarchy
- [Customer Key] Joins [Parent
Organization Key] BRIDGE
and
- FACT [Customer Key] Joins [Subsidiary

(2) Ascending up the Hierarchy
- [Customer Key] Joins [Subsidiary
and
- FACT [Customer Key] Joins [Parent

Figure 7-20 Organization bridge table for level navigation

There are two ways in which you can move through the unbalanced hierarchy
Descending the hierarchy: For this you need to make joins between
Customer, Sales, and Organization_Hierarchy_Bridge as shown in
Figure 7-20. We show one example of such a query in Example 7-1 on
page 398.
Ascending the hierarchy: For this you need to make joins between Customer,
Sales, and Organization_Hierarchy_Bridge as shown in Figure 7-20.

The bridge table is called this because, when running queries and joining the
facts table with the respective dimension, this new table will have to be placed in
the middle of the other two tables and used as a bridge, as you can see in


Figure 7-21 Graphical presentation of a bridge table

Figure 7-21 shows a graphical query where we calculate the total number of
customer sales to the IBM corporation, including all its branches and owned
companies, since the beginning of the data mart history.

And here is the SQL code produced by the tool used to design and run that
query.

Example 7-1 SQL for traversing down the hierarchy
SELECT Sum(DWH_SALES.GROSS_SALES_AMOUNT) AS Sum_Of_GROSS_SALES_AMOUNT
FROM DWH_SALES INNER JOIN
(DWH_CUSTOMER INNER JOIN
DWH_ORG_HIER_BRIDGE
ON DWH_CUSTOMER.CUSTOMER_KEY=
DWH_ORG_HIER_BRIDGE.PARENT_ORGANIZATION_KEY)
ON DWH_SALES.CUSTOMER_KEY = DWH_ORG_HIER_BRIDGE.SUBSIDIARY_ORGANIZATION_KEY
WHERE DWH_CUSTOMER.NAME Like "IBM*" AND DWH_ORG_HIER_BRIDGE.TOP_FLAG="Y";

For more detail about how to implement and traverse a parent-child hierarchy as
shown in Figure 7-19 on page 396, refer to “How to implement an unbalanced

The dimensional model we have developed to this point, after having identified
the various hierarchies, is shown in Figure 7-22 on page 399.


Bridge Table is used to traverse
the Customer Unbalanced Hierarchy

Figure 7-22 Dimensional model developed after identifying hierarchies

7.5.6 Date and time dimension and granularity
The date and time dimensions are considered the most important dimensions in
dimensional modeling. The identification of the grain is crucial for the success of
the dimensional database. For more details, see 6.4.5, “Date and time
granularity” on page 279. The granularity specifically required for the Sales
Invoicing dimensional model is a single day. Although a lower granularity would
be possible, down to the timestamp level, the information in the data warehouse
is at the date level, and a lower granularity is not considered to have any
additional business value. However, if our data mart was for the ordering
process, then a granularity at a timestamp level could have an additional
business value for analysis, for example, for orders received through the Web.

The requirements for the sales invoice process can be fulfilled by the date
dimension as shown in Table 7-24 on page 400.


Table 7-24 Example of Sales Date table dimension
Day Month Quarter Week day Year

3/10/2005 3 1 4 2005
3/11/2005 3 1 5 2005
3/12/2005 3 1 6 2005
3/13/2005 3 1 7 2005

7.5.7 Handling slowly changing dimensions
The Redbook Vineyard stores full history for product, warehouse, customer,
Sales Representative, invoice, and currency in the Enterprise Data Warehouse.
Slow changing dimensions do not normally represent a problem, and we only
need to decide what type of change management strategy we are going to use:
Type-1, Type-2, or Type-3.

In the case study, the product, warehouse, Sales Representative, currency, and
customer dimensions are slowly changing dimensions, but they have different
requirements. The change handling strategies are listed in Table 7-25.

Table 7-25 Slowly changing dimensions
Dimension Requirements Change handling
strategy

Product Keep the history on product structure and Type-2
corresponding attributes.

Warehouse No special requirement. Information can Type-1
be overwritten.

Sales Keep the history on sales structure and Type-2
representative corresponding attributes.

Invoice The only changes to this dimension will be Type-1
the population of some date fields. But
once populated, the fields will not change.

Sales date This dimension will not change. NA

7.5.8 Handling fast changing dimensions
In this section we identify the very fast changing dimensions that cannot be
handled using the Type-1, Type-2, or Type-3 approaches discussed in 6.4.6,
“Slowly changing dimensions” on page 283.


Fast changing dimensions are not so easy to handle, and normally require a
different approach than the Type-1, Type-2, and Type-3 strategies. The
information contained in fast changing dimensions is subject to change with a
relatively high frequency. Table 7-26 shows the dimension identified as fast
changing.

Table 7-26 Fast changing dimensions
Dimension Requirements Type

Customer Keep the history for payment conditions, and Fast Changing
credit and discount limits. Dimension

An approach for handling very fast changing dimensions is to break off the fast
changing attributes into one or more separate dimensions, also called
mini-dimensions. The fact table would then have two foreign keys—one for the
primary dimension table and another for the fast changing attributes. For more
detailed discussion about handling fast changing dimensions, refer to 6.4.7, “Fast
changing dimensions” on page 286.

We identify the following fast changing attributes for the customer dimension
shown in Figure 7-23:
Days for Payment
Credit
Discount
Bonus

Fast Changing
Dimension Attributes

Figure 7-23 Customer with detailed financial conditions

For the fast changing attributes, we create a mini-dimension called Payment
Terms. The Payment Terms dimension is shown in Table 7-27 on page 402. In


order to handle the fast growth, we create this mini-dimension with attributes
which have a range of values. In other words, each of the attributes has a value
in a particular band.

Table 7-27 Payment terms dimension table

Payment Surrogate key uniquely identifying a payment term dimension
Terms Key instance.

Credit Band The credit band groups different ranges of credit amounts allowed for
the customer. For example:
1 Low credit: Less than 1000;
2 Medium credit: Between 1000_ and 4999;
3 High credit: Between 5000_ and 10000;
4 Extra high credit: Greater than 10000.

Days For This band groups the different times or plans for invoice payments.
Payment For example, immediate: 1 - 30 days; more than 30 days. The
Band customer can arrange a payment period in multiples of 15 days, or
arrange for an immediate payment. In theory the customer could have
to pay inmediately, in 15, 30, 45, or 60 days.

Discount This describes the bands grouping the different discount ranges. For
Band example, no discount; less than 10%; between 10% and 30%; more
than 30%.

Bonus Band This describes the bands grouping the different bonus ranges. For
example: No bonus; less than 5%; between 5% and 10%; and more
than 10%.

Figure 7-24 on page 403 shows the new mini-dimension (Payment Terms) with
the customer financial conditions and the updated customer dimension.


Mini-Dimension with each of its
attributes having a band-range
of values

Figure 7-24 Mini-dimension for customer financial conditions

Each of the dimensions (Customer and Payment Terms) are joined separately to
the Sales fact table. The dimensional model developed at this point (after
identifying slowly and fast changing dimensions) is shown in Figure 7-25 on
page 404.


Figure 7-25 Model after identifying fast and slowly changing dimensions

7.5.9 Identify cases for snowflaking
In this section we identify potential cases for snowflaking for all the dimension
tables. Table 7-28 shows the snowflakes that we have identified and for what
reasons.

Table 7-28 Snowflaking actions per dimension
Action Objects Reason

Snowflaking Supplier There is a significant number of attributes which are
from strictly supplier dependent. In other words, the
products supplier has several attributes which are at a
different grain than the overall product table. It is
very likely that this table will be used as a dimension
by the Procurement department with their future
data mart, which will have a different fact table, but
will share some conformed dimensions. The
supplier dimension then would be shared with the
procurement business process dimensional model.


Refer to 5.3.7, “Identifying dimensions that need to be snowflaked” on page 171,
for more topics on snowflaking, such as:
When to snowflake
When to avoid snowflaking
When to snowflake to improve performance

Table 7-26 on page 401 depicts the dimensional model after we have identified
the dimensions that will be snowflaked.

Figure 7-26 Model after identifying snowflake dimensions

7.5.10 Handling other dimensional challenges
In this section we identify special types of dimensions that may be applicable to
the Sales Invoicing case study. The special dimensions that we identify are


Table 7-29 Other special dimensions to identify
Seq Dimension type How is it implemented?
no.

1 Multi-valued Description
dimension Typically, while designing a dimensional model, each
dimension attribute should take on a single value in the
context of each measurement inside the fact table.
However, there are situations where we need to attach a
multi-valued dimension table to the fact table. In other
words, there are situations where there may be more than
one value of a dimension for each measurement. Such
cases are handled using multi-valued dimensions.

Implementation
Multi-valued dimensions are implemented using Bridge
tables. For a detailed discussion, refer to 5.3.10,
“Multi-valued dimensions” on page 182.

Sales Invoicing case study
For the case study on Sales Invoicing, we do not have any
scenarios with multi-valued dimension. This is primarily
because each attribute in each of the dimensions takes on
a single value for a particular fact table row.

2 Role-playing Description
dimension A single dimension, which is expressed differently in a fact
table using views, is called a role-playing dimension. A
date dimension is typically implemented using the
role-playing concept when designing a dimensional model
using the accumulating snapshot fact table. This is
discussed in more detail in “Accumulating fact table” on
page 127.

Implementation
The Role-Playing dimensions are implemented using
views. This procedure is explained in detail in 5.3.9,

For the case study on Sales Invoicing, we do not have any
scenarios for dimensions that can be implemented using
role-playing.


no.

3 Heterogeneous Description
dimension Heterogeneous products can have different attributes, so it
is not possible to make a single product table. Let us
assume that an insurance company sells different kinds of
insurance, such as Car Insurance, Home Insurance, Flood
Insurance, and Life Insurance. Each type of insurance can
be treated as a product. However, we cannot create a
Single Product table to handle all these types of insurance
because each has extremely unique attributes.

Implementation
Here are ways to implement heterogeneous dimensions:
Merge all the attributes into a single product table and
all facts relating to the heterogeneous attributes in one
fact table.
Create separate dimensions and fact tables for the
heterogeneous products.
Create a generic design to include a single Fact and
Single Product Dimension table with common
attributes from two or more heterogeneous products.
The above mentioned concepts of implementing
heterogeneous dimensions are discussed in more detail in
5.3.12, “Heterogeneous products” on page 186.

The Sales Invoicing case study has two dimension groups,
beverages and non-beverages, that do not share all the
data in the Product dimension. For example, the fact table
contains measures only applicable to the beverages (such
as alcohol content volume and delivered liters). And, the
two main groups of products have specific attributes that
are applicable to one but not the other. We therefore
conclude that the product is a heterogeneous dimension.

However, we decided to leave both subtype attributes and
measures in the same dimension and facts table. Any
specific attributes since they are not needed by the
business. Product is a heterogeneous dimension but will
be left unchanged.


no.

4 Garbage Description
dimension A dimension that consists of low-cardinality columns, such
as codes, indicators, status, and flags, also referred to as
a junk dimension. The attributes in a garbage dimension
are not related to any hierarchy.

Implementation
Implementation involves separating the low-cardinality
attributes and creating a dimension for such attributes.
This implementation procedure is discussed in detail in
5.3.8, “Identifying garbage dimensions” on page 176.

For the case study, we do not have any situations involving
a garbage dimension.

5 Hot swappable Description
dimension A dimension that has multiple alternate versions of itself,
and can be swapped at query time, is called a hot
swappable dimension or profile table. Each of the versions
of the hot swappable dimension can be of a different
structure. The alternate versions access the same fact
table, but have different output. The versions of the
primary dimension may be completely different, including
incompatible attribute names and different hierarchies.

Implementation
The procedure to implement hot swappable dimensions is
discussed in detail in 5.3.13, “Hot swappable dimensions
or profile tables” on page 188.

For the case study about Sales Invoicing, we do not create
hot swappable dimensions. However, we could if we
decide to implement tighter security, or for performance
reasons.


7.5.11 Dimensional model containing final dimensions
Figure 7-27 shows the dimension model up to the point of having completed the
Identify dimensions phase of the DMDL.

Figure 7-27 Model after “identify dimensions” phase

7.6 Identify the facts
In the “Identify the facts” phase, shown in Figure 7-28 on page 410, the primary
goal is to provide more detail on the fact table.


Determine All
Dimensions
Processes

Identify Conformed



Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


We have identified preliminary dimensions and preliminary facts, for the sales
invoice business process, using the grain definition shown in Figure 7-29. In this
phase we will iteratively and in more detail, identify additional facts that are true
to this grain.

Customer
Time Customer Representative

Grain: 1) Price per item
2) Quantity
1 line item on an invoice
3) Discount%

Invoice Date Currency
Number

Figure 7-29 Grain definition for sales invoice business


Table 7-30 on page 411 shows the activities associated with the Identify the facts
phase.

Table 7-30 Activities in the Identify the Facts phase
S Activity name Activity description
no.

1 Identify facts Identifies the facts that are true to the grain identified
for our Sales Invoicing process. The grain was
identified in 7.4, “Identify the grain” on page 373.

2 Identify conformed facts Identify any conformed facts. That is, in this activity
we identify if any of these facts have been conformed
and shared across the organization. If yes, we use
those conformed facts.

3 Identify fact types Identify the fact types, such as:
- Additive
- Semi-additive
- Non-additive
- Derived
- Textual
- Pseudo
- Factless

4 Year-to-date facts Year-to-date facts are numeric totals that consist of
aggregated totals from start of year until the current
date. In this activity we discuss year-to-date facts to
make sure such facts are not included in a fact table
which includes data at the atomic line item level.

5 Event fact tables In this activity we describe how to handle events in
the event-based fact table, and highlight the pseudo
and factless facts that may be associated with such
tables.

6 Composite key design General guidelines for designing the primary
composite key of the fact table. We also describe
situations when a degenerate dimension may be
included inside the fact table composite primary key.

7 Fact table sizing and In this activity we describe guidelines to use to predict
growth fact table growth.

We now discuss each of the activities listed in Table 7-30.


7.6.1 Identify facts
Identifying and documenting the fact types is important for the applications and
users of the dimensional model. There are several facts, other than the
preliminary facts, that cannot be known by glancing at the grain definition. Such
facts are detailed and derived facts that need further analysis to be found.

In this activity we identify all the facts that are true to the grain. These facts
include the following:
The preliminary facts identified in “Identify high level dimensions and facts” on
page 378. Preliminary facts are easily identified just by looking at the grain
definition or the invoice.
Detailed facts, such as cost per individual product, labor manufacturing cost
per product, or transportation cost per individual product. These facts can
only be identified after a detailed analysis of the source E/R model to identify
all the facts that are true at the line item grain.

The sales fact table contains information about the customers that ordered the
products, the products ordered, the date on which the product is ordered, the
warehouse from which the products are shipped, the payment terms applied to
the ordered item, and the currency in which the ordered item is invoiced.

Table 7-31 shows the facts identified for the Sales Invoicing business process
that are true to the line item grain definition (see Figure 7-29 on page 410).

Table 7-31 Sales fact table
# Fact Description

1 Accrued An amount of money in the corporate currency indicating the
Bonus Amount bonus the customer gets for purchasing a product. Accrued
Bonus Amount= Quantity sold X Accrued Bonus per Item.

2 Alcohol Duty The amount in corporate currency to be paid for duty if the
Amount product is an alcoholic drink. [Alcohol Duty Amount= Quantity
sold X Alcohol Duty per Item].

3 Alcohol Alcohol by volume (ABV) is an indication of how much alcohol
Volume (expressed as a percentage) is included in an alcoholic
Percent beverage. This measurement is assumed to be a world
standard, although in the United States, the predominant
measurement is Alcohol by weight (also known as ABW).
Another way of specifying the amount of alcohol is alcoholic
proof.

4 Cost Of Goods Also known as cost of product in corporate currency
excluding the taxes.
[Cost of Goods = Quantity sold x Cost of good per item].


# Fact Description

5 Delivered Amount of liters shipped with a product.
Liters [Delivered Liters = Quantity sold x Delivered Liters per Item].

6 Quantity Sold The quantity of the product that has been already sold.
Note: This was a preliminary fact we identified in 7.4.5,
“Identify high level dimensions and facts” on page 378.

7 Discount The amount of money, in corporate currency, the customer
Amount gets as a discount.
Note: This was a preliminary fact we identified in 7.4.5,
“Identify high level dimensions and facts” on page 378.
[Discount Amount = Quantity sold x Discount Amount per
item].

8 Gross Sales The amount of money in corporate currency already invoiced.
Amount Note: We identified Price per item as preliminary fact in
7.4.5, “Identify high level dimensions and facts” on page 378.
Instead of storing ‘Price per item’ as a fact, we store the
Gross sales cost for the line item as Quantity sold x Price per
item.
[Gross Sales Amount = Quantity sold x Price per item].

9 Lc Bonus The amount of money, in the ordered local currency,
Amount indicating the bonus the customer gets for purchasing the
product. [Local Bonus Amount = Quantity sold x Bonus
Amount per item].

10 Lc Cost Of The net amount of money in the ordered local currency the
Goods product costs. [Cost of Goods = Quantity sold x Cost of good
per item].

11 Lc Discount The amount of money in the ordered currency the customer
Amount gets discounted on the product. [Discount Amount = Quantity
sold x Discount Amount per item].

12 Lc Duty The amount in ordered local currency to be paid for duty if the
Amount product is an alcoholic drink. [Alcohol Duty Amount =
Quantity sold X Alcohol Duty per Item].

13 Lc Gross Sales The amount of money in ordered local currency already
Amount invoiced for this product. [Gross Sales Amount = Quantity
sold x Price per item].

14 Lc Order The amount of money in the order currency for order
Handling Cost handling. This information is not kept at the order level
because certain kinds of liquors are very expensive luxury
articles that need to be treated very carefully and be specially
packaged. [Order Handling Cost = Quantity sold x Order
Handling Cost per Item].


# Fact Description

15 Order The amount of money in the corporate currency for order
Handling Cost handling. This information is not kept at the order level
because certain kinds of liquors are very expensive luxury
articles that need to be treated very carefully and be specially
packaged. [Order Handling Cost = Quantity sold x Order
Handling Cost per Item].

16 Spirit Liters Amount of alcohol in liters per item =
Delivered Liters * (Alcohol Volume Pct / 100).

17 Gross profit Profit calculated as gross sales income less the cost of sales.

18 Gross margin Percentage difference between the cost of sales and the
Percent gross sales.

19 Net sales The amount of the sale. This is a negative amount, if it is a
Amount credit note, due to returned or damaged material.

20 Net profit Net Sales - Cost of Sales - Alcohol Duty Amount
Amount

How to identify facts or fact tables from an E/R model
The following are the steps to convert an E/R model into a dimensional model:
1. Identify the business process from the E/R Model.
2. Identify many-to-many tables in E/R model to convert to fact tables.
4. Identify date and time from E/R model.

Because of the huge E/R model of the Redbook Vineyard company data
warehouse, we do not cover all the steps in detail. For more step-by-step details
about how to determine fact tables and facts in the E/R model, refer to 5.1,
“Converting an E/R model to a dimensional model” on page 104.

7.6.2 Conformed facts
For the Sales Invoicing case study, we do not have any conformed facts.

7.6.3 Identify fact types (additivity and derived types)
In this activity we categorize facts based on the way they can be added. This
categorization is especially useful in order to avoid mistakes when adding up
columns of data that cannot be absolutely or partially summed.


The important factors we consider when aggregating data by column are:
additivity, semi-additivity, and non-additivity. For a detailed definition of fact types,
refer to 6.5.1, “Facts” on page 295.

In Table 7-32 we describe all the facts by:
Categorizing them as additive, non-additive, or semi-additive. Handling
semi-additive and non-additive facts is considered to be an advanced
concept. For more on handling such facts, refer to the following:
– “Non-additive facts” on page 191
– “Semi-additive facts” on page 193
Identifying them as derived, when they can be calculated from other facts that
exist in the table, or that have been also derived.

Table 7-32 Facts identified for Sales fact table
S# Fact Description Type of fact Derived

1 Accrued An amount of money in the Additive No
Bonus corporate currency indicating [It would have
Amount the bonus the customer gets been a derived
for purchasing a product. This fact if Accrued
fact is true to the line item Bonus per Item
grain. [Accrued Bonus Amount had also been
= Quantity sold x Accrued stored in the
Bonus per Item] fact table].

2 Alcohol The amount in corporate Additive No
Duty currency to be paid for duty if [It would have
Amount the product is an alcoholic been a derived
drink. This fact is true to the fact if [Alcohol
grain. [Alcohol Duty Amount= Duty per Item]
Quantity sold x Alcohol Duty had also been
per Item] stored in the
fact table].

3 Alcohol Alcohol by volume (ABV) is an Non-additive No
Volume indication of how much alcohol
Percent (expressed as a percentage) is
included in an alcoholic
beverage. This measurement
is assumed as a world
standard, although in the
United States, the
predominant measurement is
Alcohol by weight (also known
as ABW). Another way of
specifying the amount of
alcohol is alcoholic proof.



4 Cost Of Also known as cost of product Additive No
Goods in corporate currency [It would have
excluding the taxes. been a derived
[Cost of Goods = Quantity sold fact if [Cost of
x Cost of good per item] good per item]
had also been
stored in the
fact table].

5 Delivered Amount of liters shipped for a Additive No
Liters product. [It would have
[Delivered Liters = Quantity been a derived
sold x Delivered Liters per fact if delivered
Item] liters per item
had also been
stored in the
fact table].

6 Quantity The quantity of product in units Additive No
Sold that has been already sold.
Note: This was a preliminary
fact we identified in 7.4.5,
“Identify high level dimensions
and facts” on page 378.

7 Discount The amount of money in Additive No.
Amount corporate currency the It would have
customer gets discounted on been a derived
the product. fact if [Discount
Note: This was a preliminary Amount per
fact we identified in 7.4.5, item per item]
“Identify high level dimensions had also been
and facts” on page 378. stored in the
[Discount Amount = Quantity fact table.
sold x Discount Amount per
item]



8 Gross The amount of money in Additive No
Sales corporate currency already [It would have
Amount invoiced. been a derived
Note: We identified Price per fact if Price per
item as preliminary fact in item had also
7.4.5, “Identify high level been stored in
dimensions and facts” on the fact table].
page 378. Instead of storing
‘Price per item’ as a fact, we
store the Gross sales cost for
the line item as Quantity sold x
Price per item. [Gross Sales
Amount = Quantity sold x Price
per item]

9 Lc Bonus The amount of money in the Semi-additive No
Amount ordered local currency (You cannot [It would have
indicating the bonus the add different been a derived
customer gets for purchasing currencies.) fact if Bonus
the product. Amount per
[Local Bonus Amount = item had also
Quantity sold x Bonus Amount been stored in
per item] the fact table].

10 Lc Cost The net amount of money in Semi-additive No
Of Goods the ordered local currency the (You cannot [It would have
product costs. add different been a derived
[Cost of Goods = Quantity sold currencies) fact if Cost of
x Cost of good per item] good per item
had also been
stored in the
fact table].

11 Lc The amount of money in the Semi-additive No
Discount ordered currency the customer (You cannot [It would have
Amount gets as a discount on the add different been a derived
product. currencies) fact if Discount
[Discount Amount = Quantity amount per
sold x Discount Amount per item had also
item] been stored in
the fact table].



12 Lc Duty The amount in ordered local Semi-additive No
Amount currency to be paid for duty if (You cannot [It would have
the product is an alcoholic add different been a derived
drink. currencies) fact if Alcohol
[Alcohol Duty Amount = duty per item
Quantity sold x Alcohol Duty had also been
per Item] stored in the
fact table].

13 Lc Gross The amount of money in Semi-additive No
Sales ordered local currency already (You cannot [It would have
Amount invoiced for this product. add different been a derived
Gross Sales Amount = currencies) fact if Price per
Quantity sold x Price per item. item had also
been stored in
the fact table].

14 Lc Order The amount of money in the Semi-additive No
Handling order currency that the order (You cannot [It would have
Cost handling of the product costs. add different been a derived
This information is not kept at currencies) fact if Order
the order level because certain handling cost
kinds of liquors are very per item had
expensive luxury articles that also been
need to be treated very stored in the
carefully and be specially fact table].
packaged.
[Order Handling Cost =
Quantity sold x Order Handling
Cost per Item]

15 Order The amount of money in the Additive No
Handling corporate currency that the It would have
Cost order handling of the product been a derived
costs. This information is not fact if Order
kept at the order level because handling cost
certain kinds of liquors are per item had
very expensive luxury articles also been
that need to be treated very stored in the
carefully and be specially fact table.
packaged.
[Order Handling Cost =
Quantity sold x Order Handling
Cost per Item]
(in Corporate currency)



16 Spirit Amount of alcohol in liters per Additive Yes
Liters item = (See
Delivered Liters * (Alcohol description for
Volume Pct / 100) formula)

17 Gross Profit calculated as gross Additive Yes
profit sales income less the cost of (See
sales. description for
formula)

18 Gross Percentage difference Non-additive Yes
margin between the cost of sales and (See
Percent the gross sales. description for
formula)

19 Net sales The amount of the sale. This is Additive Yes
Amount a negative amount if it is a (See
credit note due to returned or description for
damaged material. formula)
Net sales = Gross Sales minus
Amount returned and
damaged items. 7.3.6, “Gather
the requirements” on
page 362.
When the invoice is created for
a customer, the <Amount
returned and damaged items>
is 0. However, if the customer
returns goods which are
damaged, a new invoice form
is created and sent to the
customer with <Amount
returned and damaged items>
showing a negative Amount.

20 Net profit Net Sales - Cost of Sales - Additive Yes
Amount Alcohol Duty Amount (See
(in Corporate currency) description for
formula)

Note: A fact is said to be derived if the fact can be calculated from other facts
that exist in the table, or that have also been derived. You may decide not to
include the derived facts inside the fact table and to calculate them inside the
front-end reporting application.


We have identified two year-to-date facts: Gross Sales YTD and Gross Margin
YTD. These YTD facts are shown in Table 7-33.

Table 7-33 Year-to-Date Facts
S.No Name of Year-to Date Fact Description

1 Gross Sales YTD Gross Sales during the present year

2 Gross Margin YTD Gross Margin Average during the
present year

However these year to date facts are not true to this grain. Our two year-to-date
facts will not be physically implemented and will be handled out of the
dimensional model by means of one of the following of:
An OLAP application
RDMS user-defined functions
An RDMS view
A user SQL query

For more information about Year-to-Date facts, refer to 6.5.4, “Year-to-date facts”
on page 300.

7.6.5 Event facts, composite keys, and growth
For our Sales Invoicing case study, we do not have any event-based fact tables.
For more information about event-based fact tables, refer to 5.4.4, “Handling
event-based fact tables” on page 205.

The composite key of the sales fact table consists of the following list of keys, all
of which are surrogate keys.
Currency
Customer
Invoice
Payment Terms
Product
Sales Date
Sales Representative
Warehouse

In this activity, we estimate the fact table size and also predict its growth. It is
important to understand that approximately 85-95% of the space of a
dimensional model is occupied by the fact tables. Therefore, it is extremely
important to understand the future growth pattern to plan for future fact table


performance. For more information about how to predict fact table growth, refer
to 6.5.7, “Fact table sizing and growth” on page 303.

7.6.6 Phase Summary
We now have the definitive version of the logical dimensional model. The final
dimensional model is shown in Figure 7-30.

Figure 7-30 Final model for Sales Invoicing business process

7.7 Other phases
We developed a dimensional model for the Sales Invoicing business process, in
this case study. We discussed the following phases of the DMDL:
Identify the business process
Identify the grain
Identify the facts


We now briefly describe how to apply the other phases of the DMDL for the Sales
Invoicing case study. The other phases of the DMDL we consider are:
Verify the model: Here we verify the dimensional model (Figure 7-30 on
page 421) we made for our sales invoice process. We verify whether the
dimensional model designed is able to answer the business requirements
identified during the requirements gathering phase in “Gather the
requirements” on page 362. For more information about the ‘Verify’ phase,
refer to “Verify the model” on page 305.
Physical Design Considerations: In this phase we do the following:
– Identify Aggregates: In simple terms, aggregation is the process of
calculating summary data from detail base level fact table records.
Aggregates are the most powerful tool for increasing query processing
performance in dimensional data marts. The main goal of this phase is to
identify the aggregates that will help improve performance. For more
information about creating aggregates, refer to “Aggregations” on
page 308.
– Aggregate Navigation: This is a technique that involves redirecting user
SQL queries to appropriate precomputed aggregates. In this phase we
design intelligent aggregate navigation strategies. For more information
about this advanced topic, refer to “Aggregate navigation” on page 312.
– Indexing: In this phase we create indexes for the dimensions and fact
table. For more information about creating indexes, refer to:
• “Indexing” on page 314 provides guidelines about creating indexes.
• “Indexing for dimension and fact tables” on page 218 (A sample star
schema to show indexing).
– Partitions: Partitioning a table divides the table by row, by column, or both.
If a table is divided by column, it is said to be vertically partitioned. If a
table is divided by rows, it is said to be horizontally partitioned. Partitioning
large fact tables improves performance because each partition is more
manageable and smaller to enable better performance. In this phase we
create partitions. For more guidelines about creating partitions, refer to
“Partitioning” on page 319.
Meta Data management: In this phase we manage meta data for the
following phases of the DMDL:
– Identify business process: The output of this phase results in the
Requirements gathering report. The report primarily consists of the
business requirements for the selected business for which you will design
the dimensional model. In addition, the report also consists of the
business processes, owners, source systems involved, data quality
issues, common terms used across business processes, and other
business-related meta data.


– Identify the grain: The output of this phase results in the Grain definition
report (See DMDL Figure 7-28 on page 410). The Grain definition report
consists of one or multiple definitions of the grain for the business process
for which the dimensional model is being designed. Also, the type of fact
table (transaction, periodic, or accumulating) being used is mentioned.
The Grain definition report also includes high level preliminary dimensions
and facts.
– Identify the dimensions: The meta data for this phase includes the
dimension name, its hierarchies, update rules, load frequency, load
statistics, usage statistics, archive rules, archive statistics, purge rules,
purge statistics, attributes, and so on. For more detail, refer to “Meta data
management” on page 320.
– Identify the facts: The meta data for this phase includes detailed
information about the facts. For more detail, refer to “Meta data
management” on page 320.
– User Verification phase: This phase includes meta data gathered after the
user verification of the dimensional model.
– Physical design consideration: The meta data for this phase includes
aggregate tables involved, aggregate navigation strategy, index names,
index descriptions, and partitions involved.
Design the next priority business process data mart: In this activity, we
designed the dimensional model for the Sales Invoicing business process of
the Redbook Vineyard company. However, the Redbook Vineyard company
has several business processes that we identified and prioritized in the
“Identify business process” on page 355. The various business processes are
shown in Table 7-34. We identified the Sales Invoicing as the process with the
highest priority for which the dimensional model should be designed.
The next step is to select the next high priority business process and use the
DMDL to design another mart. Most likely, management will agree on either
Inventory or Sales Orders. This is because these have the highest points as
shown in Table 7-34. For more information on prioritizing business processes,
refer to 7.3.2, “Identify business process” on page 355.

Table 7-34 Business process prioritization
Name of Complexity Data quality System Strategic Final Points
Business Availability significance
Process

Distribution 6 6 1 2 15

Sales CRM/ 4 6 2 4 16
Marketing


Name of Complexity Data quality System Strategic Final Points
Business Availability significance
Process

Inventory 4 9 3 4 20

Procurement 4 6 3 2 15

Sales Orders 4 9 3 4 20

Accounting 4 9 2 2 17

Production 2 3 3 6 14

Sales 6 9 3 4 22
Invoicing

7.8 Conclusion
In this chapter, we designed a dimensional model for the Sales Invoicing
process. The objective was to put to test the DMDL developed in Chapter 6,
“Dimensional Model Design Life Cycle” on page 227.

We exercised the DMDL by executing the following phases, and completing a
dimensional model of the sales invoice process:
Identify the business process
Identify the grain
Identify the facts
Verify
Physical design considerations
Meta data management

The important point to understand is that each phase of the DMDL consists of
several activities. However, depending upon the business situation and the
design need, you may use all, or only some of the activities inside each of the
phases shown in the DMDL.


8

Chapter 8. Case Study: Analyzing a
dimensional model
In this chapter we review an existing design of a dimensional model and provide
guidelines to improve the model. So, the flow of this chapter is different because
we typically describe how to design a new model from the beginning.

In short, in this chapter the discussions focus on the following topics:
Study the business of a fictitious company.
Understand the business needs for creating a dimensional model.
Review the draft dimensional model.
Discuss guidelines for reviewing the existing dimensional model.
Make improvements to the draft dimensional model, and discuss the reasons
for doing so.


8.1 Case Study - Sherpa and Sid Corporation
In this section, we briefly describe the fictitious Sherpa and Sid Corporation and
its business. In doing so, we describe the reason that the corporation has
submitted a request to build a data mart.

8.1.1 About the company
Sherpa and Sid Corporation started as a manufacturer of cellular telephones,
and then quickly expanded to include a broad range of telecommunication
products. As the demand for, and size of, its suite of products grew, Sherpa and
Sid Corporation closed down the existing distribution channels and opened its
own sales outlets.

In the past year Sherpa and Sid Corporation opened new plants, sales offices,
and stores in response to increasing customer demand. With its focus firmly on
expansion, the corporation put little effort into measuring the effectiveness of the
expansion. Sherpa and Sid’s growth has started to level off, and management is
refocusing on the performance of the organization. However, although cost and
revenue figures are available for the company as a whole, little data is available
at the manufacturing plant or sales outlet level regarding cost, revenue, and the
relationship between them.

To rectify this situation, management has requested a series of reports from the
Information Technology (IT) department. IT responded with a proposal to
implement a data mart. After consideration of the potential costs and benefits,
management agreed.

8.1.2 Project definition
Senior management and IT put together a project definition consisting of the
following objective and scope:

Project objective
To create a data mart to facilitate the analysis of cost and revenue data for
products manufactured and sold by Sherpa and Sid Corporation.

Project scope
The project will be limited to direct costs and revenues associated with products.
Currently, Sherpa and Sid Corporation manufacturing costs cannot be allocated
at the product level. Therefore, only component costs can be included. At a future
time, rules for allocation of manufacturing and overhead costs may be created,
so the data mart should be flexible enough to accommodate future changes. IT


created a team consisting of one data analyst, one process analyst, the
manufacturing plant manager, and one sales region manager for the project.

8.2 Business needs review
In this section, we review the requirements gathering meta data. In other words,
we review what the project team defined, and felt they needed to investigate, in
order to understand the business need for the Sherpa and Sid Corporation. The
understanding of the business will greatly help in the review of the data mart.

The team identified the following areas of interest:
Life Cycle of a product
Anatomy of a sale
Structure of the organization
Defining cost and revenue
What do the users want?

8.2.1 Life cycle of a product
The project team first studied the life cycle of a product. Each manufacturing
plant has a research group that tests new product ideas. Only after the
manufacturing process has been completely defined and approval for the new
product has been obtained is the product information added to the company's
records. Once the product information is complete, all manufacturing plants can
produce it.

A product has a base set of common components. Additional components can
be added to the base set to create specific models of the product. Currently,
Sherpa and Sid Corporation has 300 models of their products. This number is
fairly constant as the rate of new models being created approximately equals the
rate of old models being discontinued. Approximately 10 models per week
experience a cost or price change. For each model of each product, a decision is
made about whether or not it is eligible for discounting. When a model is deemed
eligible for discounting, the salesperson may discount the price if the customer
buys a large quantity of the model or a combination of models. In a retail store,
the store manager must approve such a discount.

The plant keeps an inventory of product models. When the quantity on hand for a
model falls below a predetermined level, a work order is created to cause more of
the model to be manufactured. Once a model is manufactured, it is stored at the
manufacturing plant until it is requested by a sales outlet.

Chapter 8. Case Study: Analyzing a dimensional model 427

The sales outlet is responsible for selling the model. When a decision is made to
stop making a model, data about the model is kept on file for six months after the
last unit of the model has been sold or discarded. Data about a product is
removed at the same time as data about the last model for the product is
removed.

8.2.2 Anatomy of a sale
There are two types of sales outlets: corporate sales office and retail store. A
corporate sales office sells only to corporate customers. Corporate customers
are charged the suggested wholesale price for a model unless a discount is
negotiated. One of Sherpa and Sid Corporation's 30 sales representatives is
assigned to each corporate customer. Sherpa and Sid Corporation currently
serves 3000 corporate customers. A customer can place orders through a
representative or by phoning an order desk at a corporate sales office. Orders
placed through a corporate sales office are shipped directly from the plant to the
customer. A customer can have many shipping locations. So, it is possible for a
customer to place orders from multiple sales offices if the policy is to let each
location do their own ordering.

The corporate sales office places the order with the plant closest to the customer
shipping location. If a customer places an order for multiple locations, the
corporate sales office splits it into an individual order for each location. A
corporate sales office, on average, creates 500 orders per day, five days per
week. Each order consists of an average of 10 product models.

A retail store sells over the counter. Unless a discount is negotiated, the
suggested retail price is charged. Although each product sale is recorded on an
order, the company does not keep records of customer information for retail
sales. A store can only order from one manufacturing plant. The store manager is
responsible for deciding which products to stock and sell from their particular
store. A retail store, on average, creates 1000 orders per day, seven days per
week. Each order consists of an average of two product models.

8.2.3 Structure of the organization
It was clear to the team that understanding products and sales was not enough;
an understanding of the organization was also necessary. The regional sales
manager provided an up-to-date copy of the organization structure, which is
depicted in Figure 8-1 on page 429.


Sherpa & Sid
Corporate Office

Manufacturing Sales

Western Region Eastern Region
North East South East North West South West
3 Plants 4 Plant

Corporate Sales Retail Sales Retail Sales Corporate Sales
2 Offices 3 Stores 2 Stores 1 Office

Retail Sales Retail Sales
4 Stores 6 Stores

Figure 8-1 Sherpa and Sid Corporation organizational structure

8.2.4 Defining cost and revenue
The project team studied the cost and revenue components from a business
standpoint. Their goal was to be able to define cost and revenue, in order to be
able to effectively analyze those factors.

For each product model, the cost of each component is multiplied by the number
of components used to manufacture the model. The sum of results for all
components that make up the model is the cost of that model.

For each product model, the negotiated unit selling price is multiplied by the
quantity sold. The sum of results for all order lines that sell the model is the
revenue for that model.

When trying to relate the cost of a model to its revenue, the team discovered that
once a model was manufactured and added to the quantity on hand in inventory,
the cost of that unit of the model could not be definitively identified. Even though
the cost of a component is kept, it is only used to calculate a current value of the
inventory. Actual cost is recorded only in the company's financial system, with no
reference to the quantity manufactured.

The results of this determination were two-fold. First, the team requested that the
operational systems be changed to start recording the actual cost of a
manufactured model. However, both management and the project team
recognized that this was a significant change, and that waiting for it would
severely impact the progress of the project. Therefore, and based on the fact that
component costs changed infrequently and by small amounts, the team defined
this rule:
The revenue from the sale of a model is always recorded with the current unit
cost of the model, regardless of the cost of the model at the time it was
manufactured.


8.2.5 What do the users want?
Because the objective of the project was to create a collection of data that users
could effectively analyze, the project team decided to identify a set of typical
questions users wanted the data to answer. Clearly, this would not be an
exhaustive list. The answer to one question would certainly determine what the
next question, if any, might be. As well, one purpose of the data mart is to allow
the asking of as yet unknown questions. If users simply want to answer a rigid
set of questions, creating a set of reports would likely fill the need. With this in
mind, the team defined a set of questions, and they are shown in Table 8-1.

Table 8-1 Business questions
S.no Business question

1 What are the total cost and revenue for each model sold today, summarized
by outlet, outlet type, region, and corporate sales levels?

2 What are the total cost and revenue for each model sold today, summarized
by manufacturing plant and region?

3 What percentage of models are eligible for discounting, and of those, what
percentage are actually discounted when sold, by store, for all sales this
week? This month?

4 For each model sold this month, what is the percentage sold retail, the
percentage sold corporately through an order desk, and the percentage sold
corporately by a salesperson?

5 Which models and products have not sold in the last week? The last month?

6 What are the top five models sold last month by total revenue? By quantity
sold? By total cost?

7 Which sales outlets had no sales recorded last month for the models in the top
five models list?

8 Which salespersons had no sales recorded last month for the models in the
top five models list?

As well as being able to answer the above questions, the users want to be able to
review up to three complete years of data to analyze how the answers to these
questions change over time.


8.2.6 Draft dimensional model
Based on the business functions and the requirement gathering, the team came
to the following conclusion:
The Sherpa and Sid corporation is tracking the sales of its products (made in
different manufacturing plants) to different customers.
The Sherpa and Sid corporation is basically comprised of two broad
operations:
– Manufacturing products in its manufacturing plants
– Sales of these products by its sales outlets to customers
The customers of Sherpa and Sid corporation are either big corporate
companies or retailers who buy directly over the counter.
Each customer purchases one or more products through an order.
There are two types of seller outlets:
– Corporate sales office
– Retail stores
The products can be bought in the following two ways:
– In the case of retail (non-corporate) customers, the products are
purchased over the counter from retail outlets.
– In the case of corporate customers, orders can be placed over the phone
and goods are delivered directly from plant to the particular corporate
office.

Based on the above conclusions, the data mart team chose the grain for the
dimensional model to be a single line item on an order placed. The draft
dimensional model that the team built is shown in Figure 8-2 on page 432.


Figure 8-2 Draft dimensional model before review

8.3 Dimensional model review guidelines
In this phase we review the design of the dimensional model shown in Figure 8-2.
We perform this review by using the concepts presented in the Dimensional
Model Design Life Cycle. See Chapter 6, “Dimensional Model Design Life Cycle”
on page 227. The life cycle is divided into different phases. Each phase explains
different concepts that help you in the design of the dimensional model.


Determine All
Dimensions
Processes

Identify Conformed



Granularity
Accumulating)
(Source Driven

Iterate
Metadata Management


In this case study, instead of designing a new model, we review an existing one.
Below we provide general guidelines to consider while reviewing the design of a
dimensional model.

8.3.1 What is the grain?
Grain specifies what level of detailed information your star schema will have. The
grain is the most important part of the dimensional model. It is the grain definition
that sets the stage for development of the entire model. An incorrect grain may
lead to the redesign of the entire model. It is very important that enough time be
spent on analyzing the grain definition.

Review result for draft model: For the dimensional model shown in Figure 8-2
on page 432, the grain definition is right and appropriate considering the need of
the business to track the sales of the different product models on a daily basis
(see the question in Table 8-1 on page 430).


8.3.2 Are there multiple granularities involved?
There will be times when more than one grain definition will be associated with a
single business process. In these scenarios, we recommend you design
separate fact tables with separate grains and not forcefully try to put facts that
belong to separate grains in a single fact table.

You can handle differing data granularities by using multiple fact tables (daily,
monthly, and yearly tables). Also consider the amounts of data, space, and the
performance requirements to decide how to handle different granularities. The
factors that can help you to decide whether to design one or more fact tables are
discussed in more detail in 6.3.2, “Multiple, separate grains” on page 249.

on page 432, there are no separate fact granularities that are involved so there is
only one fact table needed. The business requirements (questions 1 to 8) shown
in Table 8-1 on page 430 show that only sales of products needs to be tracked for
different products. Had the business required knowing the inventory of all sold
products in different outlets, we would have had to design a separate fact table to
handle inventory in each outlet on a daily basis.

Ideally we would like to design the dimensional model at the most atomic level.
The dimensional model has been designed at the most atomic level if it cannot
be further divided into detailed information. It is important for the dimensional
modeling team not to be shortsighted. Suppose, for example, for now that the
data needs to be summarized at the monthly level. The dimensional modeling
team may, based on the current business requirements, decide to design the
dimensional model with a monthly grain. But what if in future the business asks to
see the data at the weekly level? The sooner you ask these types of questions,
the better. In other words, try to foresee the business needs. In this case, if it
seems likely that they may need to see weekly data in the future, then design the
model at the most detailed level. This way you not only give the business what
they need (monthly data) right now, but are also well positioned to proactively
help them in future with daily data.

For more discussion on the importance of having a detailed atomic grain, you
can refer to the following:
“Check grain atomicity” on page 252.
“Importance of detailed atomic grain” on page 122

Review results for draft model: For the dimensional model shown in Figure 8-2
on page 432, the grain is a single line item on an order. This is the most detailed
atomic definition, so we do not need to go further.


8.3.4 Review granularity for date and time dimension
All dimensional models will surely consist of a date dimension. Some may have a
time dimension too. Date and time dimensions play a very important role in
identifying the level of detail of information that is going to be available in your
model. Guidelines for handling date and time in your dimensional models are as
follows:
Date should be handled separately as its own dimension. We discussed the
importance of handling date separately as a dimension in “Date and time
granularity” on page 279.
Date and time dimensions should not be merged together.
If there are multiple dates involved in the dimensional model, then instead of
creating separate physical tables for these dates, use the concept of
role-playing to implement several dates as views from one main date table.
The concept of role-playing is discussed in “Role-playing dimensions” on
page 179.
Time can be handled in two ways:
– As a fact inside the fact table
– As its own dimension
We discussed this in more detail in 5.3.2, “Handling time as a dimension or a
fact” on page 139.
We may handle date and time across international time zones by storing both
the local date/time and the international GMT date/time. This is discussed in
“Handling date and time across international time zones” on page 142.

on page 432, there are two date dimensions involved. Currently, the order date
and delivery date are both incorrectly stored in an order dimension table. The
order date and delivery date should be separated into individual order date and
delivery date dimensions.

8.3.5 Are there degenerate dimensions?
A degenerate dimension is a dimension without attributes. Degenerate
dimensions are typically numbers such as a booking number, order number,
confirmation number, receipt number, and ticket number.

How to identify and handle a degenerate dimension
A degenerate dimension exists in the dimensional design if any of the following is
true:


The dimensional design consists of a dimension table which has an equal
number of rows as the fact table. In other words, if while loading your fact
table with X rows, you need to pre-insert the same number of X rows in any
dimension table, then this is an indication that there is a degenerate
dimension table in your design.
If any of your dimension tables have a nearly equal (but not equal) number of
rows as compared to the fact table, then in this case there is a possibility of a
degenerate dimension.

The concept of identifying and handling a degenerate dimension is discussed in
more detail in the following sections:
– “Degenerate dimensions” on page 266.
– “Degenerate dimensions” on page 134.

on page 432, the order table is actually a degenerate dimension. The order
dimension table is storing the order number, order date, and delivery date. We
discussed in 8.3.4, “Review granularity for date and time dimension” on page 435
that order and delivery dates belong as separate dimensions and not in the order
table. In addition to this, all order-related information, such as products ordered
and from which outlet, is stored in separate dimensions such as product,
manufacturing plant, and seller.

The order dimension therefore is degenerate and the only information remaining
is the Order_Number, which can be stored inside the sales_fact table.

8.3.6 Surrogate keys
Primary keys of dimension tables should remain stable. We strongly recommend
that you create surrogate keys and use them for primary keys for all dimension
tables. In this section we discuss surrogate keys, and why it is important to use
the surrogate keys as the dimension table primary keys.

What are Surrogate Keys?

Surrogate keys are keys that are maintained within the data warehouse instead
of the natural keys taken from source data systems. Surrogate keys are known
by many other aliases, such as dummy keys, non-natural keys, artificial keys,
meaningless keys, non-intelligent keys, integer keys, number keys, technical
integer keys, and so on. The surrogate keys basically serve to join the dimension
tables to the fact table. Surrogate keys serve as an important means of
identifying each instance or entity inside a dimension table.

For a detailed discussion about surrogate keys and their importance, refer to
“Reasons for using surrogate keys are:” on page 263.


on page 432, we observed that all the dimension tables are using the natural
OLTP id instead of using a surrogate key. It is important that all dimension
primary keys strictly be surrogate keys only.

8.3.7 Conformed dimensions and facts
A conformed dimension means the same thing to each fact table to which it can
be joined. A more precise definition is that two dimensions are conformed if they
share one, more than one, or all attributes that are drawn from the same domain.
In other words, a dimension may be conformed even if it contains only a subset
of attributes from the primary dimension.

Typically, dimension tables that are referenced, or are likely to be referenced by
multiple fact tables (multiple dimensional models) are called conformed
dimensions.

Identify whether conformed dimensions or fact exist
If conformed dimensions already exist for any of the dimensions in your data
warehouse or dimensional model, you will be expected to use the conformed
dimension versions. If you are developing new dimensions with potential for
usage across the entire enterprise warehouse, you will be expected to develop a
design that supports anticipated enterprise data warehouse needs. In order to
determine the anticipated warehouse needs, you might need to interact with
several business processes to find out how they would define the dimensions.
The same is true for conformed facts.

on page 432, we observed that this is the first effort toward building a data
warehouse for the Sherpa and Sid Corporation. Therefore, since this is the first
dimensional model of its kind, there are no previous shared or conformed
dimensions or facts we can use for this data mart.

8.3.8 Dimension granularity and quality
Dimension granularity specifies the level of detail stored in a dimension. The level
of detail of all dimensions helps to determine the overall level of information that
is available in the dimensional model. Guidelines for reviewing dimension
granularity are:
Every dimension should be reviewed to see if all attributes are true to the
grain.
Every dimensional attribute should take only one value for a single row inside
the fact table. If a dimensional attribute has two or more values for a single
fact table row, then you must treat this dimension as a multi-valued


dimension. The concept of handling multi-valued dimensions is explained in
more detail in the following sections:
– “Other dimensional challenges” on page 290
– “Multi-valued dimensions” on page 182

Important points to consider when choosing attributes for dimensions are
explained as follows:
Non-key columns are generally referred to as attributes. Every dimension
table primary key should be a surrogate key which is not considered an
attribute. Because the business users will not use this surrogate key for
analysis (because it is simply an integer with no associated information).
Use a separate field in the dimension table to preserve the natural source
system key of the entity that is used in the source system.
A schema design that contains complete, consistent, and accurate attribute
fields helps users write queries that they intuitively understand and reduces
the support burden on the organization (usually the IT department)
responsible for database management and reports.
A well-designed schema includes attributes that reflect the potential areas of
interest and attributes that you can use for aggregations as well as for
selective constraints and report breaks.
If a dimension table includes a code, in most cases, include the code
description as well. As an example, if branch locations are identified by a
branch code, and each code represents a branch name, include both the
code and the name. Avoid storing cryptic decodes inside a dimensional table
attribute to save space.
Make sure the attribute names are unique within your model. If you have
duplicate names for different attributes, use the prime term (entity name) to
create a distinction. For example, if you have multiple attributes called
Address Type Code, one might be renamed Beneficiary Address Type Code,
and another might be Premium Address Type Code.
The dimension attributes serve as report labels and should be descriptive and
easy to understand. For example, in a given situation where you want to store
a flag, such as 0/1 or Y/N, it is better to store something descriptive such as
Yes/No instead.
An attribute can be defined to permit missing values in cases where an
attribute does not apply to a specific item, or its value is unknown.
An attribute may belong to more than one hierarchy.
Use only the alphabetic characters A-Z and the space character. Do not use
punctuation marks or special characters, including the slash (/) or the hyphen
(-).


While naming your attributes, do not use possessive nouns, for example, use
“Recipient Birth Date" rather than "Recipient's Birth Date."
Do not reflect permitted values in the attribute name. For example, the
attribute name "Employee Day/Night Code" refers to code values designating
day shift or night shift employees. Name the attribute to reflect the logical
purpose and the entire range of values. For example, "Employee Shift Type
Code," which allows for an expandable set of valid values.
Do not include very large names for building your attributes.
Properly document all dimensional attributes.

Note: It is important to remember that 85-90% of space in large dimensional
models is occupied by the fact table. Therefore, saving space by using codes
in dimensional attributes does not save much overall space, but these codes
do affect the overall quality and understandability of the dimensional model.

on page 432, the following improvements need to be made:
The Customer dimension stores address in a single column. In this way we
are unable to determine the detail about addresses such as city, district,
state, and region. We will redesign the address column to include more
address details, such as street name, suite, city, district, state, and region.
The draft dimensional model shown in Figure 8-2 on page 432 consists of
codes in mostly every dimension. We need to include the detailed description
for these codes also.
All the dimension attributes have a single value for every fact table row. In
other words, we do not need to use multi-value dimensions.

8.3.9 Dimension hierarchies
basically consists of different levels. Each level in a hierarchy corresponds to a
dimension attribute.


There are three major types of hierarchies that you should look for in each of the
dimensions. They are explained in Table 8-2 on page 440.


S.No Name Hierarchy description How implemented

1 Balanced A balanced hierarchy is a See “How to implement a
hierarchy in which all the balanced hierarchy” on
dimension branches have the page 144.
same number of levels. For more
details, see “Balanced hierarchy”
on page 143.

2 Unbalanced A hierarchy is unbalanced if it has See “How to implement
dimension branches containing an unbalanced
varying numbers of levels. hierarchy” on page 146.
Parent-child dimensions support
unbalanced hierarchies. For more
details, see “Unbalanced

least one member whose parent ragged hierarchy in
belongs to a hierarchy that is dimensions” on
more than one level above the page 155.
child. Ragged dimensions,
therefore, contain branches with
varying depths. For more details,
see “Ragged hierarchy” on
page 154.

Note: A dimension table may consist of multiple hierarchies. Also, a
dimension table may consist of attributes or columns which belong to one,
more, or no hierarchies.

on page 432, we observe that the product hierarchy (Product Category →
Product Brand → Product Name) has been incorrectly snowflaked into separate
tables. It is typically bad from a performance standpoint to snowflake hierarchies.
The product, brand, and category tables should be merged into one product
table.

8.3.10 Cases for snowflaking
result in the implementation of a snowflake design. In other words, a dimension
table is said to be snowflaked when the low-cardinality attributes in the


dimension have been removed to separate normalized tables and these
normalized tables are then joined back into the original dimension table.

environment because it has drastic affects on the understandability of the
dimensional model and can result in decreased performance because of the fact
that more tables need to be joined to get the results.

We need to identify which dimensions need to be snowflaked for each of our
dimensions that we have chosen for our dimensional model. There are no
appropriate candidate snowflakes identified for the grocery store example.

The following topics on snowflaking are discussed in more detail in 5.3.7,
“Identifying dimensions that need to be snowflaked” on page 171:
What is snowflaking?
When do you do snowflaking?
When to avoid snowflaking?
Under what scenarios does snowflaking improve performance?

on page 432, we observe that the product hierarchy (Product Category →
Product Brand → Product Name) has been incorrectly snowflaked into separate
tables. It is not appropriate for the product hierarchy to be snowflaked. In addition
to this, there are no other tables that need to be snowflaked.

8.3.11 Identify slowly changing dimensions
A slowly changing dimension is a dimension whose attribute or attributes for a
record (row) change or vary slowly over time.

In a dimensional model, the dimension table attributes are not fixed. They
typically change slowly over a period of time, but can also change rapidly. The
dimensional modeling design team must involve the business users to help them
determine a change handling strategy to capture the changed dimensional
attributes. This basically describes what to do when a dimensional attribute
changes in the source system. A change handling strategy involves using a
surrogate (substitute) key as the primary key for the dimension table.

There are three ways of handling slowly changing dimensions. They are:
Type-1: Overwrite the value.
Type-2: Create a new row.
Type-3: Add a new column and maintain both the present and old values.


For more detailed discussion on slowly changing dimensions, refer to the
following:
“Slowly changing dimensions” on page 283
“Slowly changing dimensions” on page 155

on page 432, we observe that product, seller, manufacturing plant, and customer
can be treated as slowly changing dimensions and handled using a Type-2
change handling strategy.

8.3.12 Identify fast changing dimensions
Fast changing dimensions cannot be handled using slowly changing dimension
handling strategies. That is, fast changing dimensions cannot be handled using
the Type-1, Type-2, or Type-3 slowly change handling strategies. The rate of
change of these dimensions is much faster when compared to slowly changing
dimensions. It is important that we review each dimension to see if it is a fast
changing dimension and handle it appropriately. It may be also possible that a
fast changing dimension may have been handled using a Type-2 change
handling strategy. Handling a fast changing dimension using a Type-2 approach
is inappropriate, does not solve the fast changing problem, and can significantly
increase the number of rows in the dimension.

Fast changing dimensions are handled by splitting the main dimension into one
or more mini-dimensions. For more detailed discussion about handling fast
changing dimensions, refer to the following:
“Fast changing dimensions” on page 286
“Handling fast changing dimensions” on page 163

on page 432, we observe that all dimensions can be handled comfortably using
the Type-2 change handling strategy.

The period beginning at the start of the calendar year, up to the current date, is
called Year-to-Date. For the calendar year where the starting day of the year is
January 01, the Year-to-Date definition is a period of time starting from January
01 to a specified date.

In other words, Year-to-Date facts are numeric totals that consist of an
aggregated total from the start of year to the current date. For example, assume
that a fact table stores sales data for the year 2005. The sales for each month
are additive and can be summed to produce year-to-date totals. If you create a


Year-to-Date fact such as Sales_$$_Year_To_Date, then when you query this
fact in August 2005, you would get the sum of all sales to August 2005. In other
words, the Sales_$$_Year_To_Date fact stores aggregated values.

Other types of aggregated facts can be Month-to-Date facts and Quarter-to-Date
facts. Dimensional modelers may include aggregated Year-to-Date facts inside
the fact table to improve performance and also reduce complexities in forming
Year-to-Date queries. However, storing such Year-to-Date facts inside the fact
table may lead this fact to be incorrectly overcalculated if business users count it
twice. Such untrue-to-grain facts should be calculated inside the front-end
reports, and should not be stored as a fact, to avoid confusion.

Suggested approaches for handling Year-to-Date facts are as follows:
Year-to-Date facts can be easily handled by OLAP-based applications.
Year-to-Date facts can also be defined by using SQL functions in Views or
Stored Procedures.

on page 432, we observe that the fact table contains a year-to-date fact called
Year-to-Date_Total at the line item grain. In other words, the fact called
Year-to-Date_Total is untrue to the grain and should not be stored in this fact
table.

8.4 Schema following the design review
After reviewing the initial draft schema as shown in Figure 8-2 on page 432, we
made a number of improvements, resulting in the new design shown in


Date table is implemented
with Order_Date and
Delivery_Date using Views
(Role-playing)

Order_Number is a
Degenerate Dimension

Figure 8-4 Dimensional Model after review

Improvements made to the model include:
The order dimension table has been dropped and replaced by the
Order_Number which is a degenerate dimension. In other words, previously
in the draft model shown in Figure 8-2 on page 432, the Order dimension was
incorrectly represented as a dimension because it really had no attributes.
The attributes, such as order date and delivery date, actually can be well
represented as a separate dimension.
The Order_Date and Delivery_date columns have been removed from the
Order table, which has now been made a degenerate dimension. The order
date and delivery date have now been represented as separate dimensions
using the common date dimension table.This is shown in Figure 8-4. This
concept of using a common table to represent different conceptual
dimensions is called role-playing. The concept of role-playing is discussed in
more detail in “Role-playing dimensions” on page 179.
The quality of all the dimension tables has been improved by adding
descriptions for the following coded columns in Figure 8-2 on page 432:
– Product_Model_Code
– Plant_Code
– Outlet_Code
– Customer_Code


In addition to the above codes, now the dimensional model shown in
Figure 8-4 on page 444 also has descriptive columns that describe these
codes. The new descriptive columns added are:
– Product_Model_Description (To describe Product_Model_Code)
– Plant_Description (To describe Plant_Code)
– Outlet_Description (To describe Outlet_Code)
– Customer_Description (To describe Customer_Code)
In the draft design, we saw that all dimensions were using OLTP natural keys
as primary keys. However, after review, we changed all dimension primary
keys to surrogate keys.
We saw in the old draft design that the product tables hierarchy (Category →
Brand → Product name) had been incorrectly snowflaked into two separate
tables. We believe that snowflaking hierarchies can be bad for browsing
reasons because of the increased number of joins involved when selecting
different components of the hierarchy. As shown in the revised design
(Figure 8-4 on page 444), we have removed the snowflaked tables and
merged the broken hierarchy into a single product table.
We studied all the dimensions in detail and concluded that all dimensions
changing data can be easily implemented using the Type-2 change handling
strategy. Also we determined that there were no fast changing dimensions.
We have removed the two Year-to-Date facts called Year_to_date_cost_total
and Year_to_date_revenue_total from the fact table as shown in Figure 8-4 on
page 444. This is because these year-to-date facts are untrue to the grain
definition which is a single line item on the order.
In the old draft model, we saw that the customer address is stored inside a
single column called Customer_Address. It is difficult to analyze the
customers according to their city, district, state, and region if we have only a
single address field. This is because all of these details are stored inside the
Customer_Address column as a single string. To better analyze the customer
address, we split the Customer_Address column in our improved, post-review
design (Figure 8-4 on page 444) into more detailed address field columns,
such as:
– Customer_Shipping_Street#
– Customer_Shipping_Street_Name
– Customer_Shipping_Suite
– Customer_Shipping_City
– Customer_Shipping_District
– Customer_Shipping_State


– Customer_Shipping_Region


9

Chapter 9. Managing the meta data
In this chapter we provide an overview of meta data, specifically for the purpose
of implementing a data warehouse or dimensional database.

The most common and simplistic description of meta data is data about data.
Since this expression was coined, a long time has passed and the data
warehouse and business intelligence environments have become bigger and
more complex. More complex data warehouses have resulted in more complex
data and more problems with managing data. The need to know more about the
huge volumes of data stored in organizations has also impacted the definition of
meta data. Meta data is now more commonly referred to as information about
data. In other words, meta data is now considered information that makes the
data understandable, usable, and shareable.

Today, companies of all sizes are exposed to quickly evolving business
environments that challenge the way computer systems and databases are
designed and maintained. One of the most significant parts of this challenge
derives from the business intelligence needs of companies today. The volume of
data in companies is growing at a rapid rate, much of which is fueled by the
Internet. While the Internet, and particularly the Web, grow and reach more and
more locations, and computer systems become more and more complex,
generating and relying on more and more data, we have to deal with the
challenge of efficiently managing and taking advantage of volumes of data, and
converting them into useful information. That is where the Business Intelligence
and, more specifically, meta data management play their critical role.


9.1 What is meta data?
In order to explain the meaning and usage of meta data, think of an example that
refers to a typical IT (Information Technology) environment in a mid-size
company where Business Intelligence has not yet been fully exploited.

This company has recently purchased a smaller company in the same branch of
business, and that action has resulted in significant redundancy in certain types
of data. This company already had a number of separate systems dispersed
across various departments. There was some degree of integration between
them, but they were designed at different times, and followed different
implementation approaches. Sound familiar? Actually, this is a somewhat typical
situation in many companies. One of the problems the company now has to deal
with, or more particularly the IT-User community, is to understand the data
available in each system and its interrelation with the data on other systems. In
Figure 9-1, you can see some of the situations mentioned.

Figure 9-1 Data Structure of Company ITSO-CO


As examples:
These systems are not thoroughly documented anywhere, which implies that
a system developer will have to analyze the database to find out what exactly
the format is of any given field.
Many times the table or field names are not meaningful. As examples,
consider names such as T_212_XZZ_3, and P8793_DX as depicted in
Figure 9-2. There needs to be documentation indicating such things as the
primary purpose of each object, how can it be used, and from where it is
populated.

Figure 9-2 Meaningless table names

In Figure 9-3 there is a field called Purchase_Order_Number
(PURC_ORD_NBR) in two different tables. They are Orders, in the Sales
system, and Purchase_Invoices in the Purchasing system. In the Orders
table, Purchase Order Number indicates the purchase order number used by
the customer to order the goods that are being sold to him. In the Purchasing
system, the Purchase_Order_Number indicates the purchase order number
used when buying goods from a supplier. In the Sales system the first field,
ORD_ID, is more typically called Customer_Order_Number. These are
situations where a name in one system is used as synonym for a name in
another.

Figure 9-3 Same name with different meanings

Now consider the Sales systems of the two merged companies. Look for
invoice information. In Figure 9-4 on page 450, notice that the INV_ID field in

Chapter 9. Managing the meta data 449

the INVOICES table of Sales system of the Company ITSO-CO is equivalent
to the field BILL_ID of the Sales system of the Company Res-CO. As well, the
ORDERS table in the PURCHASES system, does not have the same
meaning as in the Sales System.These observations may be more easily
seen in Figure 9-1 on page 448.

Figure 9-4 Different names with the same meaning

Those familiar with these systems may think that all these clarifications are
unnecessary and irrelevant. But think of a new user entering the company. Or,
even riskier, a user changing departments who is sure they know the meaning of
all the names used in the new department. Perhaps now you can get an idea of
the confusion, and misinformation, that can occur in these situations.

Well, the meta data is precisely the information that has to be kept in order to
efficiently describe all the data objects in a company. A primary objective is to
avoid the huge pitfall of having a highly inconsistent and unreliable methodology
for the management of the information about the different data objects being
used in the business. Keep in mind that a data object is not necessarily a
particular field in a form or a table. In this context, when we say Data Object, we
mean such things as fields, tables, forms, and reports. Even entire applications,
such as Sales and Purchases, and all their associated data structures, can be
considered Data Objects subject to the requirement to be described in the meta
data. Keep in mind also that not only business users, but also system developers
and modern and conveniently designed applications, need to work and access
good meta data for efficient and reliable functioning. This is especially critical
when interacting with other users, departments, or systems.

Notice that meta data not only describes structured, database recordable
objects, it can also describe any sort of data container, such as e-mails, pictures,
sound recordings, memos, books, and disks.


9.2 Meta data types according to content
According to the data described, and the usage, we can distinguish four types of
meta data:
Business
Structural
Technical
Operational

9.2.1 Business meta data
Also called informational or descriptive meta data, business meta data describes
the contents of the business objects or any other data object in which the user is
interested. We now take a look at typical questions or problems addressed by
business meta data:
In Figure 9-5, we have two tables from Sales Orders and Purchase Orders
Systems. In both tables we find the field WEIGHT, which is used for freight
calculation purposes with, more or less, the same meaning. However, in one
system the weight is given in Kilograms. In the other system the weight is
given in pounds.

Figure 9-5 Same name, same meaning, different unit of measure

In the company reporting system, we also find a field called
YEAR_TO_DATE_ORDERS. A user wanting to use this field for reporting
purposes for the first time will have questions about it, such as:
– How is the amount given? In dollars (think of an international application)?
If it is in dollars, is it in currency units, or in, for example, thousands (K) of
dollars?


– Are Order Amounts gross or net? That is, with or without such things as
additional costs, taxes, and discounts?
– Does year-to-date include the orders of the present month? Or, the orders
of the present week? What about orders from today, until precisely the
moment of the inquiry?

Reference meta data
The reference meta data is an important type of business meta data and it
describes the lists of values, groups of codes, types, and domains, which
implicitly contain business rule information or static external information, such as
status, geographical codes, types, and categories.

9.2.2 Structural meta data
It is the meta data that describes the structure of the different data objects. It is
useful when implementing navigation and presentation of the data to the user.
For example:
An EMAIL data object is composed of a header and body. The header has
TO_ADDRESS, FROM_ADDRESS, DATE and SUBJECT. The body has
TEXT, and attachment names or descriptors.
A BOOK data object is, for instance, comprised of a TITLE, AUTHOR,
TABLE_OF_CONTENTS, PREFACE, CHAPTERS, INDEX, and GLOSSARY.
Every CHAPTER contains from 0 to many HEADERS.
A PICTURE data object is described by a FILENAME, the FORMAT of the file
(such as BMP or GIF.), and a description.
A SOUND file can be described by the FILENAME, FORMAT, DATE,
HIGHEST_TONE, and LOWEST_TONE, for example.

9.2.3 Technical meta data
This describes all technical data necessary for the development, integration,
functioning, and comprehension of the system applications. It defines source and
target systems, and their table and fields, structures and attributes, and
derivations and dependencies. For example:
Database field formats and lengths, such as CUSTOMER_NAME
VARCHAR(100)
Table physical characteristics, such as TABLE_TABLESPACE_NAME =
TS_MAIN_DATA
Control fields, such as Record_Creation date or Record deleted


This meta data is useful for Business Intelligence, OLAP, ETL, Data Profiling, and
E/R Modeling tools.

9.2.4 Operational meta data
This type of meta data manages the information about operational execution
(events) and their frequency, record counts, component-by-component analysis,
and other related granular statistics.

It is primarily used by operations, management, and business people.

9.3 Meta data types according to the format
According to the format, or type of container, we can also distinguish two types of
meta data:
Structured
Unstructured

9.3.1 Structured meta data
Structured meta data is the meta data that can be efficiently stored and
systematically accessed. For example, a relational database is a good recipient
for structured meta data, since the data will always be accessed in the same way,
for example, by inquiring using the same variables through the same paths. XML
documents are also adequate for storing structured meta data.

9.3.2 Unstructured meta data
Unstructured meta data is stored in repositories or containers that do not allow
systematic access to the information, for example, audio recordings, descriptive
written documents without any internal structure, and employees.

9.4 Design
It is very important that the design of the meta data management system is
user-oriented. Therefore, one of the first steps when building a meta data system
is to conduct a workshop with the users. The goal of this workshop is to ensure
that the business needs are satisfied, and that this is done in a comfortable way
for the user. When designing a meta data system you should consider the
following development methodology and structures:
1. Meta data strategy (Why?)


2. Meta data model (What?)
3. Meta data repository (Where?)
4. Meta data management system (How?)

9.4.1 Meta data strategy - Why?
The first step working with meta data in a Business Intelligence environment is to
develop the meta data strategy for this particular solution. It helps respond to the
question why is it needed along with a detailed description. The strategy
determines the specifics regarding meta data, including:
Is there an existing meta data strategy and direction in place?
Why is meta data needed?
What will be the business value of this information?
Who will use the meta data (including both tools and people), and for what
purposes?
What type of meta data is to be captured?
When will that meta data be captured?
Where will the meta data that is captured and collected get stored?
How and when will this capture occur?
How will the captured meta data be made available?

Once these questions have been answered, then the results must be prioritized
in order to determine the level of effort that is required. The results of this
prioritization allow the work to be directed toward the most important tasks.

9.4.2 Meta data model - What?
The next step is to determine how best to satisfy the prioritized needs identified
in the strategy by determining what specific meta data will be captured. In order
to answer what meta data to capture, an initial identification of the tools that will
be used in the data warehouse environment need to be inventoried. These
include data modeling, database management, ETL, and user reporting and
analysis tools.

Each of the specified types of tools has meta data that is needed as input to
begin the specific tasks that will be supported by the tool. In addition, each tool
produces meta data, which is needed by another tool at another point in time.
The information that should be gathered about the tools is:
A list of all types of meta data that each tool supports.


Identification of the meta data required to start the activity supported by each
tool.
Identification of the meta data that is produced by the activity supported by
each tool.
Determination of what specific types of meta data will be captured.

Note: Before attempting to do this identification, check to see if a meta-model has
been developed for each tool.

You need the previous activities to determine the technical meta data. But you
still need the business meta data, so that users can locate, understand, and
access information in the data warehouse and data mart environments.

Out of the analysis of these two meta data types, you should get a E/R logical
meta data model, normalized to a reasonable extent, that consolidates the meta
data models of the different systems.

History
An important aspect to consider when specifying the meta data structure, is the
history of the meta data. To a lesser or greater degree, the database structures of
a company evolve. This evolution can take place at an application or database
level, where new databases or applications replace old ones. Inside a database,
old tables may not be used any longer and will probably be archived and deleted
from the database. Data warehouses and data marts, because of their iterative
approach, are especially volatile and unstable. For these structures, if the data
history is kept during long periods, and the structure of the data changes, it is
very important to keep track of the changes. Otherwise, there is a high risk that
business information maintained during many years may become useless
because it cannot be understood.

9.4.3 Meta data repository - Where?
The next step will be to determine where to collect the meta data that is captured
during each activity and/or tool evaluation. This location may be in the tool itself
(data models or reports), or collected and integrated into another medium (word
document, spreadsheet, or meta data repository). As with any situation where
information resides in multiple locations, the redundancy may introduce data
integrity and data quality problems. On the other hand, integrating multiple
sources of information requires additional time and effort.

Often the decision will be made to merely capture the meta data in each
individual tool and then virtually integrate this by providing knowledge of where
that meta data resides. Most, if not all, of the meta data can be integrated into a
single meta data repository. This last method is costly but provides the highest


level of quality and should only be attempted if the organization is committed to
maintaining and using the meta data in the future.

9.4.4 Meta data management system - How?
Once the determination of why, what, and where meta data will be addressed in
this Business Intelligence effort, the actual steps to do the capture must be
integrated into the project plan. This task may be fairly straightforward, as
creating database tables and therefore producing the technical meta data for the
tables. On the other hand, ETL (extract, transform, and load) tools and user
query tools can capture a variety of meta data. A decision must be made as to
the appropriate place to capture various types of meta data.

A good example is the business rules. Even though no one would doubt that this
information is important and somehow needed, it must first be determined
whether the results of this discovery are captured. If the decision to capture is
made, then the second decision is where and when that capture will be done. Is it
captured in the ETL tool where it is used as part of the load process or in the user
query where the results are presented? The answer to these questions will
determine whether the meta data captured is considered business (the rule
defined in business terms) or technical (the rule expressed as a calculation or
edit) meta data, or both.

If the meta data will be captured and maintained in the individual tools, the time
to capture and the validation of the effort must be accounted for. If the individual
tool meta data is to be integrated into a separate repository, time must be
allocated for potential integration and resolution effort.

9.4.5 Meta data system access - Who?
The last step, and the one most often ignored, needs to address who will use the
meta data that is captured and how it will be provided to that identified user. The
approach to provide business and technical meta data is often handled very
differently (if indeed business meta data is even considered at all).

The majority of the technical meta data is captured and maintained in individual
tools. This information is most often required by developers or administrators,
and access to this information is via use of the individual tools. This access must
be identified and made available.

The type of business meta data to be captured and the users of this meta data
are much harder to identify, and often ignored. When this business meta data has
been identified, a way of accessing this information by the user must be
developed. Most recently, APIs (application program interfaces) have been used
to capture meta data from where it is stored and used to make it available


through other methods. An example of this is to take the definition of an item on a
report from wherever that definition is stored, and make it available through a
help facility as the report is being viewed.

One of the key points to remember in this step is that if the usage of meta data
cannot be identified, the value of capturing it should be questioned. Data, of any
kind, that is not used tends to decay.

9.5 Data standards
Data standards are part of the meta data and can be understood as the rules or
guidelines to be considered or followed when naming and coding the data
objects, or setting the base of the data structure for every project that creates or
handles data structures. These rules would be applicable to every project or
activity, independently of the number of people and the size of the project. Here
are examples:

Imagine that, in order not to forget your tasks, you write them down on your note
pad, and you always write one short word or code with the task description
indicating the priority and urgency. You may even document the meaning of the
codes somewhere so that you do not forget what they mean. If you do that and
do it consistently and systematically, and you always specify a code and a
description of the task, and you always use the same codes, then you have
designed your own data standards, and, although simple, they are working
efficiently for you.

From then on, you will only find good reasons for designing and maintaining
(meta) data standards. Whether you work in a small team designing data
structures or creating object information, and your work has to be combined and
consolidated with the others’ work, or whether you consider the Web, a
worldwide data structure, where everybody puts information, and from which
everybody gets information, you will find that data standards are essential to
efficiently create, interpret, and manage all the information.
The contents of the data
The naming of the data
The structure of the data
The format of the data

9.5.1 The contents of the data
Data standards referred to as structured contents, such as the ones of a
database, are equivalent to the reference meta data covered in “Reference meta
data” on page 452. In that section we learned that reference meta data contains


information about the business rules, and one of its purposes is, precisely, to
establish unique and common values for the data object contents (such as
names and codes).

9.5.2 The format of the data - domain definition
A data domain, in the context of meta data, is the definition of a common type of
data in your business. In Figure 9-6, you can see an example with the definition
of the PERCENT domain. It states that a PERCENT type attribute is defined as a
numeric field that can be nnn.nn where “.” is the decimal separator. In addition to
the attributes you see in this example, you can specify that the values this
domain can adopt range from 0.00 to 100.00.

Figure 9-6 Example of a data domain definition

Using data domains, modern case tools allow you to efficiently describe the
different percents you have in your database. For example, as you see in
Figure 9-7 on page 459, for defining the attributes AlcoholVolume%,
TotalSales%, or Tax%, you just need the data domain PERCENT.


Figure 9-7 Assigning data domains

Domains are used in data modeling to speed up the process of design, and help
to efficiently maintain the data model.

Following the previous example, if, after the data model is finished, you need to
redefine all the attributes under the same domain, you only need to redefine the
domain. The tool cascades the changes of the domain to all the attributes that
use that domain.

In the following sections we explain and recommend what to use for common
data types.

Dimension keys
For dimension keys and, generally speaking, for data warehouse keys, we
recommend using surrogate keys. We explain the benefits and reasons in detail
in “Primary keys for dimension tables are surrogate keys” on page 263.

Always choose an integer format. And choose the integer type in your database
that provides the largest range. Integer length types vary from database to
database. In DB2, for example, LONGINT, goes up to +9 223 372 036 854 775
807.


Tip: If you are concerned about performance when choosing the length, the
general rule is:

If you foresee having a maximum of 10n records, then choose the integer type
necessary to contain 102n . For example, if you are sure that your table will
never have more than 100,000 records, then choose a surrogate key format to
accommodate 1,000,000,000. With this approach, it is unlikely that you will run
out of identifier values.

Date data type
This is a data type that generates much confusion and causes trouble for people
and programs when not correctly specified, published, and accepted. Just think,
for example, about the date 10/12/05. In Europe that date corresponds to the
tenth of December. In the United States, however, that date corresponds to the
twelfth of October. And the year? Does it refer to 1905, or to 2005, or to some
other century?

Always define dates so that you use them consistently. For example,
mm/dd/YYYY, where dd is the day of the month, mm is the month of the year, and
YYYY is the year in the Gregorian calendar. These are typical values Western
countries use. It is good practice not to use a YY format for the year. You may
recall the issues caused by a two-digit format when we changed centuries from
1900 to 2000.

Exceptions
Some applications still accept years in YY format. The important thing is that you
define appropriate defaults and the year information is stored in the database in
four digit format. Depending on the business subject, you should define your
defaults in one way or the other. An example of how to process these date types
is depicted in Example 9-1.

We do not see the benefit in using a two digit data format, and recommend you to
always present the complete four digit year. Unless you are quite convinced
about the short life of your application and the evolution and situations of the
business are covered, this approach could become unreliable for assigning a
wrong century to the years. For example, with regard to birth dates, it is possible
to find people born in 2005 and 1905.

Example 9-1 Year format example
Date of first ADSL contract in YY format (in pseudocode):
if YY > 90 then ADSL_DATE = 19YY
else ADSL_DATE = 20YY


Time data type
Time is less confusing than an unspecified date format, but also prone to
misunderstandings. Always document the format you want, need, or expect to
use in your daily operations, such as hh:mm:ss or hh:mm. Also specify the
separator, as “:” or “.”, and do not forget to indicate whether it is in 12 hour format
(a.m./p.m.) or 24 hour format.

Boolean data type:
This is a type of domain that contains only two logical values: true and false.
Unless the database server you use explicitly manages boolean values, we
recommend you define this domain as a numeric field admitting the two values 0
and 1. You can say that, for the user, a YES/NO field is better to understand than
a number. But to get the user accustomed to it and gain acceptance to this
solution is relatively easy. And, the database systems can handle 0 and 1 more
efficiently, if appropriately defined. In addition, it makes additive operations
straightforward.

For example, assume that you have defined a boolean field named
HAS_NO_CAR as numeric, and this field admits the values 0 and 1. If you want
to find how many customers have not yet gotten a car, you only define a SUM
operation on this field. Whether you are using a modern reporting tool, or you are
manually entering the corresponding SQL-SELECT statement, this is a very
straightforward operation.

Now imagine that you have defined the field as CHARACTER. The selection of
the customers who have not yet gotten a car is still easy. But if you or the user
needing this information do not remember the two exact values of the boolean
field, and the database or user application makes a difference between upper
case and lower case, you can have trouble remembering which of the following
variants is the right one: Y/N, y/n, Yes/No, YES/NO, yes/no, or True/False. This
variant is also less appropriate because you have to cast between different data
types. If you do not know well the database server system and do not know how
to use its functions, or the OLAP tool you use does not efficiently support the
data type conversion needed, then you have additional work to do when using
boolean values for selection or calculation purposes.

Tip: Define your boolean domain as numeric, and use the values 0 and 1 for
false and true respectively. This will enhance the data management
performance, and simplify the arithmetic operations based on boolean fields.


9.5.3 Naming of the data
In the section 9.1, “What is meta data?” on page 448 we saw how important is to
have a common and unique understanding of the names given to every data
object. In this section we list and describe options that help you to properly name
your data and determine naming rules for your project or data structure.

Historical considerations
In the past, one of the primary characteristics of computer systems was the lack
of enough space for comfortably storing all the information handled. That
affected the way data was stored, for example, when full postal addressees had
to be stored in 20 character fields. But it also imposed limitations when naming
system component names, such as, for instance, tables or fields. Because of
that, system developers and users entering data had to start abbreviating as
many words as possible. And, due to the lack of resources that database
administrators and developers had available when working with data objects,
often the name given to the data objects had nothing to do with the contents of
the data object.

Here is an example taken from a recent project. During the physical design of the
database of a new data warehouse, the customer insisted on using the same
naming rules for the new data warehouse tables they had for their existing OLTP
environment. So, for example, the table containing the Customer Orders was
DWHTW023, where DWH was the user or schema owning the table, T indicated
that it was a table, “023” was a sequential number, and W had no special
meaning as far as we could recall. Here ORDERS would have been a more
appropriate name.

Remember that the standards are for helping business operations to work
reliably, provided the benefit the standards bring is larger than the overhead they
add to the business. It was correct to try to follow the standards, but it would have
been much better to have changed and adapted somewhat to the new
environments and business area needs, which was, reporting and data analysis
activities.

Naming: General considerations
Although it is common to use client tools for OLAP and reporting activities, often
the business user is exposed directly to the database tables of the data marts or
even to the main enterprise data warehouse tables. It is therefore important to
name the table and fields as meaningfully as possible. The name should
whenever possible refer to its contents and nothing else. The goal is that by
looking at the name of the table or field, the user has no doubt about its contents.
In the previous example, none of the parts of the name had any value or meaning
to the user.


Tip: Naming booleans. Whenever possible, use a verb in third person, such
as, HAS, IS, and OWNS. For example, HAS_ALCOHOL,
IS_PREFERRED_CUSTOMER, and OWNS_MORE_THAN_1_CAR. This is
sufficient to indicate to the user that the contents of the field is of the type
true/false.

Abbreviations
Due to system limitations, but also in normal written or spoken communications,
people tend to shorten the words for what we euphemistically call language
economy. At the end we frequently find conversations, communications, and
documents with such a high percentage of abbreviations, that somebody new to
the business or the subject probably does not understand.

And if there is a tendency to shorten what we verbally say, this tendency,
regarding computer systems is even bigger because we have to key it instead of
saying it. Fortunately any modern computer applications dealing with data at any
level, including OLAP applications, offer easy-to-use drag-and-drop features that
allow us not to have to worry any longer about our typing skills or the high
probability to misspell a field. In OLTP environments you may still consider it
advantageous to use abbreviations for naming data objects, because the user is
normally not exposed to the internal data structure. However in a business
intelligence environment we highly recommend you avoid the usage of
abbreviations. Only names that appear in practically all tables are candidates for
abbreviation. For example, Name, Description, Amount, and Number. Otherwise,
use full names, unless the database physical limitations force you to abbreviate.
And if you use abbreviations, abbreviate consistently following the standard
company abbreviation list.

Technical prefixes, midfixes, and suffixes: Substrings containing technical
or meaningless information, such as T for table, can be useful in OLTP
environments. However, avoid them in data warehouse environments.

In Chapter 7, “Case Study: Dimensional model development” on page 333, we
used the standard abbreviation list used in IBM. An example is in Figure 9-8 on
page 464.


Figure 9-8 IBM abbreviation list

Glossary
Just like abbreviations, companies develop their own language or use acronyms
typical of the respective business branch. It often happens that, during business
conversations, users working in one department do not understand the users of
another department. This situation gets even worse when the listener, or reader,
is external to the company. In order to solve or alleviate that situation it is normal
to maintain a corporate glossary, or according to the dimension and complexity
of the areas covered, specialized glossaries.

A glossary is a document or database, on paper or in electronic format, with a list
of all the business words used in the daily activities of the company and their
corresponding descriptions.

A company should always maintain a central and unique glossary. Modern tools
allow an easy integration and publishing of this data.


9.5.4 Standard data structures
In 9.5.2, “The format of the data - domain definition” on page 458, we explained
that a domain is used to describe the structure of an attribute or field for
standardization purposes. In the same way, you can use a domain to predefine,
or standardize, upper level structures, for example, the fields of a physical
dimension table, the attributes of a logical entity, and the relationship between
determinate types of tables.

Structure example
When designing a dimensional model, there are several data objects that should
be among your common objects. They are depicted in Figure 9-9.

Figure 9-9 Example of a standard basic dimensional design

These are common elements to all the logical entities: “Names”, “Descriptions”,
and “Natural Ids”. Keep in mind that this is an example for orientation purposes,
and you might need different, or additional, attributes when designing your
dimensional model. It is even possible that you may only need one attribute such
as “Name” or “Natural Id”, and no “Description”. The need to include one or the
other may change according to the business requirements. Nevertheless, at the
corporate level, it is still possible, and recommended, to agree and propose a
common fixed set of fields, even if not yet requested. Consider this from the
perspective of the conformed dimensions, (see 6.4.3, “Conformed dimensions”
on page 268), and consider how much will it take to implement a new field into
the dimensions, and how big is the probability that a business process needs that
information in the near or long-term future.


All the attributes we described above are defined at the logical levels. At the
physical level, however, we need additional physical fields, depending on the
auditing and technical controls you might need to perform on your data. An
example is depicted in Figure 9-10. There we see more fields which have no
particular business meaning but are needed for various technical purposes. For
example, the surrogate keys (Dimension Keys) and the Record Creation Date,
are necessary to know exactly when this information was written into the
database.

Figure 9-10 Example of a physical structure

Figure 9-11 on page 467 shows an example of a logical representation that can
be used as a reference. Especially the data types and formats must be defined
and consistently used across the data structure.


Figure 9-11 Reference meta data examples

Reference meta data structures
Reference tables are usually good for describing business rules, for example,
product classes, sale order types, and tax types. In Figure 9-12, entities,
including attributes, refer to typical business information, such as Product
Categories, and Customer Classes.

Figure 9-12 One physical table per meta data reference type

Although, at a logical level, every type of reference meta data deserves its own
logical entity, when physically implementing there are several approaches. The
key common ones are:
One physical table for each reference meta data type
Only one physical table for all the reference meta data types


Reference meta data in separate tables
One way to physically implement the reference meta data is by creating one table
for every group of reference codes or types. So the physical design to the group
above would convert into one such as that represented in Figure 9-13 on
page 469. This approach is convenient because of the relatively easy way to join
the tables.

And, there is another potential advantage. If you use reference codes whose
natural identifiers have significance and they are not foreseen to change either
the code or the meaning, you might consider keeping them as identifiers.
Suppose you have a Reference Table called “CURRENCIES”, and you have
values such as GBP for British Pounds or USD for US Dollars. Although we insist
on the usage of surrogate keys for uniquely identifying the information at a table
record level, in this case it might be reasonable to use this short code as the
primary key identifier. Practically any entity describing information for which ISO
codes are available (such as Countries and Currencies) is a good candidate for
this content denormalization. If considering Figure 9-12 on page 467, you need
to select the Sales Offices in a certain country, you will not need to join the
country table (assuming you know the short code of the country, such as CA for
Canada or FR for France, for example). If you are highly concerned about query
performance, you can follow this approach when populating your reference
codes.

However, this first approach presents the inconvenience of having to maintain
dozens, possibly hundreds, of additional tables. That can be a problem if you
expect frequent changes or additions to the reference meta data structure. If you
are concerned about overall database performance you may not want to follow
this approach, because having to maintain so many small tables will cause, to
more or less of a degree, extra maintenance overhead to the database server,
compared with the efficiency reached when managing fewer but larger tables.

Reference meta data in one main table

Another common solution is to put all or most of the reference meta data into one
main table, as shown in Figure 9-13 on page 469.


Figure 9-13 One table containing most or all the reference codes

This approach is particularly useful when you have numerous reference tables or
you foresee that you will add new ones frequently. In this case with just two
tables you can cover as many single reference tables as you need. However the
inquiry and usage of the reference data is not as straightforward as when you
have one table for each reference data group, or reference area. And in order to
be able to sort the reference codes correctly you will need to handle them as
numeric or alphanumeric. As you can see in Figure 9-13, the table COUNTRY
has been left apart for having a different structure and needing two additional
fields: ISO_COUNTRY_CODE and ISO_CURRENCY_CODE.

The use of reference meta data in dimensional modeling
In a dimensional data model you will want to select, filter, or group your data
using available reference information, as in the example of the figure below
where “Marketing Customer Category” can be WHOLESALER, STORE CHAIN, GAS
STATION CHAIN, RESTAURANT, or SMALL SHOP. You may decide that you just want
to have those names available in the CUSTOMER table. But imagine that you have
to perform negative querying, as an example, producing a list of all “Marketing
Customer Category” names without any sale during the previous year, or select
by other attributes that you do not want to or cannot store in the dimension
tables. In that case you may consider the use of the formal reference code


identifiers in the dimension tables, and the implementation of reference tables in
a snowflake fashion, such as Country in Figure 9-14.

Figure 9-14 Example of denormalized and snowflaked reference data

9.6 Local language in international applications
Companies are challenged by the business globalization that the Internet has
triggered, and more particularly, the Web. In order to effectively reach the
maximum number of customers, and efficiently work with the maximum number
of partners, you need to translate into other languages not only the database
contents, but also the interface component labels, such as fields and forms.
Typically the translation of the presented information label on the screen is
enough, but sometimes it may require additional comments or notes fields,
particularly when the field labels can be extended by context help.

An example is depicted in Figure 9-15 on page 471. Here we make the point that,
because the students are of a different culture from the teacher, the material is
best learned if it is presented in the language of the students.


Figure 9-15 Different cultures, different languages

In Figure 9-16 we show how it would be possible to store multilingual information
applied to the data. Here we assume we want to keep translations of names and
descriptions for the company products and translation of descriptions for
contracts in several languages. This modeling is done at the business data
content level. If, for example, you had to translate the names of the labels of the
fields you see in the screen, or the fixed headers of a report, then you would have
to provide a similar translation data structure within the meta data model so that
the names of the meta data objects, and not only the contents, also get
translated.

Figure 9-16 Example of structure necessary to store data for different languages


9.7 Dimensional model meta data
In this section, we list and describe all the aspects that have to be considered
when modeling or defining the meta data of a dimensional model. We also
indicate how they fit into the dimensional model design life cycle (DMDL).

The following meta data is collected during the phases of the DMDL:
Identify business process: The output of this phase results in the
Requirements Gathering report. This report primarily consists of the
requirements for the selected business for which you will design the
dimensional model. In addition to this, this report also consists of various
business processes, owners, source systems involved, data quality issues,
common terms used across business processes, and other business-related
meta data.
Identify the grain: The output of this phase results in the Grain Definition
report. This report consists of one or multiple definitions of the grain for the
business process for which the dimensional model is being designed. It also
contains the type of fact table (transaction, periodic, or accumulating) being
used. The grain definition report also includes high level preliminary
Identify the dimensions: The meta data documented for this phase contains
the information shown in Table 9-1. It includes such items as dimension
names, hierarchies, updates rules, load frequency, load statistics, usage
statistics, archive rules, archive statistics, purge rules, purge statistics, and
attributes.

Table 9-1 Identify the dimensions - meta data

Name of Dimension The name of the dimension table.

Business Definition The business definition of the dimension.

Alias Specifies the other known name by which the business
users know the dimension.

Hierarchy The hierarchies are present inside the dimension. They
could be balanced, unbalanced, or ragged.

Change Rules Specifies how to handle slowly changing dimensions
(type-1, type-2, or type-3) and fast changing dimension.

Load frequency The frequency of load for this dimension is typically daily,
weekly, or monthly.



Load Statistics Consists of meta data such as:

Usage Statistics Consists of meta data such as:

Archive Rules Specifies whether or not the data is archived.

Archive Statistics Consists of meta data such as:

Purge Rules Specifies any purge rules. For example, Customers who
have not purchased any goods in the past 48 months will be
purged on a monthly basis.

Purge Statistics Consists of meta data such as:

Data Quality Specifies quality checks for the data. For example, when a
new customer is added a search is performed to determine
if we already do business with this customer in another
location. In rare cases separate branches of a customer are
recorded as separate customers because this check fails.

Data Accuracy Specifies the data accuracy. For example, incorrect
association of locations of a common customer occur in
less than .5% of the customer data.

Key The key to the dimension table is a surrogate key.

Key Generation Method This meta data specifies the process used to generate a
surrogate key for a new dimension row.
For example, when a customer is copied from the
operational system, the translation table (staging area
persistent table) is checked to determine if the customer
already exists in the dimensional model. If not, a new key is
generated. The key, along with the customer ID and
location ID, are added to the translation table. If the
customer and location already exist, the key from the
translation table is used to determine which customer in the
dimensional model to update.



Source This includes the following meta data:

→ Name of the source system table: <Table Name>

→ Conversion Rules: These specify how the insert/update
to the dimension table occurs. For example, rows in each
customer table are copied on a daily basis. For existing
customers, the name is updated. For new customers, once
a location is determined, the key is generated and a row
inserted. Before the update/insert takes place a check is
performed for a duplicate customer name. If a duplicate is
detected, a sequence number is appended to the name.
This check is repeated until the name and sequence
number combination are determined to be unique. Once
uniqueness has been confirmed, the update/insert takes
place.

→ Selection Logic: Only new or changed rows are
selected.

Conformed Dimension This specifies if the dimension being used is a conformed
dimension.

Role-playing Dimension This specifies if the dimension is being implemented uses
the role-playing concept. Role-playing is explained in 5.3.9,



Attributes (All columns The meta data for the all the dimension attributes includes
of a dimension) the following:

→ Name of the Attribute: <Attribute Name>

→ Definition of the Attribute: Attribute definition

→ Alias of the Attribute: <Attribute Name>

→ Change rules for the Attribute: For example, when an
attribute changes, then use Type-1, Type-2, or Type-3
strategy to handle the change.

→ Data Type for the Attribute: Data Type, such as Integer
or Character.

→ Domain values for the Attribute: Domain range such as
1-99.

→ Derivation rules for the Attribute: For example, a system
generated key of the highest used customer key +1 is
assigned when creating a new customer and location entry.

→ Source: Specifies the source for this attribute. For
example, for a surrogate key, the source could be a system
generated value.

Facts This specifies the facts that can be used with this
dimension.
Note: Semi-additive facts are additive across only some
dimensions.

Subsidiary Dimension This specifies any subsidiary dimension associated with
this dimension.

Contact Person This specifies the contact business person that is
responsible for maintaining the dimension.

The meta data information shown in Table 9-1 on page 472 needs to be
captured for all dimensions present in the dimensional model.
Identify the facts: The meta data documented for this phase contains the
information as shown Table 9-2.

Table 9-2 Identify the Facts - meta data

Name of Fact Table The name of the fact table.



Business Definition The business definition of the fact table.

Alias The alias specifies the other known name by which the
business users know the fact table.

Grain This specifies the grain of the fact table.

Load frequency The frequency of load for this fact table, such as daily, weekly,
or monthly.

Load Statistics Consists of meta data such as:

Usage Statistics Consists of meta data such as:

Archive Rules Specifies whether or not data is archived. For example, data
will be archived after 36 months on a monthly basis.

Archive Statistics Consists of meta data such as:

Purge Rules Specifies any purge rules. For example, data will be purged
after 48 months on a monthly basis.

Purge Statistics Consists of meta data such as:

Data Quality Specifies data quality checks for the data. For example,
assume that we are designing a fact table for inventory
process. Inventory levels may fluctuate throughout the day as
more stock is received into inventory from production and as
stock is shipped out to retail stores and customers. The
measures for this fact are collected once per day and thus
reflect the state of inventory at that point in time, which is
the end of the working day.



Data Accuracy This specifies the accuracy of fact table data. For example,
assume that we are designing a fact table for the inventory
process. We may conclude that the measures of this fact are
97.5% accurate at the point in time they represent. This may
be based on the results of physical inventories matched to
recorded inventory levels. No inference can be made from
these measures as to values at points in time not recorded.

Grain of the Date Specifies the grain of the date dimension involved. For
Dimension example, the date dimension may be at the day level.

Grain of the Time Specifies the grain of the time dimension involved. For
Dimension example, the time dimension may be at the hour level.

Key The key of the fact table typically consists of concatenation of
all foreign keys of all dimensions. In some cases, the
degenerate dimension may also be concatenated to
guarantee the uniqueness of the primary composite key. This
is shown in Figure 6-31 on page 302.

Key Generation The key generation method specifies how all the foreign keys
Method are concatenated to create a primary key for the fact table.
Sometimes a degenerate dimension may be included inside
the primary key of fact table to guarantee uniqueness. This is

Source The meta data for the source includes the following:

→ Name of the Source: <Source Name>

→ Conversion rules for the source: Rules regarding the
conversion. For example, each row in each inventory table is
copied into the inventory fact on a daily basis.

→ Selection Logic: The logic for selecting the rows.

Facts This specifies the various facts involved within the fact table.
They could be:
→ Additive
→ Non-additive
→ Semi-additive
→ Pseudo
→ Derived
→ Factless Fact
→ Textual

Conformed Fact This specifies if any conformed facts are in the fact table.



Dimensions This specifies the dimensions that can validly use these facts.
Note: Some facts are semi-additive and can be used across
only some dimensions.

Contact Person This specifies the contact business person responsible for
maintaining the fact table.

Verify the model: This phase involves documenting meta data related to user
testing (and its result) on the dimensional model. Any new requests made or
changes to the requirements are documented.
Physical design considerations: The meta data documented for this phase
contains the information as shown Table 9-3.

Table 9-3 Physical design considerations - meta data
Physical design Description
consideration meta data

Aggregation The aggregation meta data includes the following:

→ Number of Aggregate tables

→ Dimension tables involved in creating aggregates
→ Dimension Hierarchies involved in creating
aggregation

→ Fact table and facts involved in creating aggregates

Information relating to the aggregate tables can include
such items as the following:
→ Load frequency
→ Load statistics
→ Usage statistics
→ Archive rules
→ Archive statistics
→ Purge rules
→ Purge statistics
→ Data quality
→ Data accuracy

Indexing This specifies the indexing strategy used for dimension
and fact tables.

Based on all the information listed in this section, we have built the meta data
model shown in Figure 9-17 on page 479. It properly describes the relationships
between the different components. Notice that it is a logical view, and that, for the


sake of understanding of the data structure, some subtypes have not been
resolved.

We have specifically tailored this meta data model to the star schema meta data.
However it could be easily integrated in other more generic meta data structures.

Figure 9-17 Meta Data model


9.8 Meta data data model - an example
In Figure 9-17 on page 479, the data model is focused on the star schema meta
data. In Figure 9-18, we show a more generic meta data model which covers not
only data stored in databases, but also that stored in other types of data
containers. We have chosen a traditional fixed explicit data structure to show the
possibility of storing meta data about such types as documents, web pages, and
computer office files. If the business you want to model is rather dynamic, and
the data structures or business processes change with a relatively high
frequency, you might decide to design a more abstract data model where the
attributes and entities do not need to be specified in advance.

paper file
is implemented as domain attribute representation
file id
material attribute representation DAR id
area
MAR Id DAR name
section
cabinet attribute MAR name DAR description
folder data format identifier (FK) data format identifier (FK)
attribute identifier (FK)
standard name Glossary
attribute domain id (FK) describes / is described by MAR: Material Attribute Representation
is stored in / MAR description
attribute name Field: Attribute representation in not
physically contains attribute identifier (FK)
standard name structured files, like ASCII or Text files,
DAR id (FK)
attribute description data format etc.
comments
file document comments data format identifier
questions
MAR Id (FK) issues data format name
file id (FK) Describes the data of / data format description Other data source could also be modelled:
comments Is described by comments e-mail distribution lists, newsgroups, fax
on-demand, radio and TV programs
(weather forecast, teletext), etc.
is the master reference of /
is the image of

field
document MAR Id (FK) web page
MAR Id (FK) MAR Id (FK) column GUI
source field name
MAR Id (FK) MAR Id (FK)
document code file sequence number WWW address
document name length web page name column name GUI field internal name
document description decimal places web page description standard name GUI field external name
source field description column description field description
contains / data type data type
extracted from is the master reference of / lenght lenght
is the image of decimal places decimal places
text extract
spreadsheet extract system file field table
MAR Id (FK) displays / is displayed in
MAR Id (FK) MAR Id (FK) table identifier
chapter system file id (FK) table column
section from cell table name
to cell table identifier (FK) program field
column comments logical entity identifier
table description MAR Id (FK) MAR Id (FK)
paragraph
comments program id (FK)
line comments
word VALEX Instructions comments
document id (FK) SQL Code
system file
uses / is used by
system file id
database program
file name
server name database table database identifier program id
text extract end domain name database identifier (FK) database name program name
MAR Id (FK) disk table identifier (FK) database description library name
directory comments
chapter comments module name
section system name
column comments
paragraph
line
word

Figure 9-18 Different types of data objects described by the meta data


9.9 Meta data tools
Years ago there were not many types of data management tools, apart from the
database management systems. And these systems were the only ones offering
a meta data system, which was called the Data Dictionary. This was just an
externalization of the user data structure descriptions needed by the database
management system itself in order to function. And it was covering a limited
scope of what we understand today as meta data.

Since then, as data management has evolved, new tool groups, offering to a
greater or lesser degree of meta data management functionality, are available.
These groups, according to their data management focus, are divided into
categories such as ETL (extract, transform, and load), data modeling, data
warehouse management, content management, and OLAP.

Today, every data management tool, regardless of the type, normally provides
meta data management functions. The problem is that typically the functionality
is focused on a specialized area. That is, ETL tools will be very technical and will
describe the transformation of data between systems, but often will not
implement more abstract business meta data. Or, looking at the other extreme,
some OLAP tools will be prolific in mapping and describing business data, but
will not describe the meta data present in OLTP systems.

The primary issue resides in the integration of the tool with all the other
processes and data structures of the business. Today, almost every company has
to deal with information of the types presented in 9.2, “Meta data types according
to content” on page 451, that is, business, technical, structural, and reference
meta data. Business meta data, in certain companies, not only covers the data
but also the processes responsible for the generation and transformation of data
at any level. In these situations, it is typical for companies implementing a meta
data system to focus on the most relevant area and choose the meta data
management tool based on the coverage of that area.

9.9.1 Meta data tools in business intelligence
There are three primary approaches for managing meta data:
Manually
Customizing a meta data tool in order to collect information from other
systems, or to transfer information by using adequate meta data information
exchange protocols.
Using the available tool modules in order to directly exchange meta data.


Each tool that supports the effort in a business intelligence environment has a
need for information to begin. This is often meta data that has previously been
collected or produced. How this information is provided (for example, by manual
input or tool to tool exchange) must be identified. As the tool is used, additional
meta data is created. When the usage of the tool or activity is completed, there is
meta data that has been produced as the results of that effort. That meta data
should be considered the minimum requirements.

To prevent administrative inefficiency, a mechanism must be found for tools from
different vendors to share common information. This can be a very difficult task.
For example, each tool defines the same object in a different manner and
therefore the objects are not easily shareable. Furthermore, the same meta data
is redundantly defined and stored in multiple meta data stores, data dictionaries,
repositories, database catalogs, and copy libraries, each with a different API.
Typically, the tools cannot exchange meta data among themselves because the
meta data is stored in a proprietary format understood only by a specific tool. As
a result, changes to meta data in one meta data store are not easily changed in
the others.

The work of the Object Management Group (OMG), of which IBM is a major
contributor, has concentrated on minimizing these differences through the
adoption of a standard meta model for BI efforts and the recommendation of a
standard Extensible Markup Language (XML) interface. The results of these
efforts should minimize the difficulty in the future.

9.9.2 Meta data tool example
We now look at the information we have been discussing, from a tool
perspective. For this discussion, we have chosen MetaStage®, the meta data
management tool from Ascential™ Software, an IBM company.

MetaStage introduction
This tool enables integration of data from existing data warehousing processes.
A special program module monitors these processes that generate operational
meta data. The tool then stores and uses the operational meta data. This
enables the collection of detailed information about what has happened across
the enterprise.

Most data warehousing products generate their own meta data, with no common
standard for exchanging the meta data. MetaStage integrates all the meta data in
a centralized directory, from which you can share the meta data with other tools.
No additional work is required by the tool vendors. This means you can use
unrelated data warehousing tools in a single environment.


By integrating meta data from the entire enterprise, MetaStage can provide
answers to such questions as where a piece of data came from, what
transformation and business rules were applied, who else uses the data, and
what the impact would be of making a particular change to the data warehouse. It
also quickly provides detailed hyperlinked reports about items of data and their
relationships with other data.

In summary, MetaStage serves as an enterprise-wide tool for many data
warehousing functions, such as:
Synchronizing and integrating meta data from multiple data warehouse tools
Automatically gathering operational meta data from operational systems
Sharing meta data, job components, and designs
Browsing, querying, searching, and reporting on all the meta data of the data
warehousing assets from a single point
Understanding the sources, derivation, and meaning of data in the data
warehouse
Assessing the impact of changes on data warehousing processes

Integrating meta data
You can integrate your own tools and procedures with MetaStage. It enables you
to transfer and translate meta data between different external tools, including
modeling, design, extraction, transformation, and data analysis tools. For
instance, you can collect meta data from these other tools and use it to design
data warehouse processes. Or you can export meta data directly to data analysis
tools to avoid tedious manual data entry.

MetaStage does not impose a single common model on the meta data you want
to share. A common model gives you only the lowest common denominator of
the meta data that your data warehousing tools can share. Instead, this tool
breaks down the meta data from a data warehouse tool into semantic atoms that
can then be reconstituted for use by other data warehouse tools. Individual
MetaBrokers, described in “MetaBrokers” on page 492, provide an interface
between each tool and the semantic units of meta data held in a particular
directory.

This tool captures and stores the following types of meta data:
Design meta data, used by designers and developers to define
requirements. It includes data models, business meta data, and
transformation job designs.
Physical meta data, created, managed, or accessed by tools when they are
executed.


Operational meta data, which tells you what happens when a data
warehouse activity runs, particularly with regard to the way it affects data
warehouse resources.

Note: MetaStage uses MetaBrokers to import and export meta data
directly from and to data warehousing tools.

Capturing operational meta data
MetaStage can automatically capture the meta data that describes events that
occur when a data warehouse process is running. It then uses this operational
meta data to build a view of the relationships between the resources in the data
warehouse. By enabling the combination of the operational meta data with the
design meta data for these activities, this tool provides you with extensive query
capabilities.

Sharing meta data
Meta data can be shared throughout the enterprise with users of the leading data
warehousing tools with MetaStage. The major steps to do this are:
1. Create the meta data using an appropriate tool, such as DataStage™ or
Platinum ERwin.
2. Import the meta data into this tool via a tool-specific MetaBroker® or capture
it with the Process MetaBroker.
3. Publish the selected meta data, so that interested users can subscribe to it
and export it for use with their data warehousing tool. Note, this does not have
to be the same tool used to create the meta data.

For example, you can import meta data from a database design tool into a
particular directory, then export that meta data to DataStage to define the tables
and columns to be populated by a DataStage job. Or you might export the meta
data for use by a business reporting tool, such as Business Objects. You can also
use the Send to Database functionality to export data to a set of relational tables.

Analyzing and reporting on resources
After you import meta data into a particular directory, or capture operational meta
data from data warehousing activities, you can browse, query, or search the
complete meta data structure in the directory, following relationships between the
objects. These functions are all accessed from a tool component called the
Explorer, which is described in “Explorer” on page 495.

You can do the perform the following functions:
Inspect the attributes and relationships of one or more objects in the tool
directory.


Specify which person or organization is responsible for an object.
Associate the object with a business term, the term with a glossary, and the
glossary with a business domain.
Browse from an object, following its relationships to other objects it contains
or depends on.
Run a simple search of a directory.
Investigate the lineage of data in the data warehouse.
Investigate the execution history of processes, such as DataStage jobs.
Use cross-tool impact analysis to investigate the impact of making a change
to a data warehouse resource or process.
Build and run customized and predefined queries for more complex searches.
Document your meta data by creating reports in a variety of formats, including
HTML and XML.

Inspecting objects
This tool makes it easy for you to inspect any object in the tool directory—a
DataStage table definition, for example—viewing its attributes and relationships,
its overlap in other data warehousing tools, and the queries and templates you
can run to generate more information about it.

Browsing from an object
Using the navigation pane in the Content Browser, you can follow the
relationships from any meta data object—a DataStage project, for example—to
all the objects that populate those relationships, up and down the hierarchy. You
can also display the same object from the perspective of a different data
warehouse tool—using a different MetaBroker view—and follow the different
relationships available in that tools view. An example content browser is depicted
in Figure 9-19 on page 486.


Figure 9-19 Explorer: Content Browser

Performing a simple search
Simple searches let you scan a subset of a particular directory for objects that
meet criteria based on name, description, date, and meta data type. You can
save the properties of the objects found by the search to a text file that can be
read by Microsoft Excel, for example.

Investigating data lineage
Data lineage investigations of operational meta data enable you to find the
history of a data item, for example, its source, status, and when it was last
modified. When you combine data lineage and impact analysis, you can answer
such questions as “From where is a particular column of a particular physical
target table derived?” and “When was it last derived?” The tool, Process
MetaBroker, gathers the operational meta data to answer questions about data
lineage by monitoring activity that affects the data warehouse. It records the data
resources affected by, for example, a job run, and whether they were written to or
read from. This tool can also import operational meta data from DataStage
mainframe jobs. An example of a data lineage report is depicted in Figure 9-20
on page 487.


Figure 9-20 Data Lineage report

Performing process analysis
Process analysis investigations are similar to data lineage investigations, but
they give information on process execution rather than data movement. Process
analysis lets you look at the history of process executions by examining
operational meta data. You can answer questions such as “What are the details
of the last run of each of these executables?”, “Which parameters were used
each of the last three times this process ran?”, and “What information is available
about the job that generated this fail event?”

Performing impact analysis
The impact analysis function enables you to answer questions such as “What
else will be affected if I make this change?” and “What else is this object
dependent upon?” For example, you can find out the effect of deleting a
particular transform definition from a particular directory.

This tool provides impact analysis by following relationships within the meta data
stored in a directory. The cross-tool support means these relationships can span
tools and therefore the whole realm of your data warehouse implementation. For
example, you can use this tool to find out not only which DataStage jobs are
affected by a change to a routine, but also which CASE diagram is associated
with the routine. An example of a cross-tool impact analysis is depicted in


Figure 9-21 Cross-tool impact analysis

Running queries
To fully explore the meta data in a particular directory, you can build your own
customized queries using the Query Builder. This lets you answer questions that
involve following complex relationships between data warehouse components.
The result of the query is a collection of objects, sorted according to your
specified criteria. You can refine a search by using the result of one query as the
starting point for another.

A query can be saved, then reused to examine different collections of objects.
You can also export queries to a text file, from which they can be imported into
other tool directories.

Documenting your meta data
This tool lets you almost instantly create complex, hyperlinked reports that show
the relationships and attributes of an object or a collection of objects in a
particular directory. You can create these reports in HTML, XML, RTF, and text
format and print them out or display them in a Web browser, Microsoft Word, or a
text editor. An example of the MetaStage Browser is depicted in Figure 9-22 on
page 489.


Figure 9-22 MetaStage Browser

MetaStage components
MetaStage consists of client and server components. The primary components,
and the flow of the meta data, is depicted in simplified form in Figure 9-23.

Data Business
Modeling Tool SAP BW
Intelligence

Metadata Metadata Metadata

MetaStage MetaStage
Clients MetaStage Directory Browser

ProfileStage DataStage
Impact
Analysis QualityStage

Figure 9-23 MetaStage components

The tool has a client component that lets you browse and search a directory, and
provides administration functions. The directory and the Process MetaBroker are


server components. Meta data is added to a particular directory in the following
ways:
Importing meta data from external tools via MetaBrokers
Using the Process MetaBroker to capture operational meta data generated by
data warehousing activities

The sections that follow provide overviews of the MetaStage components.

Directory
This tool holds meta data in a particular directory. You can import static meta
data from external tools, such as ERwin or DataStage, and capture operational
meta data automatically from data warehousing activities. A directory is a server
component. It is built using relational database tables on a Windows® NT or
UNIX® server, and is accessed via ODBC. The tool administrator can create
more than one directory in a particular system, and users can specify the
directory in which their meta data is stored.

Meta data is held in a directory as objects, which are derived from a hierarchy of
classes. Objects represent both physical and nonphysical components of the
data warehouse. They are cross-linked in a variety of ways, with links being built
from imported meta data and from the events that occur when data warehousing
activities are executed. The object attributes and relationships enable this tool to
provide information about data warehousing processes and assets. Whenever
the content of a directory changes significantly, for example as a result of meta
data being imported, the tool creates a new version of the directory. Versioning
provides a history of directory changes, and lets the tool administrator roll back a
directory to a known point if a problem occurs. An administrator can label
versions, for example to document the change history of a collection of objects
created or modified by particular imports.

MetaStage provides three ways of viewing the structure of objects in a given
directory:
Relationship hierarchy. Most objects in a directory, that represent an
element in the data warehouse with which you can interact directly, have a
variety of potential relationships with other objects. An object can contain
other objects, much as a table contains columns, or can depend on another
object, much as a column depends on a data element. You can track
relationships between objects by browsing this hierarchy of relationships.
Class hierarchy. All the objects in a directory are instances of a class, and
many classes are subclasses of other classes. The class hierarchy clarifies
what components are in a data warehouse. For example, it can show you a
collection of all the computers. The MetaBroker for each external tool has a
different class hierarchy— derived from its data model—that determines its


view of the data. The tool itself also has a class hierarchy that encompasses
most of the classes in the different MetaBrokers. You can view the objects in
the tool directory through the view of the MetaBroker for any tool. The names,
attributes, and possible relationships for each object depend on the view. For
example, a table object seen in a particular product view, becomes a table
definition in the DataStage view and a data collection in the tool view. In this
way attributes and relationships an object has can be seen if you export it to
another tool.
Categories. This tool uses categories to organize objects in a directory. The
category types include:
– User-defined categories, which let you organize objects in a way that
reflects your own view of your data warehousing resources.
– Import categories, which hold objects imported from external tools.
– Publication categories, which contain objects that have been made
available to export to an external tool.
– Business Domain and Glossary categories for holding information on meta
data objects associated with business terms.
– Steward categories for designating that an individual or organization is
responsible for particular meta data objects.
An example of the category browser is depicted in Figure 9-24 on page 492.


Figure 9-24 MetaStage Explorer: Category browser

MetaBrokers
MetaBrokers provide an interface between a particular directory and the external
data warehousing tools. MetaBrokers can also be used with the Ascential tools
DataStage and QualityStage.

There is a different MetaBroker for each external tool, or version of that tool. The
MetaBrokers are implemented as part of a hub and spoke architecture, in which
a MetaBroker and its associated tool are on each spoke and the hub is the
directory holding the integrated meta data. For example, this tool can import
meta data from Tool A and subsequently export it to Tool B in a form that is
meaningful to Tool B. Tool A might be a data warehouse design tool, and Tool B a
data extraction and transformation tool. The MetaBroker lets you view data in a
directory from the perspective of any installed MetaBroker so the data can be
previewed before it is exported to the tool with which that MetaBroker is
associated. MetaStage is supplied with a number of MetaBrokers, and others
can readily be installed to provide interfaces to additional tools. See Figure 9-25
on page 493.


DB2
ERwin Cube Views
4. Import
1. Export OLAP Center
ERwin XML DB2 Cube
Views XML

XML MetaBroker MetaBroker XML
MetaStage

2. Run 3. Run DB2
ERwin Cube Views
MetaBroker MetaBroker

Figure 9-25 MetaBrokers: Meta data flow example

The Publish and Subscribe functions, used with the appropriate MetaBroker,
enable movement of meta data from MetaStage to external tools. They can be
executed from the Explorer, and provide meta data export functions to enable
you to publicize meta data that you want to share with other users. A collection of
meta data objects in a directory can be made available to other users by defining
it as a publication. A user interested in the meta data in the publication creates a
subscription to that publication. The subscription specifies that the user wants to
export the meta data through a specified MetaBroker, and perhaps to be
informed of any changes to the publication.

The meta data does not have to be exported to the same type of tool from which
it was imported. It can be exported to a different tool or to a different version of
the same tool as long as there is a MetaBroker for the target tool. In such cases,
the target tool has its own view of the published meta data.

DataStage users, for example, might do the following:
1. Import a transform from a DataStage project.
2. Publish the transform.
3. Subscribe to the publication on behalf of another DataStage project.
4. Export the transform from this tool into the project.

The administrator allocates rights to importing and exporting, and so has control
over who can share meta data.

MetaArchitect
When you need information from a tool for which a standard MetaBroker does
not exist, MetaArchitect® lets you easily create custom MetaBrokers for
importing and exporting meta data in CSV or XML format. You can create custom
MetaBrokers based on the Common Data Subject Area model, which contains a


full range of OLAP, Entity Relational Model, Relational Data Model, and ETL
meta data classes, or is based on the DataStage 7 MetaBroker model, which
includes ETL and operational meta data classes. An example of a MetaArchitect
is depicted in Figure 9-26.

Figure 9-26 MetaArchitect

Process MetaBroker
MetaStage uses a Process MetaBroker to capture operational meta data. A
Process MetaBroker can be installed on a UNIX or Windows NT® system.

Typically, data warehouse users may run activities to populate part or all of the
data warehouse, or to perform other related tasks such as rollback or
postprocessing. The Process MetaBroker keeps track of an activity, gathers
operational meta data, and delivers that meta data to a particular directory. The
operational meta data includes the times an activity started and ended, its
completion status, and the resources that it accessed, such as files and tables.

To capture meta data, users access the Process MetaBroker via a particular
proxy. Proxies are UNIX shell scripts or DOS batch files that run data
warehousing activities and use the Activity Command Interface to define
operational meta data about those activities.


MetaStage contains a template proxy shell script from which you can build
customized proxies.

Explorer
The Explorer lets you navigate a particular directory, view the attributes of objects
in a directory and, where appropriate, edit the attributes. You can follow the
relationships between objects and run queries to explore those relationships in
more detail. The Explorer also includes the following components:
Category Browser, for organizing meta data and launching imports and
exports.
Class Browser, for viewing class hierarchies and overlap between classes
and the models for different tools.
Directory Search tool, which enables you to find objects in a directory that
match certain criteria.
Query Builder, to build more sophisticated queries to locate meta data in the
directory.
MetaArchitect, is for building custom MetaBrokers.
Object Connector and connected to functionality lets you connect objects that
are essentially identical but which may have been imported from different
tools. For example, a DataStage table definition that was created from an
ERwin table.
The ability to create Data Lineage, Process Analysis, and Impact Analysis
paths in order to inspect and analyze the containment structure,
dependencies, and sources of each object in the directory.
Send to database functionality, which lets you export meta data to a separate
set of tables in the tool directory so you can query these tables using common
SQL query tools.
Online Documentation tool, which lets you create, view, and share detailed
documents describing the attributes and relationships of MetaBroker models
and objects.

If you have administrative rights, functions in the Explorer let you handle such
tasks as configuring the tool options, deleting data from a directory, and defining
the access that users have to the tool functions. You can control the interchange
of meta data among users by restricting access to the tool functions, and by
monitoring user requests to share meta data. You can find examples of Explorer
views in Figure 9-19 on page 486 and Figure 9-24 on page 492.


Managing meta data in the data warehouse
Define data once and use it many times is the basic principle of managing meta
data. Figure 9-27 illustrates a recommended overall flow of meta data among
tools as you build the data warehouse.

Meta Data Validation
Updates Tables

Quality Data Extract, Business
Modeling Transform, Intelligence
Load
Quality Erwin
Manager ER/Studio BusinessObjects
DataStage and Cognos Impromptu
Oracle Designer
PowerDesigner Quality Manager

3. Job
1. Standard 2. Table 4. Table 5. BI Tool
Design and
Data Subscription Meta
Runtime Subscription
Definitions Data
Meta Data

Publish Subscribe Notify
MetaStage
Distribute

To Web

Figure 9-27 Managing meta data

Start with the data model of the databases that will be populated by the ETL
(extract, transform, and load) processes. After importing meta data to this tool,
publish the table definitions from the physical data models and then run
subscriptions to export the table definitions to DataStage.

Load the definitions into the DataStage job stages from the DataStage Manager
to preserve meta data continuity once you import the job designs into this tool.
You can also export the published table definitions to a business intelligence tool
repository, such as a Business Objects Universe or Cognos Impromptu Catalog.
If necessary, you can import Universe and Catalog meta data back into
MetaStage.

And, to finish, you can publish any collection of meta data—data models, table
definitions, impact analysis and data lineage reports— to the Web in a variety of
formats.


10

Chapter 10. SQL query optimizer: A
primer
In this chapter, we discuss a specific Structured Query Language (SQL) and
relational database server program capability. In particular, we discuss the role,
behavior, and design algorithms (the technology) of query optimizers. This is an
important topic, particularly for those who are developing and implementing a
data warehousing environment to support your BI initiatives. Why? Because BI is
heavily dependent on query and analytic applications, typically operating on
huge volumes of data, for supporting data analysis.

We need to understand the optimization process to enable acceptable
performance and response times. And, there are other components and
technologies involved that also must be understood. As examples, the data
warehousing environment and the data models used to define it.

Data warehousing is all about getting your valuable data assets into a data
environment (getting the data in) that is structured and formatted to make it fast
and easy to understand what data is stored there, and to access and use that
data. It is specifically designed to be used for decision support and business
intelligence initiatives. This is vastly different from the more traditional
operational environment that is primarily used to capture data and monitor the
execution of the operational processes.


The emphasis with BI is on getting the data out of the data warehouse and using
it for data analysis. As such, you will typically be working with huge volumes of
data. You can get data out by using stored applications or ad hoc access. In
either case, the technology, and methodology, most used for this purpose is the
query.

Therefore, there is an emphasis on developing queries that can get the data out
as fast as possible. But, having fast queries is dependent on such things as how
and where the data is stored and accessed, which is determined by the data
model. And they are dependent on the capabilities of the relational database
being used, in the form of the query optimizer.

Though the primary focus of this redbook is on dimensional modeling for BI, we
have included information on query optimizers. It is through understanding these
two areas, and enabling their integration, that you can develop a robust business
intelligence environment.

How do you develop such an environment? You start by understanding
dimensional modeling to enable you to create the appropriate data environment.
Then you read this chapter to help you understand query optimizers so you can
make best use of the data environment. Understanding the optimizer technology
is the focus of part one, and in part two, we focus on how to apply it.

In Chapter 11, “Query optimizer applied” on page 599, we discuss the application
of this technology This is to enable you to perform more accurate data modeling
and develop more performant data models. In fact, it can enable you to
determine the performance characteristics of your data model before you even
create it. This will save significant time, and money, in such activities as data
model creation, physical data placement and loading, index creation, and
application development. The objective will be to “get the data model right”
before building on, and investing in, it.

Not only is it important to know the facts and figures of a given technology, in this
case query optimizers, it is also important to know how to work with it. After a
brief introduction, we present a real world query optimizer example. This
example is presented in a narrative style, and we discuss information as it was
discovered during the resolution of an actual customer problem.


10.1 What is a query optimizer
Before we define the term query optimizer, there are other supporting terms that
we need to clarify. The term server is often used in an imprecise manner. It is
used to refer to both a hardware server and software server packages and
systems. Examples of software server packages and systems include relational
database servers, electronic mail servers, print servers, and Web servers. The
IBM pSeries®, formerly known as IBM RS/6000®, is one example of a hardware
server.

Process, memory and disk components
Much like any computer program, a software server has a process component, a
memory component, and a disk component, as depicted in Figure 10-1.

Processes
Shared
Memory
Disk

(Program Memory)
Figure 10-1 Software server architecture

In the case of a relational database (software) server, the components are
defined as follows:
The disk component, or disk architecture, refers to those structures that the
relational database server software maintains on disk. You might initially think
this is the whole point of a relational database server. However, in the case of
a relational database server, we are typically discussing entities such as
dbspaces, tablespaces, chunks, extents, and pages. The SQL hierarchy of
objects, such as SQL tables and SQL indexes, reside inside these disk
component structures.
The term disk component refers to that portion of the server which occupies
or refers to persistent storage. The term disk architecture refers to the design
of that portion of the software server.
The process component, or process architecture, refers to the collection of
computer programs that execute the workload on behalf of the software
server. The two most common relational database process architectures are
a two-process architecture, and a multi-threaded process architecture.
– A relational database server with a two process architecture maintains a
number of persistent daemon processes which generally listen for new

Chapter 10. SQL query optimizer: A primer 499

database connection requests, manage the shared memory structures
(the buffer pool), and so on. The phrase two process refers to user
processes. Each active user has a dedicated back-end server process. If
there are 200 users, then there are 200 dedicated database server
processes (plus the daemon processes). If eight additional users connect
to the server, then you have 208 connections, one connection per user.
IBM DB2 UDB and Oracle 9/10g use a two process architecture.
– A relational database server with a multi-threaded process architecture
also maintains a number of persistent daemon processes for tasks such
as new connections requests and buffer pool management. User
activities, however, are handled by a finite number of constant and
dedicated server processes. Based on the precise database server
process architecture, this number of processes might be related more
towards the number of (hardware) server CPUs than the number of users.
For example, 100 users and four CPUs may be handled by three to four
user server processes. However, 200 users and four CPUs may still be
handled by three to four user server processes. After all, a hardware
server can only execute as many concurrent processes as there are
CPUs.The users are represented by threads, in a pool, that are then
executed on an available user server process. IBM/Informix IDS,
IBM/Informix XPS, and Sybase use multi-threaded process architectures.
– IBM/Informix IDS and IBM/Informix XPS specifically use a non-blocking,
pipelined parallel, multi-threaded process architecture. This means that
the server processes dedicated to user activities do not call for program
interrupts and can execute with 100% (or very nearly 100%) efficiency, in
other words, the non-blocking part. The pipelined part means that one
thread can send work to the next thread to perform activity
asynchronously.
The memory component, or memory architecture, refers to those structures
that the software maintains in memory. Since servers can execute operations
in parallel, this memory is most often shared memory, or, memory that is
available to numerous and concurrent server processes. Shared memory is
the conduit between processes, and between process and disk.


Note: You might wonder why all of this attention to memory and processes
architectures in a data modeling book. The reason is performance. You could
create the most accurate and elegant data model possible, but if the
supported business application system does not perform adequately then the
project may not be successful. One other point to consider is that modern
relational database servers observe read cache ratios in the high 90% range.
This means that 95 to 98 times out of 100 that a data page needs to be read,
the page is already located in a buffer pool and does not need to be retrieved
from disk. Modern parallel database servers, in addition to being disk
management systems, are CPU and data calculation engines.

The relational database server multi-stage back-end
Given all of this, what is a query optimizer? A query optimizer is a subsystem in
the process architecture of a relational database server. Most relational database
servers employ a multi-stage back-end, as shown in Figure 10-2.

New
Connection Listen
Requests Thread

SQL Poll Thread

User Thread

Parser
Query Optimizer
Query Processor
Other Libraries
Figure 10-2 Relational database server, multistage back-end

The relational database server multistage back-end is delineated in Figure 10-2
by the four boxes entitled; Parser, Query Optimizer, Query Processor, and Other
Libraries. While a given relational database vendor may state that they have a
four stage back-end or a seven stage back-end, each database is performing the
same work regardless of how these stages are drawn into various boxes.

The work performed in the multistage back-end is detailed in the following list:


Parser: Typically, the first stage in the back-end is a command parser. Each
inbound communication request received from the user, and each request for
SQL service, must be identified. The inbound command is read character by
character until it is known to be, for example, an SQL CREATE TABLE
request or an SQL SELECT (table) request. This parser may also be
associated with an inbound SQL statement command cache, as a possible
source of performance tuning.
There are two types of SQL commands that are observed by the parser,
those that need to examine existing data records in some table and those that
do not. Examples of SQL statements that do not examine existing data
records include CREATE TABLE, INSERT (new record), and CREATE
SCHEMA. The three most common examples of SQL statements that
examine existing data records are the SELECT (zero or more records),
UPDATE (zero or more records), and DELETE (zero or more records).
SQL statements that need to examine existing data records move on to the
next stage in the multi-stage back-end, the query optimizer. SQL statements
that do not need to examine existing records can skip the query optimizer and
query processor stages. In the diagram depicted in Figure 10-2 on page 501,
this is the area entitled “Other Libraries”.
Query optimizer: Inbound communication statements, requests for SQL
service, that arrive here in the multi-stage back-end are those that are
somehow examining existing records in a single or set of SQL tables. SQL
SELECT statements read data records, and their need to examine existing
records is obvious. SQL UPDATE and SQL DELETE statements also read,
but they are reading before the expected writing, so that they may find the
records they are to UPDATE or DELETE. After UPDATE and DELETE find
(read) the records they are to process, they write.
In each of these cases, the relational database server automatically
determines how to access and locate the records that need to be read. This
means that the invocation and use of the query optimizer is automatic.

Note: The query optimizer is a relational database server subsystem,
residing in the process architecture, that automatically determines the most
optimal method to retrieve and assemble the data records that are required
to process a request for SQL service.

Part of what the query optimizer determines is the (table) access method per
each SQL table involved in the request. That is, should the given table be
read sequentially or via an available index? In a multi-table query, the query
optimizer also determines the table order and table join method to be used.
The product of the query optimizer stage is a query plan. This is the plan for
executing the request for SQL service that is being received and processed.


Query processor: The query processor executes the query plan that was
created by the query optimizer. In some advanced relational database
servers, the query plan is dynamic and may change as the query processor
detects errors in the query plan, or the opportunity for improvement.
The query processor may also make other run-time decisions. For example, if
a given subset of data records is too large to sort in the available memory,
these records may be written to a temporary table which is used as a form of
virtual or extended memory.
While the query optimizer and query processor are two distinct stages in the
multi-stage back-end, you can treat them as one logical entity, which is what
we do in this chapter.
Other Libraries: If nothing else, the remaining stages of the relational
database server multi-stage back-end contain the shared libraries and
software product code. These designs are often proprietary and there is little
value in the current context to offer or review any particular library structure.

10.2 Query optimizer by example
At this point in this chapter, we have only defined the term query optimizer. We
have not discussed much or anything about how a query optimizer behaves. Still,
we are now going to review a real world example involving a query optimizer. In
this real world example, a customer is migrating from a proprietary hardware and
software system and is observing poor (unacceptable) relational database server
performance. Why discuss a technology in an example without first
understanding the technology?
It has been proven that persons retain new information better, and with better
comprehension, when this information is presented in narrative versus
statistical presentation. That is, you learn better by example.
The value of query optimizer knowledge is in the application of this
knowledge, so you need to know how to use this information. Therefore, we
present query optimizer information by example.

10.2.1 Background

Note: This example is real. The business description, use, and intent of the
tables have been changed to protect the privacy of the actual customer. This
example uses three SQL tables. The described purpose of these three tables
may be variably changed in the example text to make specific points.


The real world example we are reviewing happens in the time just before Internet
applications. This company has a toll free call center, where their customers can
place new sales orders via the telephone after reading from a printed catalog to
determine what they want. A human sales agent listens on the phone, interacts
with the customer, accesses various parts of a business application to gather
data for the new sales order form, and generally works to complete and then
place the order. Approximately 70 customer sales agents are active on the
system, 10-12 warehouse workers are using the system to perform order
fulfillment, and two or so additional users are running reports.

The proprietary hardware and software system is being replaced to reduce cost.
A new hardware and software system has been purchased from an application
vendor who has focus on (sells) this type of retail sales application. This
packaged application had been modified to a small extent to meet whatever
specialized needs the given company required. The new hardware and software
system was many times the size of the older proprietary hardware and software
system, with many years of capacity planned for growth.

After months of planning and modification, sales agent training, and so on, a
weekend cutover was used to move data from the old system to the new. On
Monday morning, the new system was fully in use for the first time. After the first
few new sales orders were received, it became painfully obvious that the new
system was performing very poorly. If the customer on the phone asked to buy a
computer peripheral, for example a mouse, the drop down list box offering the
various mice for sale would take minutes to populate with data, where it needed
to be populated in under a second.

After three to four hours of panic, the problem is assumed to be caused by the
new relational database. The relational database software vendor is called to
identify and repair the defect in the database product that is causing this poor
performance.

10.2.2 The environment
The problem observed seems to be one of performance. At this point, it is not
known whether the source of performance is the (hardware) server, the network,
the operating system, the (user) business application itself, the relational
database server, unreasonable user expectations, or some other source. Here
we take the role of the database vendor. Since we have already been called in,
we begin to investigate the database server.


Note: There are two areas of relational database server tuning. One involves
tuning the relational database server proper, its processes, memory and disk,
and the other tuning the user (SQL programming) statements that run inside
that relational database server. Tuning user (SQL programming) statements
could mean tuning the actual SQL commands, or the run time (data model and
indexes) that supports execution of these commands.

Tuning SQL database servers
As mentioned, there is tuning the actual relational database server, and then
there is tuning the user SQL statements that operate within the relational
database server. Tuning the processes, memory, and disk of a relational
database server is a relatively well known procedure. In our example, a database
administrator with six months of experience could validate the performance
tuning of a single (non-clustered, non-distributed) relational database server
system with 500 GB of disk storage and 8 CPUs (just to give you an idea of the
size of the system), in well under two hours of work. Among other things, you
check the following:
The read and write buffer pool cache ratios.
The disk operation request queues, wait times, and high water marks.
Disk layout. Is the disk I/O activity balanced across devices? Are data
segments nice and contiguous or highly fragmented? and other.
Data server events; Is activity being held or delayed for tape backup
operations or related activities?
User status lists; Are users waiting on locks or other resources?
A few sanity checks; Can we connect to the database server and select data
with reasonable performance?
Other procedures, as determined by your relational database vendor.

Managing customer expectations
After basic checking to review adequacy of the relational database server tuning,
there does not appear to be a relational database server tuning problem. The
buffer pool cache is fine, all other observed statistics are fine, yet the user is
getting performance that is unacceptable. In the actual example we reviewed,
there were many barriers that prevented the reviewing engineer from diagnosing
this problem in an efficient manner.

As examples of these barriers, there was a undue amount of tension in the work
area, and the environment discouraged rational thought. Due to these pressures,
the customer was asking for the problem to be immediately solved, and asking
for this to become a training event for his personnel. Even before the source of


the problem had been identified, the customer was looking to assign blame. The
following is a list of actions that may aid in such an environment:
1. Information Technology (IT) engineers sometimes rush off and begin solving
the wrong problem. Repeat the observations and assumptions back to the
customer. For example; “If I can make the performance of this subroutine go
from five minutes to two seconds, would that solve your problem? Is this the
task you would like me to work on first?”
2. Every IT issue has three elements; technical, business, and emotional.
Technical elements are those such as, “What protocol do we use to
communicate between a client application and the server?” Business
elements are those such as, “How much will this solution cost and what is the
expected return on investment?” Emotional elements are those such as, “Bob
is stupid and should not be trusted to solve this problem.”We recommend that
you use these three indexes in communications with the customer to help
quantify and prioritize the tasks that need to be completed. In our actual case,
we chose to focus on the technical challenge of identifying the source of the
performance problem.
3. Work to set reasonable expectations. In this actual case, it is paramount that
we solve the performance problem as quickly as possible. In most cases,
performing training while solving the problem will slow your resource down by
a factor of three of four. So do not use this time for training, that is not the
primary goal. The primary goal at this time is solving the performance
problem quickly.
4. In our actual case, there was no test system on which samples or trial
solutions could be explored. All trial solutions had to be performed on the live,
production system. We would lose credibility if we made the problem worse or
brought the production system down. Advise the customer of the additional
challenges these various conditions cause, and ask for permission to proceed
as the current environment best allows. In our actual case, we created a
second database within the production server so that we could partially
isolate the impact of our investigations from the production system. The
database server software allowed for this option.

10.2.3 Problem identification and decomposition
At this point in the example, we have:
Performed a basic review of the database server tunable parameters and
found no obvious issues. We did this work with little or no prompting because
the task was easy, and is a likely source of error, capable of producing a
quick solution or return.
Worked to start confirming assumptions, asking for and then setting priorities,
and setting customer expectations.


With more attention and detail than before, we ask the customer to demonstrate
the problem. In our actual case, an operation performed inside the business
application is to populate a drop-down list box (a specific graphical control within
a data entry screen form) with data. The customer sales agents logged in to the
database with distinct user names. This made identifying and tracing the specific
SQL commands being executed by a given user and at a specific time, an easy
task. When users log on to the database using a connection concentrator, or
database session pooling mechanism, this task becomes much more difficult.
When pooling is in effect, distinctly named users are recognized by the database
server software by a single or smaller set of shared names. Figure 10-3 displays
some of the information that is discovered at this time.

146 Columns A 54 Columns C
70,000 Rows 637 Rows
1223 Byte Row 507 Byte Row
Length Length
4 Available Indexes 9 Columns
B 4 Available Indexes
- 2 unique 840,000 Rows - 2 unique
- 2 duplicate 94 Byte Row Length - 2 duplicate
1 Extent 7 Available Indexes 1 Extent
- 1 unique
- 6 duplicate
1 Extent

Figure 10-3 Results of investigation at this time

Gather statistics on tables involved in the query
By using the database server software diagnostic utilities, we confirmed that
opening the graphical user application program drop-down list box executes a
three table SQL SELECT statement. We captured the specific SQL SELECT
statement syntax. Then we executed the same SQL SELECT statement on a
non-graphical, non-networked, terminal window and observed the same poor
performance. This was a significant achievement. We removed many variables
from the test case, and captured a repeatable test. Our test produced the same
data in the same amount of time as the users. Add the following:
The actual table names and column names will be obscured from this
example. The actual table and column names are those such as
“dsxggprtst11”. Names that made no sense to a person who has no
experience with this application.


The number of records (rows), columns, indexes, and physical disk space
allocations (extents) is displayed in Figure 10-3 on page 507. This is all data
we were able to discover, given the SQL SELECT statement, the table
names, and other information.
The query returned about 700 records in an average of three to five minutes,
which equals two to three records a second. Given the small table sizes, two
to three records a second is very poor performance.

Note: Page Corners. All relational database vendors have the concept of page
corners, although the specific term used by a vendor for this condition may
vary. All vendors recognize some term which represents the smallest unit of
I/O they will perform when moving data to and from memory and disk. We call
this unit a page. Generally, a vendor will place as many whole data records on
one page as will fit. The data record has a length in bytes, the page has a
length in bytes, and only so many whole records can fit on a page. The vendor
will not unnecessarily split a single record across two pages to make
maximum use of (disk) space. Whatever space is left unused due to this
condition is called the page corner. Why not split data records across pages
unnecessarily? That would subsequently require two physical I/O operations
to retrieve a record when only one could have sufficed. The answer is to
preserve system performance (over disk consumption).

In an SQL table with fixed length records (no variable length columns), the
page corner will be the same on all full pages. We say full pages, because a
page with unoccupied record locations will have additional unused space.

Figure 10-3 shows that table-A has a record length of 1223 bytes. If the page
size for this system is 2 KB (2048 bytes), then 800 or so bytes are left unused
on every data page allocated for that table. If there were 300 or so fewer
commonly used bytes in a table-A record, we could possibly gain some
performance by expelling those bytes to a second table, and observing more
table-A records per page on the more commonly used columns. This physical
data modeling optimization is generally referred to as expelling long strings.

Variable length records have issues too. Some vendors will leave as much free
space on a data page to allow for any single data record on the page to
achieve its maximum length. If the page size is 2 KB, and the maximum length
of a given record is also 2 KB, you may get one data record per page even if
the average variable record length is only 80 bytes.

Reading a text presentation of a query plan
Most relational database software includes graphical and/or character-based
administrative interfaces to retrieve the query plan of a given SQL statement.


Example 10-1 displays the query plan from the current three table SQL SELECT
that is being reviewed.

Example 10-1 Character-based administrative interface showing this query plan
0001
0002 -- This file has been edited for clarity.
0003
0004 QUERY:
0005 ------
0006 select
0007 ta.*, tb.col001, tb.col003, tc.*
0008 from
0009 table_a ta,
0010 table_b tb,
0011 table_c tc
0012 where
0013 ta.col001 = “01”
0014 and
0015 ta.col002 = tb.col001
0016 and
0017 tb.col003 = “555”
0018 and
0019 tb.col003 = tc.col001
0020
0021 OPTIMZER DIRECTIVES FOLLOWED:
0022
0023 OPTIMZER DIRECTIVES NOT FOLLOWED:
0024
0025 Estimated Cost: 407851
0026 Estimated # of Rows Returned: 954
0027 Temporary Files Required For:
0028
0029 1) Abigail.ta: INDEX PATH
0030
0031 (1) Index Keys: col001 col002 (Serial, fragments: ALL)
0032 Lower Index Filter: Abigail.ta.col001 = ‘01’
0033
0034 2) Abigail.tb: INDEX PATH
0035
0036 Filters: Abigail.tb.col003 = ‘555’
0037
0038 (1) Index Keys: col001 col002 col003 (Key-Only)
(Serial, fragments: ALL)
0039 Lower Index Filter: Abigail.ta.col002 = Abigail.tb.col001
0040 NESTED LOOP JOIN
0041
0042 3) Abigail.tc: INDEX PATH
0043


0044 (1) Index Keys: col001 (Serial, fragments: ALL)
0045 Lower Index Filter: Abigail.tb.col003 = Abigail.tc.col001
0047

By reading the query plan displayed in Example 10-1 on page 509, we can see
the following:
Lines 006-0019; The SQL SELECT statement, which we captured earlier, is
repeated.
– table_a.col002 joins with table_b.col001, line 0015.
– table_b.col003 joins with table_c.col001, line 0019.
– All three tables join, A with B, and then B with C. That is good because
there are no outer Cartesian products, a condition where some tables are
not joined.

Note: It is hard to find a real world application of an outer Cartesian product,
since this results in a costly projection of one data set onto another without
relation or care. For example, an outer Cartesian product of the
Valid-Instructors table onto Valid-Courses table returns Instructors with
Courses they do not have any relation to. This is a good way to generate
voluminous data that is of little use. Bob cannot teach the Home Economics
course, because Bob is a Computer Software Instructor.

Outer Cartesian products can exist for valid reasons, but their presence is
extremely, extremely rare. They are so rare, that their presence should be
considered a possible indicator of an error in the data model. And, a relational
database product could show an application error upon receiving an outer
Cartesian product. Outer Cartesian products generally occur between two
tables, and are optimized by creating a third table to act as an intersection. For
example, a Courses-Taught-by-these-Instructors table indicating which
Instructors can teach which Courses, from the example in this block.

– Lines 0029, 0034, and the 0042 give us the table join order; that is, that
table A will be processed first, then table B, then C.
– Tables A and B both have filter criteria. For example, WHERE Ship_date
equals TODAY, WHERE Customer_num = ‘100’.
– The table A filter is displayed on line 0013. The table B filter is displayed
on line 0017. Both filters are based on equalities.
– Filters and/or joins could be based on equalities, ranges, or other set
operands (strings and wildcard meta-characters for one example). A
range criteria would be similar to: WHERE Ship_date > “09/02/1974” AND


Ship_date < “09/09/1974”. Generally, equalities are preferred over ranges
because fewer data records satisfy the criteria, and therefore fewer
records are processed.
Lines 0021-0023; query optimizer directives.
– Be default, the query optimizer uses the algorithms it contains, and certain
information about the tables and resultant data sets involved to form the
query plan. This work is automatic.
– Optionally, the relational database vendor may offer the ability to override
or influence query plans by a technology generally referred to as (query)
optimizer directives, or (query) optimizer hints. Some vendors offer the
ability to promote and/or discourage certain query paths, other vendors
can only promote or discourage, but not both. Optimizer hints are used
during those few occasions when the automatically determined query plan
just does not perform, and a human operator can specify a better plan.
– Generally, optimizer directives are specified by embedding specifically
formatted strings inside the SQL command text.
– The topic of optimizer directives is expanded upon in “Query optimizer
directives (hints)” on page 584.
– No query optimizer directives were in effect for the execution of this query,
as you can see by the absence of observed text between lines 0021-0023.
Lines 0025-0026 display the estimated cost and number of rows (records)
returned. Generally, if the cost goes down while you are tuning a specific
single SQL statement, that is good. It is often less productive to try and
compare costs between differing SQL statements or attempt to calculate
expected run times from this data.
Line 0027; temporary objects.
– Generally, each relational database vendor has two categories of
temporary objects. Here we define these objects to be temporary tables
and temporary files.
– We define temporary tables to be those temporary structures which are
explicitly requested. For example, CREATE TEMPORARY TABLE, or
SELECT ... INTO TEMP TABLE.
– Temporary tables tend to be private in scope, meaning they are viewable
(readable, writable) within the single user session which created the
object. Temporary tables are generally released when the user session is
terminated (closed), but the capability of a specific relational database
vendor may vary here.
– We define temporary files to be those temporary structures which are not
explicitly requested. The most common use of temporary files tends to be
for sorting. For example, a given user session needs to sort 1 GB of data,


but is only allowed to use 5 MB of memory, so data overflow will occur to
some sort of non-memory related structure. That non-memory structure
has to be a disk space allocation. We are calling these allocations
temporary files, although your specific vendor’s name for this resource
may vary.
– Sorting data is generally the cause of temporary files (implicit disk space
allocation used for query processing).
– Sorting inside an SQL statement can occur for many reasons.
Line 0029 lists the first table to be processed in the execution of this query,
table A. Table B is second, line 0034, and table C is last, line 0042. This
condition is referred to as the table order.
Line 0031 details how table A is accessed, and this data is referred to as the
table access method (for table A).
– The two most common table access methods are sequential scan and
indexed retrieval.
– The text “Index keys: col001 col002” means that this table is being
accessed by a SQL index which exists on columns col001, col002 in this
table. This index happens to prevent duplicate key values, as detailed in
Figure 10-3 on page 507, above.
– Database servers can perform operations using multiple concurrent
processes (threads) or not. Non-parallel operations are called serial
operations. Access of this table, table A, is done serially as denoted by the
presence of the “Serial” keyword on line 0031.
– Parallel operations are not automatically faster than serial operations. In
the case of this query, the data sets as so small (as determined by the
small records counts), it is likely this query would not run faster if executed
in parallel. This could however be an area to research in the solution of
this problem.
– Relational database vendors can also physically place records on the hard
disk via a more optimal algorithm than random placement. This specific
vendor can create distinct, pre-organized allocations of data records by
key value and other schemes. This vendor calls these individual physical
allocations fragments. Fragment elimination is a term for the nearly zero
cost ability to reduce the effective physical size of the table, by
determining that given fragments do not satisfy the query criteria by the
rules in effect for the storage of that table. For example, given a list of
world citizens organized by continent where each continent’s list is stored
on a separate disk (fragment), we need only examine one of seven
fragments if we know the records we need reside in the Czech Republic
(Europe). The database server knows these constraints without having to


examine actual data records. The database server knows this by the rules
it maintains.

Note: The term query criteria refers to those conditions in the WHERE
CLAUSE of an SQL statement, both the joins and the filters.

– Line 0013 contains the text, “fragments: ALL”, which means that
organization of data for the table did not make use of this feature, or that
no fragments could be, or needed to be, eliminated based on the query
criteria.
– Indexes contain key columns, the members or columns which reside in
that index. Indexes contain data that is presorted in ASCII collation
sequence, for example, alphabetically or in number sort order. The data in
the table is generally assumed to be in random sequence. Indexes are
always presorted based on an algorithm determined by the index type.
– Indexes may be read in ascending or descending order. Line 0032 states
that this index is to be read in ascending order. That specific point is of
little value.
– Line 0032 also states that this index is being read in order to satisfy a filter
criteria equal to “ta.col001 = ‘01’”.
– From the data in Figure 10-3 on page 507, we know that the SQL table
column ta.col001 is of data type CHAR(02). Generally, relational database
servers join or evaluate strings (character columns) by performing two
byte character comparisons for the length of the string. This operation is
measurably less CPU efficient than say, joining integer values. If ta.col001
can be restricted to containing integer data over character, a degree of
performance could be gained when performing certain operations. Having
to join character data versus integer data adds to the join cost.
Line 0034 begins the description as to how table B is accessed, also a table
access method. This next list of subpoints mentions only new technologies,
those which were not discussed when reviewing table A.
– Line 0038 contains the text, “Key-only”, and this refers to a modifier as to
how the given index on table B is being used.
– Generally, indexes are read to find an index key value (Customer_num =
‘101’, for example). This single entry in the index data structure also
contains the physical address on disk of the associated full data record in
the SQL table proper, (where the remaining columns of this full record
reside for Customer_num 101). In most cases, the index is read to find the
address of the remainder of the data record, and then the full data record
is read.


– The term key-only means the query processor did not need to perform this
work. All of the columns needed for the given task were members of the
given index, the query processor does not need to go to the table proper to
get (other) columns.
– Obviously key-only (index) retrievals are more efficient than standard
index retrievals, as less work needs to be performed. The work in question
is a reduced amount of physical or logical disk I/O.
Line 0040 contains the text “NESTED LOOP JOIN”.
– Thus far we have discussed table access methods and said there is a
sequential scan disk access method, and an indexed table access
method. There are also other table access methods, discussed in “All
table access methods” on page 537.
– A table access method details how the given records contained within one
SQL table are processed.
– Line 0040 displays the table join method. A table join method details how
records from one table are then matched with their counterparts in a
second, or subsequent table.
– From our current example, and in a nested loop (table) join method, table
A is read via an indexed table access method. Although it could have been
indexed or sequential, it does not impact the table join method in this
context. As each row is read via the index on table A, the filter criteria is
applied (line 0013). As each row passes this filter (test), the row is then
matched to any and all rows that match in table B, based on the join
criteria line 0015, however table B is being accessed.
– The nested loop table join method gets its name from the software
language programming construct which is used to create this program
code, namely a recursive (nested) loop.
– Table A is joined to table B using a nested loop join method, as stated on
line 0040. Table B is joined to table C also using a nested loop join
method, as stated on line 0046.
– Generally, a given relational database vendor has three or more table join
methods. Nested loop join method is the oldest, and under the proper
conditions, the fastest. Sometimes nested loop is not the fastest, which is
documented later in “All table join methods” on page 541.
Line 0042 begins the listing for table C. No new technology is displayed after
line 0042.
To summarize, here is the primary data we need to track:
– The tables are processed in the order of table A, then B, then C.


– Tables A and C are accessed (table access method) via a unique index
scan. That is, records in tables A and C are located via a unique index and
an equality in the SQL SELECT WHERE clause.
– Table B is also accessed via an index, although that index is an index that
permits duplicate values. As a modifier, the indexed retrieval into table B is
key-only. A key-only retrieval is one where all of the necessary columns
(all of the columns we are being asked to read) are members (elements,
attributes) of the given index. Normally, an index retrieval accesses the
key columns of an index, in order that we may locate the full width data
record in the table proper. A key-only retrieval has all of the columns it
needs in the index itself, thus avoiding a table lookup.

Create a query plan diagram
Figure 10-4 provides a graphical depiction of the query plan for this example.
Only the most important elements of this query are displayed. It would be an
important skill to be able to draw a diagram of this type from your specific
relational database server query plan display.

The query plan followed the table order, Tables “a”, “b”, then “c”.
“c”.

C

A FILTER, a.col1 = LITERAL

col1 CHAR(10)
col1 CHAR(02) 1
col2 CHAR(30)
4 Unique Index
(col1)
Unique Index
(col1,col2) B
col1 CHAR(30) JOIN, b.col3 = c.col1
2 col2 SMALLINT
JOIN, a.col2 = b.col1 col3 CHAR(10) 3
Duplicate Index FILTER, b.col3 = LITERAL
(col1,col2,col3)

Figure 10-4 Graphical depiction of query plan

Thus far we have determined:
The query has three tables of small size. (Finding the query through
administrative tools, reading the tables’ sizes, specific structure, and disk
space allocations through administrative tools.)


All three tables are joined. (Reviewing the syntax of the SQL SELECT.)
All three tables are accessed using an indexed retrieval. (Reviewing the
query plan, specifically the table access methods.)

Note: We could check to ensure that these indexes are not corrupt, by a
number of means. Generally, corrupt indexes do not occur in modern
relational database server systems. Most commonly, corrupt indexes give
incorrect results, or never return from processing. Corrupt indexes do not
display poor performance and correct results. One easy way to check to see
whether or not an index is corrupt is to perform a specific query using a
sequential (non-indexed) retrieval. Then compare these results to those using
an indexed retrieval.

We see that there are not temporary files or tables being used, which can
delay query processing. Absence of these structures observed from the query
plan.
We have executed the query in a simple manner (from a character-based
terminal program). We have observed the server returns on average 700
rows in three to five minutes, in other words, two rows per second, on
average, from very small tables.
We have already eliminated the likelihood that slow network traffic is to
blame. If there were remote SQL tables, we would copy all of the tables to a
single, local database server and repeat the perform tests. (The tables were
small and we have available disk.) Also, the test was, as stated above, run
from the hardware server console.
We have already eliminated the likelihood there is multiuser delay due to
locks, lock waits, or dead locks, by viewing the database server
administrative status reports.
We still think we can do better.

Note: When you have something that is not working, remove conditions from
the test case until something changes. In our case, we have a three table SQL
SELECT that is not performing adequately. The third table, table C, provides
little value. For experimentation only, remove table C from the query and see
what happens.

10.2.4 Experiment with the problem
At this point we are going to experiment with conditions in the run-time
environment, specifically those called for in the syntax of the SQL SELECT
statement. Because this is an operational and production system, and because


these tables are small (we have the available disk space), we create a copy of
these three tables somewhere outside the current database. (The server product
allowed for numerous and concurrently operating databases inside one software
server. Before proceeding, we re-execute the test case (the SQL SELECT) and
compare data results, query plans, and execution times to help indicate that we
copied this example run-time environment correctly.

Drop criteria from the query
The current problem is that a small three table SQL SELECT seems to be
performing below expectations. We do not know what is causing the problem, so
we are going to change the SQL SELECT statement and see what happens.
Table C is joined to the products of table A and B, via an equality on a single
column that has a unique index placed upon it. Table C is processed last and has
only 600-700 records. As a result, table C will be removed from the query for the
purpose of research. If we cannot get the two table SQL SELECT of just tables A
and B to perform adequately, then we have little hope of making the original
three table SQL SELECT perform.

Figure 10-4 on page 515 shows the query plan as it currently executes, and with
the final criteria it needs to logically contain. Figure 10-5 displays the query plan
after table C was dropped for investigatory purposes.

The query plan followed the table order, Tables “a”, “b” then “c”.
“c”

C


col1 CHAR(10)
col1 CHAR(02) 1
col2 CHAR(30)
4 Unique Index
(col1)
Unique Index
(col1,col2) B
2 col2 SMALLINT
(col1,col2,col3)

Figure 10-5 Altered query, just tables A and B

As expected, the following was observed:


Table C is truly low cost. The total query execution time went from an average
of five minutes, to four minutes and 40-50 seconds. This is not a significant
change, considering the customer needs sub-second performance.
The table order remained table A then table B. This may also be useful. If the
table order changed to table B first then table A, and also produced an
acceptable performance result, then we may have trouble recreating this
behavior when we reattach table C to the query. One could join table B to
table A, then join this product to table C. With several (ten or more) tables and
several discontiguous table join orders, this may encourage the implicit use of
temporary files which would negatively impact performance.
What we have now is a two table SQL SELECT that does not perform
adequately, perhaps a simpler problem to research.

Measure filter selectivity
With the more simple two table SQL SELECT, we are going to measure the
tables separately and then measure them together. From the data presented in
Figure 10-3 on page 507, we already have some knowledge of table A.
Figure 10-6 displays what is learned when we measure the selectivity of the filter
on table A.

FILTER, a.col1 = LITERAL
A

col1 CHAR(02) 1
col2 CHAR(30)
98%
Unique Index
(col1,col2)

Figure 10-6 Measuring just the filter on table A

Table A is listed as having 70,000 rows. This can be confirmed by many means
including an SQL SELECT similar to:
SELECT COUNT(*) FROM table_a;

Figure 10-6 displays what we learned from executing a SQL SELECT equal to:
SELECT COUNT(*) FROM table_a WHERE col001 = ‘01’;

Basically, we just copy the predicates (filters) which are in effect on table A, and
change the column list to perform a count. This query informs us that the single


filter on table A is not very selective, returning on average 98% of the entire
contents of table A.

Index utilization guidelines
As we mentioned previously, part of what the query optimizer determines is table
order. In this example, should the query processor read table A first, then B, then
C, or perhaps tables C, then B, then A, or another choice? Part of what the query
optimizer considers when determining table order is selectivity of the filters.
Table order is affected by many things. However, in the area of just selectivity of
filters, read the following:
Here we are using the example of a telephone book, a sorted listing of
persons. It contains last name, first name, gender, their telephone number,
and perhaps other data.
One can sequentially read the telephone book, page after page, to find the
entry needed. This is a sequential scan, one of the table access methods.

Note: In this context, index utilization guidelines does not refer to when the
data modeler should use (create) indexes. Index utilization guidelines refer to
the conditions that allow for or prevent the query optimizer from using indexes.

Data in the phone book is sorted, and for this example we will imagine this
condition of having the presorted data to be our index. Most phone books are
sorted by last name, then first name, and so on. For this example, we will
state that the phone book can only be sorted on one column of data.
If the phone book is sorted on last name, but not any other column (because
we are imposing a single column sort order limit), we would locate a male
named “Bob Jones” in a manner similar to:
– This phone book is many hundreds of pages thick. Knowing that the last
name is Jones, and that this phone book is sorted by last name, we open
the phone book somewhere in the middle. We know the letter J is close to
the middle of the alphabet.
– Paging not by single pages, but by dozens of pages we locate the start of
the J page. Even then, the last name starts with “Jo” (Jones), not “Ja”, or
“Jb”, so we still page many pages at a time to find the listing for all persons
with a last name of Jones.
– This phone book is sorted only by last name, so while we are at the start of
a list of male and female persons with the last name of Jones, the name
we seek could be last in this list. This list is not then subsequently sorted
by first name or by gender. We perform a sequential scan within Jones to
find Bob, and then Bob Jones, a male.


– In an actual SQL database, we have used the index to position at the start
of the list of all persons named Jones, then we had to go to the SQL table
proper to read each first name and gender. In this example, the first name
and gender were not members of the given index.
– While this phone book contained 1,000,000 persons total, there were only
10,000 persons with a last name of Jones. By using the last name index,
we avoided some amount of workload between sequentially scanning
1,000,000 records, or performing an indexed read of 10,000 index key
columns, and then looking up the remaining (non-key) columns in the
table proper by their absolute disk address.
Now let us change the example to one with a different selectivity.
– Let us say that the phone book is not organized by last name but is
organized by gender (male or female). We know Bob is a male so we can
position at the start of the list of males, then go to each record sequentially
and see if the name is Bob Jones.
– This option would give us measurably more work to perform. As a filter,
the gender code is measurably less selective than last name. The amount
of physical or logical disk I/O to perform on a list that is sorted (indexed) by
gender, a less selective filter, is higher than a list that is sorted (indexed)
by last name.
– We need to tie selectivity of filters and their effect on table order with other
concepts which are not defined here. This work will be done shortly. At this
time we can state that for now, all things being equal, preference is given
to the filter with the greatest, most unique, selectivity.
If you wish to state that the phone book example is too simplistic, or too rigid,
to be of real value, consider the following:
– Why not just create a phone book which is sorted by last name, then first
name, then gender (this is called a concatenated, or multi-column index)?
Concatenated indexes have issues too. Consider a phone book that is
sorted by last name, then first name, then whatever else. To find a person
in that phone book whom you only know has a first name of Bob, (you
don’t know the last name), you lose the index and have to perform a
sequential scan.
– In short, concatenated indexes do not solve every problem. Concatenated
indexes have limits equal to or greater than non-concatenated (single
column) indexes.

As stated in the note box above, the subsection title “index utilization guidelines”
does not refer to when you should create indexes. It refers to when the query
optimizer can make use of indexes.


Note: The specific condition that was just discussed, (finding a person in a
phone book by first name when the phone book was sorted (indexed) by last
name, then first name), is called negation of index, non-anchored
composite key. The last name, then first name index can not serve a first
name only filter criteria because this index is not headed (anchored) by first
name. Imagine trying to find all first names of Bob in a phone book. You would
have to sequentially scan the entire book. A negation of an index means that
an otherwise usable index is disallowed due to a condition.

Another negation of an index is negation of index, non-initial substring.
That is, given a phone book sorted by last name try finding everyone in that list
where the last name ends with “son”, such as in Hanson, Johnson,
Williamson, or Ballardson. To perform this search, you would have to
sequentially read the entire phone book. The index (in this case sort) column
is not organized by the last three characters of the last name. It is anchored by
the leading character of this string. When you lose the leading portion of an
indexed string, you negate the effectiveness (use) of the index.

This is another reason not to violate the first rule of third normal form, which is
do not concatenate related columns. If the last three characters of this column
have meaning, then it is a separate entity attribute (column) and should be
modeled as such. As a separate column, it can be indexed. (To be fair, some
relational database vendors allow you to index a non-initial substring of a
given column. Still, this feature often needs to reference a static range. For
example, index character positions five through eight, for example; not the last
three characters of a variable length data value.)

B-tree+ indexes defined
We have been discussing index use, without really defining what indexes truly
are or how they function. There are several types of indexes, and thus far we
have been discussing a type of index called a b-tree+ index without bothering to
label it as such. Figure 10-7 on page 522 provides a graphical depiction of a
b-tree+ style index.


--
505
490
Index 487 89
Retrieval -- 485
485
476 300
436 4
-- Sequential
505
-- Scan
378 292
485 300 156
376 --
292 -- 436
-- 180 292 476
-- 209
-- -- 143
89 -- 180 21
89 156
55 143 ...
21 --
--
ROOT 89
70
LEVEL 2 LEVEL 1 LEVEL 0

Figure 10-7 Graphical depiction of a b-tree+ type index

B-tree+ style indexes resemble a tree, and represent a standard hierarchy. All
access of a b-tree+ style index begins at the root level. B-trees use the analogy of
natural wooden trees and have terms such as root, branch, twig, and leaf
(nodes), as in the following list:
Given a list of numbers in the range of 1 to 1,000,000, the root would contain
as many entries as can fit on the (index data) page. The root node would say,
for example, rows with a key value of 1 to 100,000 are located next at some
address, rows 100,001 to 200,000 are located at some other address, and so
on.
Generally any entry in a b-tree+ index contains a single set of column values
which are the keys, and an absolute disk address of the structure to which
this entry in the index refers.
Prior to the final (leaf) level of a b-tree+, entries point to further deepening
entries in the tree; roots, then branches, then twigs, and then leaves.
By traversing a b-tree+ index, we eventually arrive at the leaf node, which
then points to the address of a row in the SQL table proper.
The SQL table proper contains all columns found to exist within the table, not
just the non-indexes column. In this manner, the table can also be
sequentially scanned to arrive at any row or column.


As any single page in the index structure becomes full, it cannot accept
additional entries. At the time of this event, the page will split leaving two
pages where one existed previously. Consider the following:
– By default, a single full page may split into two equally (half) full pages,
leaving room for growth on both.
– If however, you are only adding new records at the end of a range of data
(you are only adding increasingly larger numbers), you would be leaving a
wake of half full and less efficient index data leaf pages.
– All modern relational database servers are intelligent enough to detect
index data leaf pages that are split in the middle or at the end of a range of
data, and to behave optimally. This specific capability is referred to as
having a balanced b-tree+.
Because of the cost of having to perform logical or physical disk I/O on the
b-tree+ with its many levels, and then still having to go to the SQL table
proper to retrieve the remaining columns from the table, the query optimizer
considers the selectivity of the index when deciding whether or not to use the
index or to discard the index use going instead with a sequential table scan.
As a very rough guideline, if the query optimizer believes that the query
criteria (join or filter) of an indexed column will return one-third or more of the
contents of a table, that index is not used in favor of a sequential scan. This is
a working guideline only.
The “+” in b-tree+ refers to a capability that the index can be read in both
directions, both ascending and descending.
– B-tree+ index is either sorted in ascending or descending order at the leaf
level.
– Many years ago, an ascending index could only be read in ascending
order, not descending. Descending indexes could read in descending
order, not ascending.
– Index data leaf pages contain a pointer (address) listing the next leaf page
in the contiguous chain of index data.
– Many years ago, these index data leaf pages only contained a pointer to
the next leaf page. Now they contain a pointer to the next and previous
index data leaf page. This capability adds the “+” to the b-tree+ index
name.

Note: The specific condition that was just discussed, (not using an index
based on its poor selectivity relative to the query criteria), is called negation
of an index, (poor) selectivity of filter. An otherwise usable index may not
be used because it would be more efficient to sequentially scan the SQL table
proper.


Data distributions
We have seen that the table access method for table A is indexed retrieval. And
we have seen that the selectivity of that filter is poor, meaning that better
performance would be observed if the index were not used in favor of a
sequential scan. We can force a sequential scan of table A using (query)
optimizer hints, but that still will not solve our overall problem. We seek to get the
total execution time of this query from three to five minutes to sub-second. A
sequential scan of table A will save some time, it will not satisfy our objective.

Why did the query optimizer make this error? The query optimizer uses general
rules, and observed statistics about the involved tables to determine the query
plan. The query optimizer generally gathers these table statistics on command,
not automatically. Gathering these statistics can be costly. While statistics had
been gathered prior to executing the query in our test case, we were gathering
the minimal amount of statistics. Because of the cost of gathering statistics, most
relational database vendors will gather increasingly complete levels of statistics.
Figure 10-8 provides a representation of data by distinct value and the count of
records per value.

Count
250
200
150
100
50
0
SP SC MG RJ CE RS
Data Value
Figure 10-8 A distribution of data by value and count

From the diagram in Figure 10-8, we see that the data value “CE” carries the
majority of data by count, while the data value “SP” carries little or no data by
count. Refer to the following:
If we were to evaluate a filter criteria equal to “WHERE value = ‘CE’”, that
application is not very selective. If we were to evaluate a filter criteria equal to
“WHERE value = ‘SP’”, that application is very selective.


When query optimizer statistics are gathered at the default level in this
vendor’s relational database server product, it gathers, among other things,
the following:
– The number of records in a table.
– The number of unique values in each index.
– The maximum, second maximum, minimum, and second minimum value
in each index key value range.
Under default conditions, the vendor would know the total number of records,
and the number of distinct values. In this case, six.
Under default conditions, most relational database server vendors assume an
even distribution of data. Meaning from our example, we would expect that
each of the values in Figure 10-8 on page 524 carries 1/6 of the total number
of records on the table. Assuming an even distribution of records is fine if the
records are evenly distributed.
Data distributions is a relational database server capability where the data
is actually sampled to determine the spread (selectivity) of values across a
range. Most vendors allow the administrator to determine the percentage of
data records that are read to gather a distribution.
In this example, the server could know in advance to use a sequential scan
table access method when receiving a “WHERE value = ‘CE’”, and possibly
to use an indexed retrieval when receiving WHERE value = ‘SP’”, or at least,
given the choice of which filter to execute first, which is more selective.
Data distribution can be gathered on any column, whether or not it is an
element of an index.

Query optimizer statistics
In the three table SQL SELECT example, table A is accessed via a unique index.
That is, one that does not permit duplicate values. Table A contains 70,000
records, so this index contains 70,000 unique values. The entire index is unique,
yet the SQL SELECT, counting the filter criteria displayed in Figure 10-6 on
page 518, returns 98% duplicate values on an equality. How is this possible? The
filter criteria for this SQL SELECT accesses a subset of this composite index,
only the leading single column of a two column index. Column one of this index is
98% duplicate. The entire width (both columns) of this index is unique.

This vendor default query optimizer statistics gathering behavior looks only at the
uniqueness of the entire width of a composite index. Some vendors look at as
many as six column deep into a composite index by default. In either case, a
query optimizer statistics gathering that examines the data distribution of the
leading column of the index solves the problem of whether to use an index or go
sequential on the current query plan for this query.


Note: There are many specific applications of data distributions, to include:
To examine the uniqueness of a subset of a composite (multi-column)
index, greater than what the default query optimizer statistics will gather.
To examine the uniqueness of columns used in filters and joins. For joins,
this uniqueness is a large indicator of the cardinality observed between the
two joining tables. These columns may, or may not, be indexed.

Measuring the cardinality between tables that are joined
Currently we know that the table access method being used for table A could use
improvement. Still, that correction alone will not solve the overall problem.
Therefore, we continue to perform additional investigation. We could either
measure the selectivity of each of the remaining filters, or simply follow the table
join order as the SQL SELECT would be executed by the query processor. Both
activities are valid in either order. Since the join may also restrict record counts
(there may not be a matching record found to exist in table B that matches with
table A), we will follow the query plan table join order.

To measure the cardinality between two tables, execute a SQL SELECT in the
manner shown in Figure 10-9.

SELECT a.col1, COUNT(*) ... GROUP BY 1


col1 CHAR(02) 1
col2 CHAR(30)
2 JOIN, a.col2 = b.col1
Unique Index
(col1,col2)
B

(An actual case) P-Key Count
Key 1 32
Key 2 111
Key 3 9
Key 4 3201
...

Figure 10-9 Query to measure cardinality between two SQL tables


Given states or provinces within a given country, the SQL SELECT displayed in
Figure 10-9 on page 526 would return the state/province primary key, and a
count of cities in that state/province. This is a basic construct to measure
cardinality between two tables.

In our current example, table A contains 70,000 records, and table B contains
840,000 records. This record counts are accurate from the real world example. In
the real world example, each table A join column value would match with
somewhere between 8 and 4000 records from table B, producing an
intermediate product set, the product of joining table A to table B, of millions and
millions of records. If millions and millions of records are produced from this join,
and the final query returns only 700 records, where do these records go? Where
are they discarded? Figure 10-10 answers that question.

SELECT a.col1, COUNT(*) ... GROUP BY 1


col1 CHAR(02) 1
col2 CHAR(30)
2 JOIN, a.col2 = b.col1
Unique Index
(col1,col2)
B

FILTER, b.col3 = LITERAL 2%

Figure 10-10 Count of both tables with all filters

Figure 10-10 displays the two table query, with tables A and B joined, and with all
filters applying to tables A and B being present. The filter on table B is very
selective and reduces the number of records produced when joining tables A and
B to a final result size of 700 records. The filter on table B is definitely the filter
that should be executed first, before table A. Beginning the query processing with
a table order of table B before table A, reduces the number of total records
processed by several orders of magnitude.

Note: It is important to understand that the query optimizer made the optimal
query plan given the conditions that exist in the run-time environment. The
primary fault of the query execution time is related to the data model.


Before moving on to a solution, Figure 10-11 displays an interesting behavior to
this query that was not previously discussed.

LOOP
250 FETCH DATA
END LOOP
200
150 I/O Reads
100
Iterations
50
0
Seconds
Figure 10-11 Disk I/O versus record retrieval graph

Explaining misleading server tuning behavior
As part of the earlier investigations, we had examined disk I/O rates and queues.
The system in this real world example had eight (count) SCSI type III hard disks.
All three of the tables in this example were located on one disk. We could have
achieved a small amount of improvement in performance by spreading these
tables across several disk drives or controllers.

The database server software was reporting a very consistent rate of 150
physical disk I/O operations per second, which is average for a SCSI type III
drive. What was odd was another behavior. We created a ten line computer
program to execute a SQL FETCH loop for this three table SQL SELECT. This
SQL FETCH loop would return a few records and then pause for many seconds
or sub-seconds; the loop returned one record at a time as SQL FETCH loops do.
Until the data gathered in Figure 10-10 on page 527, we were at a loss to explain
this stop and start SQL FETCH loop behavior.

The query was executing at a constant and fast rate. The delay in query
processing (as witnessed by the delays in the iterations of the SQL FETCH loop),
was evidence of the number of records examined, filtered, and joined between
tables A and B, only to be discarded by the filter on table B. We are processing
hundreds of thousands of records only to produce 700. This is our major
optimization that will reduce the total query execution time from minutes to
sub-second.


10.3 Query optimizer example solution
Here we present the solution to the query optimizer example that began in 10.2,
“Query optimizer by example” on page 503. Figure 10-12 once again displays the
original query plan for this example.

The query plan followed the table order, Tables “a”, “b”, then “c”.
“c”.

C


col1 CHAR(10)
col1 CHAR(02) 1
col2 CHAR(30)
4 Unique Index
(col1)
Unique Index
(col1,col2) B
2 col2 SMALLINT
(col1,col2,col3)

Figure 10-12 Original query plan from example

It is the filter on table B which should be executed first based on the selectivity of
that filter. This same filter greatly reduces the number of records that need to be
joined from table B into tables A and C. The query optimizer is prevented from
taking this query plan because of the run-time environment.

Note: The query optimizer will do nearly anything so that it may complete your
request for SQL service. One thing the query optimizer will not do is
sequentially scan two or more tables inside one SQL SELECT, nor will it
sequentially scan the second or subsequent tables in the table order.
Sequentially scanning the second or subsequent tables in a SQL SELECT
means that table, or tables, would be sequentially scanned numerous
(repetitive, perhaps hundreds of thousands or millions of) times. That is
generally not happening, because performance would be grossly
unsatisfactory.


The join criteria from table B to table A is on table A col003. Table A col003 is not
indexed. If the table join order in this query were table B, then table A, then each
of 700 records in table B would cause a sequential scan in table A. In order for
the query optimizer to consider a table B then table A table join order, the join to
table A would have to be served via an indexed retrieval. Currently, there is no
index on table A that supports this behavior, and that is why the query optimizer
will not choose table B first.

10.3.1 The total solution
The total solution to the three table SQL SELECT example begun in 10.2, “Query
optimizer by example” on page 503 contains the following:
The table A, concatenated index on col001, col002 is next to useless. This
index is removed and replaced with an index on table A col002, then col001.
This new index preserves the unique data integrity constraint that this index
serves.
Because of the nearly total lack of unique values in col001 of this index, we
have little fear in removing this index. Any other portion of the system that
was using this index for processing will run faster without this index.
This concatenated index leads with col002 to support the join from table B.
This correction is displayed in Figure 10-13.

CREATE UNIQUE INDEX Becomes CREATE UNIQUE INDEX
(indexName)
indexName) (indexName)
indexName)
ON tableA ON tableA
(col1,col2); (col2,col1);

Figure 10-13 Correction to table A

The table B index on col001, col002, col003 is also reversed to contain
col003, col002, then col001, as displayed in Figure 10-14 on page 531.


CREATE UNIQUE INDEX Becomes CREATE UNIQUE INDEX
(indexName)
indexName) (indexName)
indexName)
ON tableB ON tableB
(col1,col2,col3); (col3,col2,col1);

Figure 10-14 Correction to table B

There was never anything wrong with table C.
The new query plan is displayed in Figure 10-15.

The query plan followed the table order, Tables “b”, “c”, and then “a”.
then
SELECT * WHERE ...
B

col1 CHAR(30)
FILTER, a.col1 = LITERAL col2 SMALLINT JOIN, b.col3 = c.col1
col3 CHAR(10)
C
A
Duplicate Index
2
4 3
(col3,col2,col1)
1
col1 CHAR(02) col1 CHAR(10)
col2 CHAR(30) JOIN, b.col1 = a.col2 Unique Index
Unique Index (col1)
(col2,col1) FILTER, b.col3 = LITERAL

Figure 10-15 Optimal query plan for this example

The old query plan processes hundreds of thousands of records as table A is
joined with B, and then the restrictive filter is applied. The new query plan begins
by reading table B, retrieving only 700 records which are then joined with tables
A and C. The new query plan is detailed below:
Table B is processed first using a duplicate index, but in a key-only retrieval.
Then the join is done to table A via the unique index on table A.
The join to table C is done as before, via a unique index via an equality.


There are still no temporary files.
The query execution time completes in under one second.

Before releasing this solution to the customer, we must check both data sets.
The one that was produced by the current (under performant) production system,
and the one which is created by the (performant) test system. Because the two
data set sizes were manageable, we unloaded the results to ASCII text files and
used operating system utilities to compare these results and prove them to be
identical.

10.3.2 Additional comments
Further comments on this example are contained in the list below:
Upon critical examination, we see that the index on table B contains three
columns; col001, col002, and col003. Col002 is never mentioned in the
example. Although col002 is never referenced, it was left as a member in the
new solution index in the event that other portions of the application system
needed this column in this index.
The original index on table A could definitely be removed; it was useless.
You should generally lead a concatenated index with the most unique, most
selective columns.
The original index on table B may have been left in place in the event some
other portion of the system needs it. A new second and corrected index could
have been added. But, indexes are not free resources. In addition to the disk
space they consume, they must be maintained. Here are a few examples, for
instance:
– An index is like a mini, vertical slice of a table. It contains a subset of
columns and all rows.
– When a table has (n) indexes, a single new record insertion must be
recorded in (n+1) places. This event must be recorded the SQL table
proper, and then in each index. If a table has four indexes, there is the
SQL table proper and four vertical slices of that table, the indexes, that
must receive the new record insert.
– This (n+1) guideline is also true for record deletes.
– Record updates affect the table proper, and then any indexes containing
columns that are changed.
In many of the program code listings and diagrams in this example, the
column list of the SQL SELECT statement lists all columns. Retrieving all
columns from table B would have prevented a key-only retrieval, since all
columns from that table are not members of the index being used.


Table B has nine total columns and seven indexes. The original data modeler
had violated the first rule of third normal form entity relationship modeling,
because there was a repeating group of columns that were folded into table
B. These columns should be exported as elements of a fourth table. The
reason for so many indexes was to overcome index negation, non-anchored
composite key:
– An easy example to describe what was taking place would be an order
header table, and a detail table to the order header with the various order
line items. The order header lists the unique attributes of order date,
customer number, and others. The order line items lists that dynamic
count of items which may appear on the order, such as the stock unit
identifier, the item count, and cost.
– This modeler was new to relational database and SQL and could not get
over the cost of having to join tables, compared to older database
technologies which model data differently.
– To reduce the number of join operations, this modeler folded the order line
item (child) table into the order header (parent) table. This not only limits
the number of line items that can be on one order (the column list is now
hard coded as a repeating set of columns), it causes a number of
performance issues.The most obvious performance issue raised is
non-anchored composite key.
– When order line items appear as a repeating set of columns in the order
header table, it becomes very hard to answer the question “which orders
contain item number ‘444’, for example. This data could be in any of a
repeating set of columns. Not only do you now require a larger number of
indexes to support that query, the SQL SELECT syntax is also unduly
complex.
Why was the leading column in the table A index on col001, col002 a
duplicate:
– As mentioned, this application (and its associated data model) was sold to
the customer. This is a packaged application and was sold door to door, to
dozens of consumers.
– Basically, the table A column col001 stored the activity location identifier,
similar in use to a warehouse code. This customer only has one
warehouse.
– We need not have changed the columns in the table for this one customer,
but we could have changed the index plan. Changing columns will likely
cause the application program code to change. No one would want this,
since it increases cost and allows for divergent code streams in the
application. Changing the index plan does not call for the application
program code to change.


– For those who are big on maintaining separate logical and physical data
models, perhaps the index plan should always be part of the physical data
model; at least when a given data model is delivered to numerous physical
implementations.

10.4 Query optimizer example solution update
At the beginning of 10.2.1, “Background” on page 503, we stated that this real
world example takes place several years ago. Since that time, relational
database servers have increased their program capabilities. The current release
of the database server software product referenced in this example would likely
still not serve this three table SQL SELECT example in sub-second time, given
the errors in the run-time environment. Read the following:
We define the run-time environment as, the SQL SELECT, the available
indexes, the version of relational database software, and other conditions.
The original run-time environment with its older software executes the query
in five minutes, on average. (Remember the query needs to execute in
sub-second time.)
Just updating the version of relational database software improves execution
time of this query from five minutes to five or 10 seconds.
– While the query execution time is greatly improved, it is still not
sub-second as required.
– The improvement in execution time is accomplished by a better query
plan. This better query plan is still not as good as that provided in the
solution in “Query optimizer example solution” on page 529.
– The new query plan uses a lot more memory and CPU resource to
accomplish the improved execution time.
We can use query optimizer hints to force the updated relational database
software to execute the original (least optimal) query plan. Purely as a result
of the newer and improved software, query execution time moves from five
minutes to 10 or 20 seconds. In other words, just updating the software allows
for most of the performance gain, not the new query plan.

The current release of this product would move execution time from five minutes
to somewhere around five or 10 seconds using new technology. The modern
query plan, from the same run-time environment is listed in Example 10-2 below:

Example 10-2 Query plan of same three table SQL SELECT and run-time environment,
with newer release of database server software
0000


0002
0003 QUERY:
0004 ------
0005 select
0006 ta.*, tb.col001, tb.col002, tb.col003, tc.*
0007 from
0008 table_a ta,
0009 table_b tb,
0010 table_c tc
0011 where
0012 ta.col001 = “01”
0013 and
0014 ta.col002 = tb.col001
0015 and
0016 tb.col003 = “555”
0017 and
0018 tb.col003 = tc.col001
0019
0020 DIRECTIVES FOLLOWED:
0021
0022 DIRECTIVES NOT FOLLOWED:
0023
0026 Temporary Files Required For:
0027
0028 1) Abigail.tb: INDEX PATH
0029
0030 Filters: Abigail.tb.col003 = ‘555’
0031
0032 (1) Index Keys: col001 col002 col003
(Key-Only) (Serial, fragments: ALL)
0033
0034 2) Abigail.ta: SEQUENTIAL SCAN
0035
0036 Filters:
0037 Table Scan Filters: Abigail.ta.col001 = ‘01’
0038
0039
0040 DYNAMIC HASH JOIN
0041 Dynamic Hash Filters: Abigail.ta.col002 = Abigail.tb.col001
0042
0043 3) Abigail.tc: INDEX PATH
0044
0045 (1) Index Keys: col001 (Serial, fragments: ALL)
0046 Lower Index Filter: Abigail.tb.col003 = Abigail.tc.col001
0048


10.4.1 Reading the new query plan
By reading the query plan displayed in Example 10-2 on page 534, we see the
following:
Lines 005-0018 list the very same SQL SELECT we have been reviewing
throughout these last sections. All of the joins are the same, and all of the
filters are the same.
Lines 0020-022 lists the query optimizer directives in effect. There are none.
Line 0024 lists the estimated cost of this query. This estimated cost
(1,608,720) is higher than the cost from Example 10-1 on page 509, which
was (407,851). The amount of resource, both memory and CPU, is higher
which will be seen in this list, but the execution time (which is not displayed) is
much lower.
Lines 0028, 0034, and 0043 display the table order for this query plan.
– This query plan accesses the tables in the order of B, then A, then C,
which is our preferred order.
– We wish to process table B first because of its highly selective filter
criteria. (Remember that the filter criteria on table B returns 700 records,
on average, from a total record count of 840,000. This means we only
have to perform 700 joins to both tables A and C, not 840,000 joins from
table A to B.)
Line 0028 lists the table access method for table B.
– The first operation we wish to perform on table B is application of the filter
criteria. The filter criteria on table B effectively reduces the size of this
table, as records are rejected when they do not match the filter,
– WHERE col003 = ‘555’;
– The table access method on table B would normally have been a
sequential scan, because no usable index is headed with col003 (the
column referenced in the filter). The concatenated index on table B
columns col001, col002, col003 would be disallowed because of a
negation of index, non-anchored composite key.
– However, we can use the index on table B, col001, col002, col003
because of an exception. In table B, we are only reading columns col001
and col003 which are both members of the col001, col002, col003 index.
The only exception to the negation of index, non-anchored key
behavior is when that index may be used in a key-only indexed retrieval.


– An index is a vertical slice of a table, containing all records but only a
subset of the columns. The index on table B, columns col001, col002,
col003 will be read sequentially from top to bottom, because of the
non-anchored condition. Reading this index is still more efficient than
sequentially scanning the SQL table proper, as the index is presorted and
measurably more compact.

Note: Another technology, not previously mentioned, will also be employed
here. That is read ahead scans, also known as pre-fetching. When the
database server detects (SQL table proper, data) sequential scans, or a full
range (all records, sequential) index leaf page scan, which is mentioned
above, or a bounded (data range driven) scan of leaf pages within an index,
the query processor reads numerous pages in one physical I/O call, knowing
in advance that these pages are needed and will be consumed.

In this example, we are sequentially scanning 840,000 records in an index.
Because the index is small (3 columns, 44 or so bytes, with duplicate entries
which reduce the overall size, perhaps 14 MB, per our example), it is possible
we will read this table in as few as 8-20 physical I/O calls because of read
ahead scans.

Line 0034 lists the table access method for table A, and there is much new
technology to discuss here.
– It was mentioned many times that the original query plan table order could
not be table B then table A, because the join to the second table, table A,
is not served via an index. No query optimizer would allow a second or
subsequent table in the processing of a query to be read sequentially. If
this were allowed, it would mean that each of 700 records from table B
would cause a sequential scan into table A. That absolutely would not
perform well.
– But, the new query plan does read table B, then table A, and without
preexisting support of an index in place on table A. How is this allowed?
The query processor will build a temporary index to support an indexed
retrieval into table A. An indexed retrieval will be the access method for
table A, even though no index had previously existed.

We are going to pause the review of the new query plan to complete a discussion
of table access methods.

10.4.2 All table access methods
Most vendors recognize four or five table access methods, which include:


Sequential (table) access method is used when reading the records data
page after data page, and applying filters as required by the query criteria.
Read ahead scans are generally applied during sequential scans, unless
record locking would discourage this behavior (you do not want to lock an
unnecessary number of records and hurt multi-user concurrency and
performance.)
Indexed (table) access method is used when reading a single or set of
matching key values from an index structure, and then possibly retrieving the
remaining (non-indexed) table columns from the SQL table proper.
– B-tree+ indexes are most common. B-tree+ index technology is
discussed briefly with Figure 10-7 on page 522.
– Hash indexes (also known as hash tables) are another indexing
algorithm. Actually, hash indexes are many times more performant than
b-tree+ indexes under the proper conditions.
– Hash indexes lose nearly all performance if any significant amount of the
index data changes, which is common in relational database servers.
While b-tree+ indexes perform poorly (with all of their hierarchy levels,
branch and twig overhead, page splitting, and so on), when compared to
hash indexes, b-tree+ indexes do not lose performance with volatile data
sets.
– During the execution of a SQL SELECT (or any query, which is meant to
include the read phase of SQL UPDATE and SQL DELETE), the data is
non-volatile; that is, the data is read as a moment in time. Under these
conditions, a hash index may be used and will be more performant than a
b-tree+ index.
– Most relational database server vendors have hash indexes, but not
permanent hash indexes. They use hash indexes during query
processing, and nowhere else.
– Perhaps the quickest way to describe hash indexes is to compare them to
an arithmetic modulus operand; given (n) records, and a known
distribution of key values, you can determine where a given key value is in
this list via a calculation. A small amount of (index leaf page data)
scanning may be necessary to correct for any error in this calculation.
Hash indexes have essentially a single leaf level, to use the b-tree+ index
analogy. This is why hash indexes do not handle volatile data well; they
have no branch and twig levels to split for new data insertion in the middle
of their bounded ranges. Volatile data would soon cause the entire hash
index data structure to have to be rewritten.
– Bitmap indexes are another indexing algorithm. Bitmap indexes best
serve highly repetitive demographic type data, and are used often in
decision support applications.


– Perhaps the quickest way to describe bitmap indexes is to envision a two
dimensional matrix of all SQL table proper record physical disk addresses
(the x-axis dimension), and then all possible key values (the y-axis
dimension). Any point in the two dimensional graph contains a single bit (a
very small amount of storage) which is boolean; a true or false condition to
answer, is this record a member of this key value group?
– The performance of bitmap indexes lies in the manner that complex
groupings of boolean logic (AND and OR conditions) that bitmap indexes
can process with highly efficient bit-wise boolean software operands.
– Bitmap indexes can handle volatile or static data.
– R-tree (Region tree) indexes, or Bounded box indexes, are used often
for geographic data to form associations between near or like data.
– Perhaps the easiest manner to describe R-tree indexes is to envision a
single point on a map, and then a listing of every other point located within
every gradient of one kilometer. There would be a list of every other point
within 1 kilometer, 2 kilometers, and so on. The R-tree data can also
encode direction, such as North of the given point and 1 kilometer away.
– Star indexes are not a specific term we define here, but more a grouping
of index technologies to which various relational database vendors give
many names. Generally, star indexes are those used to list key columns
from two or more SQL tables; representing a pre-joined set of records to
serve a frequently observed query.
– Star indexes can have great difficulty with volatile data, since they
represent records from numerous tables, that have also to be evaluated
for a join criteria. (Add a new record to one table in the joined table list,
and you have to perform the joined select to see which new records have
to be added to this multi-table index; each record add calls for execution of
a joined SQL SELECT.)
Remote path (table) access method is used when one or more of the
involved tables is located in a remote database server. Given a two table SQL
SELECT, where the tables are joined and each table has a single filter
criteria, here is what is likely to happen:
– The portion of the query involving the remote SQL table will be sent to the
remote server for processing; that is, just the filter criteria is applied to the
remote table.
– The filtered (remote) result set is copied to a temporary table on the local
database server.
– Query processing will continue as though the remote table were local,
which it essentially now is.


– Most vendors will not attempt to join records over a network as the
communication latency would prove too slow to the overall execution of
the query.
– A good query optimizer will have knowledge of the remote table, and
based on selectivity of filters and cardinality of the join condition, may filter
the local table, export it to the remote server, do the join, then copy the
results back. Obviously with that additional communication overhead, the
local database server would require strong evidence to take this query
plan. The IBM DB2 UDB query optimizer excels at this capability.
Auto-index (table) access method involves something similar to the three
table SQL SELECT example we have been examining.
– As mentioned, the query optimizer will generally disallow having a
sequential table access method on two or more tables inside a single SQL
SELECT.
– Given a two table SQL SELECT where neither side of the join pair is
indexed, what can the query optimizer do. Years ago, the query processor
would create a full multi-user mode b-tree+ style index on the larger of the
two tables, then sequentially scan the smaller table, and join to the now
indexed (what was largest) table. This index was then discarded
immediately after query processing. (The index was discarded because it
was built at a moment in time, like a hash index. That is, the index was not
maintained during query processing.)
– Because creation of b-tree+ style indexes is very expensive, especially
during query processing, this behavior has been changed (improved). The
query optimizer will create a hash index, versus a b-tree+, to perform the
query processing.
– While technically a table access method, creating a hash index to perform
a table join is often recorded as a table join method. (If we do not have a
join, we do not auto-index.)
ROWID (table) access method: ROWID is generally a term used to refer to
the special column in each table which is the absolute disk address of a given
record on the hard disk. A ROWID access method is the most efficient since
the physical location of a given record is known. As a table access method,
the term ROWID is rarely seen; its use is implied when query processing
through the given tables and records.

10.4.3 Continue reading the new query plan
Now that all table access methods have been fully defined, we continue on line
0034 in Example 10-2 on page 534, to review that access method on table A:


As displayed, the query processor will perform a sequential scan on table A,
but only so it may build a hash index on table A, column col002. The table
access method on table A is actually (hash) indexed retrieval; that is, as soon
as the index is built.
– The query optimizer detects that a filter is in effect on column col001 of
table A but that this filter is not very selective, and would be better served
by a sequential scan (to build the index).
– The hash index created on table A col002 may fit entirely in memory, or
may be large enough that it will overflow to temporary storage on disk.
While the query plan contains a section for temporary files (objects), it
often does not report temporary files required for hash index creation.
Another administrative tool is needed to uncover this activity.
– After the hash index on table A is built, query processing continues joining
table B to table A.
– The newly created hash index contains the table A col002 key value, in
ascending key value sort order, along with a physical disk address of the
remainder of the table record in the SQL table proper.
– Line 0037 displays the filter on table A.
– Lines 0040-0041 display the join method between tables B then A. The
join method listed is dynamic hash join, but it is more commonly referred to
as a hash join method.
Based on the level of query optimizer statistics gathered, we were seeing
indexed retrieval on table A to build the hash index, or then in other query
plans, sequential. Even though we do not like the current index on table A,
it is a (small) vertical slice version of table A and may be a more performant
method to retrieve the data for the hash index.
Lines 0043-0047 detail table C.
– Table C is accessed via an indexed retrieval and joined via a nested loop.
– Because the join to table C is via an equality to a single column in a unique
index, table C can only return one record per join. For this reason, no read
ahead scan behavior is needed for table C.

10.4.4 All table join methods
Most relational database software vendors have at least three or four (table) join
methods. Listed below are comments regarding (table) join methods:
Nested loop (table) join method:


– Nested loop (table) join method is the oldest join method. The newer
(table) join methods have only arrived in the last five to seven years with
parallel process architecture relational database servers.
– The nested loop (table) join method is dependent on at least one of the
two join tables having a preexisting index. If one of two tables is indexed,
the non-indexed table is read first via a sequential scan, while qualifying
rows are read from the (second) table via an indexed read.
– If both join tables lack an index, then one of two tables will have a hash
index created on the join columns, of the pre-filtered records; this behavior
is called an auto-index. Technically this is an indexed table access method
and a hash index (table) join method.
– Nested loop (table) join method is often viewed as the most efficient join
method based on total execution time of smaller data sets, but this is
somewhat misleading. The nested loop (table) join method depends on
indexes (a workload with an associated execution time) that are created
and maintained before query execution.
– The nested loop join (table) join method is probably less CPU and disk I/O
efficient than other (table) join methods when larger-sized data sets are in
effect. This is somewhat dependent on index efficiency, selectivity of
filters, and other run-time conditions.
Hash index (table) join method:
– The hash index (table) join method is somewhat like a nested loop (table)
join method in function; an iterative (reentrant) software programming loop
is employed. However, the hash index is so much more efficient based on
its function and smaller size, that this comparison is rarely made. (Efficient
once the hash index is created.)
– Few if any relational database server vendors have permanent hash
indexes, and prefer instead to use them exclusively during query
processing.
– A hash index table join method is reviewed in the query optimizer example
detailed in “Query optimizer by example” on page 503, and will not be
expanded upon further here.
– Use of a hash index has two phases. During the build phase, the hash
index is created. During the probe phase, a second table uses the hash
index to perform the join.
Sort merge (table) join method:
– Sort merge (table) join method is definitely the most technically advanced
(table) join method. It is generally most used in the decision support
application area.


– To better understand sort merge (table) join method, first we are going to
offer a little more detail regarding the nested loop (table) join method so
that we may compare and contrast. Figure 10-16 on page 543 displays a
simplistic rendering of a nested loop (table) join method.

8
1 3
11 4
4 4
7 4
9 6
6 7
… 7
7
8
11
11
11
…

Figure 10-16 Graphical depiction of nested loop (table) join method

Concerning Figure 10-16:
The query processor only joins two tables at a time, not three or more. To join
three or more tables, first two tables are joined, and then the third table is
joined to this two-table (now joined) product, and so on until the query is
complete.
It has been stated that the query optimizer will not allow for a query plan
wherein the second table is accessed via a sequential table access method.
If there is no index to support accessing the second table, one is built during
processing of the query. Following this idea, the table on the right of
Figure 10-16 is shown to have data in sort order, where the data in the table
on the left may be in random or ordered sequence.
The sort order present in the second table is meant to imply some form of
indexed retrieval.
The lack of sort order in the first table is meant to imply that this table may by
read via a sequential scan. (Data in an SQL table is generally not presorted,
and is assumed to be in random sequence.)


The table on the left will be read one time only, from top to bottom via
whatever table access method is in effect. If any filter criteria is in effect for
the left table, this filter criteria will be applied. Only those records which satisfy
the filter criteria are then joined to table two. Each single row from table one
which meets the filter criteria will cause table two to be accessed. If 20
records from table one satisfy the filter criteria, then table two will be
accessed 20 times. If each record from table one joins with 10 records in table
two, 20 accesses of 10 records each equals 200 total accesses of table two.
The table on the right is accessed via an indexed retrieval. The arrows
between the two tables in Figure 10-16 on page 543 are meant to signify
program flow control; that is, a recursive loop starts by positioning within table
one, and then each row in table one meeting the filter criteria can access all
or a subset of table two.
Given two tables with all other factors being equal (table size, selectivity of
the filters being applied to each table are equal or nearly equal, and certainly
other factors), the query optimizer follows a convention of worst table first;
that is, given two operations, one which is efficient and one inefficient, and
one of these operations drives the second which must be repeated, lead with
the inefficient operation which causes a repetition of an efficient operation. In
other words, given one indexed table and one which is not, lead with the
non-indexed table, suffer through the sequential scan (which is done once),
and repeat an operation which is a small, targeted (indexed) read of the
second table.

Concerning Figure 10-17 on page 545:
The nested loop (table) join method from Figure 10-16 on page 543, has no
startup costs; we can just begin by sequentially scanning table one, and join
into table two via an indexed read. If neither table is indexed, then one table
will be indexed during query processing using, most likely, a dynamic hash
join into table two. However, that is a hash join, not a nested loop join.
A sort merge (table) join method is generally used in business intelligence
and/or when the following run-time conditions are in effect:
– Neither table from the join pair is indexed, or such a large volume of these
two tables is to be read that indexed (table) access methods would prove
too costly. Filters, while present, are still expected to produce a larger
percentage of records from the tables than not.
– This raises a general condition of business intelligence (BI) queries over
those seen in online transaction processing (OLTP); that is, BI queries
read sweeping and voluminous amounts of data with little or no support
from indexes, while OLTP read few records and via known indexed table
access methods.


– Also, BI queries are often accompanied with set operands; that is,
aggregate computations on received rows, and computations such as
AVG (average). These aggregate calculations require the data set that
they report on to be sorted by several columns which are often not
indexed.
The query from Figure 10-16 on page 543 was expected to process 20 times
10 records (200 records). The query in Figure 10-17 is expected to process
all or most of both tables (many hundreds of records, and then with a heavy
data sorting requirement).

3
1
4
4
4
6
4
7
6
8
7
9
7
11
7
…
8
11
11
11
…

Figure 10-17 Graphical depiction of sort merge (table) join method

Figure 10-17 displays two data sets, table one data and also table two data.
Table one data is read sequentially, has any filter criteria applied, and then
sorts the data by join column key value. The same is done to table two, either
concurrently or in sequence.
Then, table one is read from top to bottom one time only. As a single record is
produced from table one, it is joined to any matching records in table two. As
the join pairs from table two are exhausted (we have reached the end of the
range of records in table two which do in fact join), processing returns to table
one. The primary point here is, table one is read from top to bottom one time
only, and the same is true for table two.
The sort merge (table) join method has measurably larger startup costs; that
is, filtering and sorting both tables before any processing is able to be
performed. However, after this preparatory processing is complete, the
sort-merge (table) join method is very fast, measurably faster than a nested


loop (table) join method, since significantly less post-preparatory work needs
to be performed.

The final comparison on nested loop versus sort-merge (table) join methods is
listed below:
If all run-time conditions are met, a nested loop join method will produce the
first row first; meaning, nested loop has little or no preparatory work, and will
output a joined record before sort-merge (table) join method.
If all run-time conditions are met, a sort-merge (table) join method will
produce the entire data set first; meaning, even with all of the preparatory
work to perform, sort-merge join is measurably more CPU efficient than a
nested loop (table) join method. Nested loop (table) join method accesses the
second table multiple times, sort-merge (table) join method accesses the
second table once. It is a trade-off of waiting for preparatory work to
complete, versus not waiting.

Note: If you have seen query optimizer influencing statements that call for
first row first or total data set (all rows) performance; this is an area where
these types of statements are considered.

There are times where the run-time environment does not support a nested
loop (table) join method, where you must employ a dynamic hash (table) join
method or sort-merge (table) join method.

10.4.5 Continuing the list of all table join methods
The fourth, and perhaps most interestingly named, table join method is the push
down semi-hash table join method. The push down semi-hash join is used
almost exclusively in the area of business intelligence and decision support. The
run-time environment that calls for the execution of a push down semi-hash join
is listed below:
As a decision support (business intelligence) query, the tables being
referenced in this type of query are expected to be data modeled using the
dimensional data modeling; a large central fact table containing millions of
records which represent some event or intersection of data. A number of
small dimension tables contain the indexes to this data. (Index as in
demographic vector; indexes such as time, location, and product
categorization. Index in this context does not mean database index, as in
hash or b-tree+.) Figure 10-18 on page 547 displays this kind of query.


Dimension table
Dimension table
Geography Index
Time index
48 Rows
120 Rows
Fact table

SELECT *, SUM(s.revenue) IS_BEST Sales Event
Dimension table
FROM time t, sales s, geography g, customer c 56,777,323,892 Rows Product Index
WHERE t.key = s.t_key AND g.ket = s.g_key AND 120 Rows

p.key = s.p_key AND c.key = s.c_key AND
s.salesQuarter = 4 AND s.salesYear = 2008 AND
Dimension table
s.productSKU = “IDS” AND g.geoCode = “WI”
Customer Index
AND p.productGroup = “MGD” 88,892 Rows
GROUP BY s.salesYear, s.salesQuarter
ORDER BY s.productSKU, IS_BEST;

Figure 10-18 Example query referencing dimensional modeled tables

In Figure 10-18, the Sales table has nearly all of the records by count, and
that number is massive compared to the remaining tables. All of the
dimension tables, in this case there are four, join to the Sales (central fact)
table, which is typical in a dimensional model. The problem is, any normal
query processing is going to hit the Sales table within the first or second table,
and processing Sales is going to produce a result set of thousands or millions
of records, thousands or millions of records which then must be joined to the
three remaining dimension tables; (this results in lots of looping and CPU
consumption.)
Figure 10-19 on page 548 is a graphical depiction of a nested loop join
method of this query. If Product-A represents the join of the first two tables,
Customer and Sales, this result set will have millions of records; millions of
records which then have to be joined to the three remaining tables; Product-B
will have millions of records, Product-C, and so on.


Dimension table
Prod-C JOIN
Geography Index
48 Rows
SCAN
Prod-B Dimension table
JOIN
Product Index
Prod-A 120 Rows
SCAN
JOIN
Dimension table
Time index
120 Rows
SCAN SCAN

Dimension table
Fact table
Customer Index
Sales Event
88,892 Rows
56,777,323,892 Rows

Figure 10-19 Nested loop join of earlier query, also referred to as a left deep tree join

A push down semi-hash join is designed to significantly reduce the number of
joins that have to be performed by processing the (large) central fact table last.
How to do this is not immediately apparent, because all of the dimension tables
join to the single fact table, the dimension tables do not join to themselves.

A push down semi-hash join accomplishes this goal (processing the fact table
last) in the following manner:
Using the transitive dependency query optimizer query rewrite capability,
copy as many of the filter criteria to the dimension tables as possible. The
query in Figure 10-18 on page 547 displays five filter criteria. Example as
shown:
WHERE s.salesQuarter = 4 AND s.salesYear = 2008 AND
s.productSKU = “IDS” and g.geoCode = “WI” AND
p.productGroup = “MGD”;
The two filters on s.salesQuarter and s.salesYear can be copied and applied
to the Time table, thus reducing at least a small percentage of records that
will need to be processed.
The phrase push down refers the following; join all of the dimension tables
into one large result set, applying any applicable filters as this is done. Since
the dimension tables do not relate to one another, join these tables as an
outer Cartesian product. In other words, create a large intersection data set of


these four tables. What is remaining then is a standard two table join, the
single table which is the intersection of all of the dimension tables, and the
original fact table as it existed.
These are two options:
– Standard nested loop; 88,892 records joined to 56,777,323,892, then 120,
48, and 120, resulting in several billion joins.
– Or, push down semi-hash join; join 88,892 records to 120, 48, and 120,
then join to the large 56,777,323,892 record table, resulting in fewer joins.
The term hash in push down semi-hash join refers to the hash index that is
built on the outer Cartesian result set of the dimension tables.
In a clustered database environment (the database is spread across several
physical hardware nodes operating in a single cluster), the outer Cartesian
product of the dimension tables is often broadcast to each node in whole, so
that all joins may be performed locally on each separate node.

Example index needed for push down semi-hash join
Consider the SQL SELECT listed below. It is a typical business intelligence, star
schema type SELECT:
SELECT SUM(o.price)
FROM orders o, customers c, suppliers s, product p, clerks c
WHERE o.custid = c.custid AND o.suppid = s.suppid AND
o.prodid = p.prodid AND o.clerkid = c.clerkid AND
c.zipcode = “77491-6261” AND s.name = “KEENE” AND
p.type = “WEB DATABLADE” AND c.dept “ISV”;

The optimal query plan is to perform a push down semi-hash join on; Customers,
Suppliers, Product, and Clerks. The query optimizer has rules when to consider a
a push down semi-hash join. One rule is related to having filters present on some
or all of the detail tables (such as Customer and Suppliers). Or, the rules may
entail maximum record counts that may exist in any of these (detail) tables.
Further, in order for the push down semi-hash join to be available, there needs to
be a specific index on the fact table, in this case, Orders. The Orders table
requires an index on the Orders table join columns:
CREATE INDEX i44 ON orders (custid, suppid, prodid, clerkid);

A more verbose depiction of a nested loop join method
Consider the SQL SELECT statement listed below:
SELECT c.custno, o.orderno


FROM customer c, orders o, items i
WHERE c.custno = o.custno AND
o.orderno = i.orderno;

The first thing to discuss is that there are several more table join orders than you
may realize:
Join table Orders to Customer, then join the result to Items
Join Customer to Orders, then Items
Join Order to Items, then Customer
Join Items to Orders, the Customer
(Not likely, but still briefly considered) Join Customer to Items, then Orders
(Not likely, but still briefly considered) Join Items to Customer, then Orders

The program logic for a table order of Customer to Orders, then Items, is:
FOR EACH ROW IN customer TABLE DO
READ THE ROW INTO c
FOR EACH ROW IN orders TABLE DO
READ THE ROW INTO o
IF o.custno == c.custno THEN
FOR EACH ROW IN items TABLE DO
READ THE ROW INTO i
IF i.orderno == o.orderno THEN
RETURN THIS ROW TO THE USER APPLICATION
END IF
END FOR
END IF
END FOR
END FOR

From this sample program code, perhaps you can see the CPU intensive nature
of the nested loop (table) join method. Also, the ability for the nested loop (table)
join method to return the first record after just a few statements can also be seen
(three record reads, and two IF statements).


10.5 Rules and cost-based query optimizers
After the example query reviewed in 10.2, “Query optimizer by example” on
page 503, through 10.4, “Query optimizer example solution update” on
page 534, now we are going to perform a step-by-step detailed review of the all
query optimizer algorithms. We cover single and simplistic examples for each
data point. After this section, you might then apply all of the points covered here
towards the previous example.

The very first query optimizers relied more on (heuristic) program logic to
determine the most optimal query plan. As relational database software products
became more capable, this logic was enhanced with data related to the tables
themselves, and the system which contains them to include; the distinctiveness
of duplicate key values, the expected physical performance of the hard disks
these tables reside on, the current CPU load or network transfer rates, the
amount of available sort memory, and more. As you can see, the query optimizer
considers more than table size or the syntax of the SQL SELECT to determine
query plans.

When software vendors added more and more awareness of the run-time
environment to their query plan calculation, they began to call their query
optimizers cost-based query optimizers. Cost-based query optimizers did not so
much replace rules-based query optimizers as they did enhance them.
Rules-based query optimizers are based on sound logic. A cost-based query
optimizer adds more knowledge to the already sound algorithm of a rules-based
query optimizer. Without being overly simplistic, all query optimizers result in a
large case statement (with perhaps tens of thousands of lines of code) which
determine the query plan.

While different vendors may list more or fewer primary query rules, the list below
is generally what most vendors recognize. The primary rules of a query optimizer
(which are supplemented with cost-based data) include:
Rule 1: Outer table joins
Rule 2: (Non-outer) table joins
Rule 3: Filter criteria
Rule 4: Table size
Rule 5: Table cardinality

Note: As a given SQL SELECT is received by the query optimizer, these five
rules are applied in the order listed above until, among other things, a table
order is determined; in other words, the order in which the tables will be
accessed and then joined.


Each of these rules is expanded upon in the sections that follow.

10.5.1 Rule 1: Outer table joins
If you execute a two table SQL SELECT, and the tables are joined, a record from
one table is only returned where there is a matching record in the second table;
that is, if you join the Customer and Customer-Order tables, no records from
Customer are returned if a given customer has never placed an order. This can
be both useful and not so useful. If you need to send a holiday card to each
customer thanking them for the (n) orders they placed last year, no card is going
to the customer who has yet to place orders.

An outer join is one that returns records from joined SQL tables where the stated
join criteria exists, and returns only the dominant table record columns where
the relationship does not exist; that is, where there is a relationship between two
tables, the outer acts like a normal join. When there is not a matching
subordinate record, those columns are populated with SQL NULL values.

The term outer join is not entirely industry standard. Some relational database
vendors may refer to an outer join as a left-handed join, right-handed join,
conditional join, and other similar terms.

Example 10-3 displays an outer join, a join between a Customer and (Customer)
Orders tables.

Example 10-3 Example of SQL SELECT with an outer join

0000
0001 QUERY:
0002 ------
0003 select
0004 *
0005 from customer t1, outer orders t2
0006 where
0007 t1.customer_num = t2.customer_num
0008
0011
0012 1) Abigail.t1: SEQUENTIAL SCAN
0013
0014 2) Abigail.t2: INDEX PATH
0015
0016 (1) Index Keys: customer_num (Serial, fragments: ALL)
0017 Lower Index Filter: Abigail.t1.customer_num =
0018 Abigail.t2.customer_num


0020
0021

In an outer join, one table is referred to as the dominant table, while the other is
subordinate. The dominant table is the one that will return records whether or not
a matching record exists in the other subordinate table. Consider the following:
There are two table join orders to the above query; table one then two, or
table two then one.
If the dominant table is read first, then we have all of the records we need
from table one to satisfy the query. We join to the subordinate table and return
joined records when found. If a table two join record is not found, then one
record from table one is returned with SQL NULL values filling the columns
that normally come from the subordinate table matching record.
If we read the subordinate table first, we process the join as though
performing a normal (non-outer) SQL SELECT; reading the subordinate table
first gives us join column key values as they as found to exist in the
subordinate table. However, after query processing, we somehow have to
produce those rows (the join column key values) from the dominant table that
do not exist in subordinate table. We would have to execute a query similar
to:
SELECT dominant-table-primary-key
FROM dominant-table
WHERE dominant-table-primary-key NOT IN
(SELECT subordinate-table-foreign-key FROM subordinate-table);
The above would be a very expensive query; a NOT IN to a nested (one SQL
SELECT embedded within another) SQL SELECT can produce many, many
disk I/O operations.

Note: In an SQL SELECT with an outer join criteria, the dominant table is
always processed before the subordinate table. Obviously, the mere presence
of an outer join criteria greatly affects table join order, which can then greatly
and negatively affect overall query performance. If you can avoid use of the
outer join criteria, do so.

From the text presentation of the query plan in Example 10-3 on page 552:
Line 0005 displays that the Customer table is dominant, because the outer
keyword appears before the Orders table; the presence of the outer keyword
before the Orders table makes it subordinate.
Line 0012 displays that the dominant table is read first, as expected.


Because there is no filter criteria on t1 (the Customer table), t1 is processed
with a sequential table access method; line 0012. (We are returning every
record from table t1. It is more efficient to access this table sequentially than
via an index; that versus using an index and then a retrieval of the full record
from the SQL table proper.)

Example 10-4 displays the same query as in Example 10-3 on page 552,
however, the outer keyword is applied to the Customer table, not the Orders table
as before:

Example 10-4 Example of outer join with outer keyword having been moved
0000
0001 QUERY:
0002 ------
0003 select
0004 *
0005 from outer customer t1, customer t2
0006 where
0008
0011
0013
0015
0020
0021

From Example 10-4:
The table order has changed with the movement of the outer keyword. With t2
now being the dominant table, that table is processed first.
Without the presence of filters and other query criteria, nearly every aspect of
Example 10-3 on page 552 and Example 10-4 remain the same; the costs,
and others. In a more realistic SQL SELECT, the affect of the outer keyword
on query cost and amount of resource to process the query would become
evident.


The next example, Example 10-5, displays what occurs if we attempt to use
query optimizer directives to force processing of the subordinate table before the
dominant table. (If allowed, this query would perform very poorly relative to the
other choice for query plans.)

Example 10-5 Use of query optimizer directives and outer join
0000
0001 QUERY:
0002 ------
0003 select --+ ordered
0004 *
0005 from outer customer t1, customer t2
0006 where
0008
0009 DIRECTIVES FOLLOWED:
0010 DIRECTIVES NOT FOLLOWED:
0011 ORDERED Outerjoin nesting not compatible with ORDERED.
0012
0015
0017
0019
0024

From Example 10-5:
The query optimizer directive we used was ordered, as shown on line 0003.
The ordered query optimizer directive keyword states that the table join order
must follow the names of the tables as they are listed in the SQL FROM
clause.
Line 0011 shows that the query optimizer directive was not followed, because
it is expected that execution of this query plan would not be performant.

As stated, the query optimizer applies these five rules in order until a table
processing order is determined. In each of the three queries above, rule one fully
determined table processing order. The query optimizer did not need to refer to
rules two through five to determine table order.


10.5.2 Rule 2: (Non-outer, normal) table joins
Outer joins are not that common. Out of 100 queries, you can see outer joins two
or four times out of 100. Joins (non-outer or standard joins) usually are expected
to be everywhere. In any given query, each table is expected to be joined to at
least one other table. If any given table is not joined to at least one other table,
this results in an outer Cartesian product.

Earlier, when discussing nested loop (table) join methods, we introduced the
concept of worst table first. Given a two table SQL SELECT, where the tables are
joined, and both sides of the join pair are supported via an index, make the first
table processed to be the least efficient of the two choices; that is, generally
make the second table (the table access operation that is performed numerous
times), the more efficient of the two.

Example 10-6 displays three queries, two with two tables and the last with three
tables.

Example 10-6 Three examples of queries with joined tables
0000
0001 QUERY:
0002 ------
0003 select
0004 t1.*, t2.*
0005 from customer t1, orders t2
0006 where
0008
0011
0013
0015
0020
0021
0022 QUERY:
0023 ------
0024 select
0025 t1.*, t2.*
0026 from orders t1, items t2
0027 where


0028 t1.order_num = t2.order_num
0029
0032
0034
0036
0037 (1) Index Keys: order_num (Serial, fragments: ALL)
0038 Lower Index Filter: Abigail.t1.order_num =
0039 Abigail.t2.order_num
0041
0042
0043 QUERY:
0044 ------
0045 select
0046 t1.*, t2.*, t3.*
0047 from customer t1, orders t2, items t3
0048 where
0050 and
0051 t2.order_num = t3.order_num
0052
0055
0057
0059
0061 Lower Index Filter: Abigail.t3.order_num =
0062 Abigail.t2.order_num
0064
0066
0071


You can easily witness outer joins and their impact. Non-outer (normal) joins are
perhaps less obvious because their presence is largely expected. From the first
query in Example 10-6 on page 556, lines 0000 through 0020, see the following:
Rule 1, an outer join is not present, so query optimization continues with rule
2, normal (non-outer) joins.
The two tables are joined, as shown on line 0007. With just two tables, the
presence of the join criteria alone is not enough to determine table join order.
Actually, a variation of rule 3 determines table join order for this query.
– Rule 3 is discussed below. For now, we will state that table join order was
determined by the principle of worst table first.
– On lines 0017 through 0018, the join columns are stated to be:
t1.customer_num = t2.customer_num
– T1 is the Customer table, which has a unique index on customer_num (not
shown in the example). T2 is the Orders table, which has a duplicate index
on customer_num (not shown in the example).
– As a general rule the better index is the unique index, since the duplicate
index permits numerous values. It is better to process the duplicate table
first, and use the unique index to process the repetitive operation. The
worst table is that with the duplicate index.

From the second query in Example 10-6 on page 556, lines 0022 through 0041,
see the following:
The second query is much like the first, a unique index column joined to a
duplicate indexed column. The query plan is essentially the same between
the two queries.
Evaluating the selectivity, or uniqueness of an index is, of course, a heuristic
(greedy, general guideline) type of rule. These rules apply when all other
criteria are essentially the same. See the following:
– Unique indexes generally are more selective (return fewer records) than
duplicate indexes.
– Duplicate indexes with more columns (a composite index) are generally
more selective than a duplicate index with fewer columns.
– The query optimizer considers the lower cost of joining integer columns
over character columns, which are more expensive.
– Given two indexes, both of equal selectivity, the query optimizer will give
preference to any index which may also satisfy the SQL ORDER BY
clause. Example as shown:
SELECT * FROM one_table WHERE
ixcol1 = 10 AND ixcol2 = 15


ORDER BY ixcol1;
Using the index on ixcol1 to process the query also returns the records in
a presorted manner, and is subsequently more efficient.

From the third query in Example 10-6 on page 556, lines 0043 through 0071, see
the following:
This query gives us the first true glimpse of rule 2.
The first table to be processed in this query is t3, the Items (order line items, a
detail table to customer orders) table, determined by rule 3.
The remaining table order is then determined by rule 2; table 3 joins with table
2 (Orders, the customer orders table), then table 2 with table 1 (Customer).

10.5.3 Rule 3: (Presence and selectivity of) Filter columns
Because outer joins (Rule 1) are rare, and all tables are expected to be joined
(Rule 2), Rule 3 is often the rule that determines most table orders in the
calculation of a query plan. Generally, the data regarding the selectivity
(cost-based) of these filters is the most important determinant in the creation of
the query plan; that is, which filter, if executed first, will reduce the largest
number of records to be processed, which then reduces query cost.

Previously we were using the phrase worst table first, and that concept still
applies (if all other options and costs are equal). However, with the introduction
of filters to the query, we will now use the phrase, as efficiently as possible,
reduce the size of the tables that are being examined. In other words, give
preference to the best filter to effectively reduce the size of the tables being
examined (reduce the amount of disk I/O to be performed).

Further, we also introduce the acronym, FJS (filter, then join, then sort), which
is the high level summary of processing and processing order that occurs in the
execution of a query.

The following queries include a number of criteria meant to explore Rule 3, filter
criteria, as well as reinforce Rule 1, outer joins, and Rule 2. These queries are
depicted in Example 10-7.

Example 10-7 Several queries, demonstrating Rules 1 through 3
0000
0001 QUERY: the first query
0002 ------
0003 select
0004 o.order_num,
0005 sum (i.total_price) price,
0006 paid_date - order_date span


0007 from
0008 orders o,
0009 items i
0010 where
0011 o.order_date > ‘01/01/89’
0012 and
0013 o.customer_num > 110
0014 and
0015 o.order_num = i.order_num
0016 group by 1, 3
0017 having count (*) < 5
0018 order by 3
0019 into temp t1
0020
0023 Temporary Files Required For: Order By Group By
0024
0025 1) Abigail.o: SEQUENTIAL SCAN
0026
0027 Filters: (Abigail.o.order_date > 01/01/2089 AND
Abigail.o.customer_num > 110 )
0028
0029 2) Abigail.i: INDEX PATH
0030
0032 Lower Index Filter: Abigail.o.order_num = Abigail.i.order_num
0034
0035
0036 QUERY: the second query
0037 ------
0038 select
0039 c.customer_num,
0040 c.lname,
0041 c.company,
0042 c.phone,
0043 u.call_dtime,
0044 u.call_descr
0045 from
0046 customer c,
0047 cust_calls u
0048 where
0049 c.customer_num = u.customer_num
0050 order by 1
0051 into temp t2
0052


0055 Temporary Files Required For: Order By
0056
0057 1) Abigail.u: SEQUENTIAL SCAN
0058
0059 2) Abigail.c: INDEX PATH
0060
0062 Lower Index Filter: Abigail.c.customer_num =
Abigail.u.customer_num
0064
0065
0066 QUERY: the third query
0067 ------
0068 select
0069 c.customer_num,
0070 c.lname,
0071 o.order_num,
0072 i.stock_num,
0073 i.manu_code,
0074 i.quantity
0075 from
0076 customer c,
0077 outer (orders o, items i)
0078 where
0079 c.customer_num = o.customer_num
0080 and
0081 o.order_num = i.order_num
0082 and
0083 manu_code IN (‘KAR’, ‘SHM’)
0084 order by lname
0085 into temp t3
0086
0090
0091 1) Abigail.c: SEQUENTIAL SCAN
0092
0093 2) Abigail.o: INDEX PATH
0094
0096 Lower Index Filter: Abigail.c.customer_num =
Abigail.o.customer_num
0098
0099 3) Abigail.i: INDEX PATH
0100
0101 Filters: Abigail.i.manu_code IN (‘KAR’ , ‘SHM’ )


0102
0104 Lower Index Filter: Abigail.o.order_num = Abigail.i.order_num
0106
0107
0108 QUERY: the fourth query
0109 ------
0110 select
0111 *
0112 from
0113 stock
0114 where
0115 description like ‘%bicycle%’
0116 and
0117 manu_code not like ‘PRC’
0118 order by
0119 description, manu_code
0120 into temp t5
0121
0125
0126 1) Abigail.stock: SEQUENTIAL SCAN
0127
0128 Filters: (Abigail.stock.description LIKE ‘%bicycle%’
AND Abigail.stock.manu_code NOT LIKE ‘PRC’ )
0129

The first query in Example 10-7 on page 559 is contained between lines 0000
and 0034. See the following:
This query has two tables, as shown on lines 0008 and 0009. The tables are
joined on line 0015.
Rule 1 (outer joins) is not in effect, since there are no outer joins.
Rule 2 (table joins) does not help determine table order since there are only
two tables, and both sides of the join pair can be served via an indexed table
access method (there is an index on o.customer_num and i.order_num, not
shown).
The Orders table has two filters, as shown on lines 0011, and 0013. The
Items table has no filters.
There is also a filter on an aggregate expression, as shown on line 0017. If all
things are equal, this filter has a lower priority since we must filter and then
join these two tables before we can consider a filter on an aggregate
expression.


– This filter is on the cardinality between the two tables. That is, how many
join pairs have fewer than five records. We must fully process this query in
order to satisfy execution of this filter criteria.
– If the aggregate filter were on a MIN() or MAX(), or an indexed join column
or similar filter criteria, there would then be more processing choices that
the query optimizer could consider since MIN() and MAX() can be known
before the join under the very best conditions.
Line 0025 displays that the table join order begins with the Orders table,
(table “o”). If table o has 100 records, if the filter on table o reduces these 100
records to just 10 records, and if table i has 100 records, effectively the two
choices were:
– Read table “o” at 100 records, and apply the filter to produce 10 records.
Then join 10 filtered “o” records to 100, an operation of 1000 records.
– Or join table “i” at 100 records to table ‘o” at 100 records, an operation of
10,000 records, only to filter 10% of table “o” to produce the final 1000
records.
Due to rule 3, the presence of filter records on table “o”, the table join order is
table “o” then “i”.
Because there are two filters on columns in table “o”, and these columns are
not both members of one (composite) index, this table is read sequentially.
Other choices were:
– Read table “o” via an index on order_date, assuming one is present, and
apply the filter criteria for order_date. But, any qualifying records in table
“o” must also meet the criteria on customer_num. Since customer_num is
not a member of any order_date index, we would have to access the
record in the SQL table proper to evaluate the filter criteria for
customer_num.
– The converse of the above is read any index for customer_num, then
process the filter for order_date.
– Or, read two indexes, one for customer_num and one for order_num.
Compare both sets of results and produce an intersection, in other words,
return rows which are common to both result sets. Creating an intersection
of both result sets means both lists would have to be sorted, then
compared key by key, and so on.
– For cost reasons, the query optimizer chose to sequentially scan the
Orders table. (There was no index on order_date. The query optimizer will
create indexes to join tables. The query optimizer will not create indexes
just to process filters; creating an index requires reading the SQL table
proper, which would then allow for filter criteria examination without the
overhead of building the index.)


The table access method in table one (Orders) is sequential, line 0025.
A nested loop method, line 0033, joins table two (Items) to table one (Orders).
The access method on table two (Items) is indexed, line 0029.
A temporary file is required to process the SQL GROUP BY clause. A SQL
GROUP BY clause needs to sort records to perform its data sub-grouping.
The SQL ORDER BY clause is on a differing column set. Therefore, a second
sort operation will be performed to process the SQL ORDER BY. This too
requires a temporary file.
– If the SQL GROUP BY and SQL ORDER BY were on the same columns
and column orders, this second sort is not required.
– The SQL GROUP BY clause is on columns 1, then 3. The SQL ORDER
BY clause is on column 3. If the SQL HAVING clause had a condition
where column 1 equaled a single value, then the value in column 1
essentially equals a (new) constant, whatever that value is. For example:
SELECT * FROM t1 ORDER BY “A”, column2;
is logically equivalent to:
SELECT * FROM t1 ORDER BY column2;
This is true because “A” is constant, it is unchanging. Consider the
following:
SELECT * FROM t1 WHERE column5 = 10 ORDER BY column5, column6;
is logically equivalent to:
SELECT * FROM t1 WHERE column5 = 10 ORDER BY column6;
From the example query, line 0017, we see that the aggregate
expression column can have values in the range of 1 to 4, which is not
a single constant value. Therefore, the SQL ORDER BY clause
provides different sort criteria than the SQL GROUP BY clause, two
different sorts.


Note: The query rewrite capability of the query optimizer would detect these
types of logical equalities and rewrite the query to have the SQL GROUP BY
clause match the SQL ORDER BY clause, thus reducing a now redundant
sort operation. Example as shown:
These two queries are logically equal because column5 is constant.
As specified by the filter criteria “column5 = 10”

This specific case of the query rewrite capability is called the pseudo
order-by clause.

Line 0019, the retrieve results into a temporary table is used only for the
benefit of the Redbook authors. This call did not affect query processing. All
of the queries in this section are retrieved into a temporary result sets table.
Rule 3 determined the table order for this query.

The second query in Example 10-7 on page 559 is contained between lines 0036
There are two tables in this query, as shown on lines 0046 and 0047.
The tables are joined, line 0049, and there are no filters on either table.
An SQL ORDER BY clause exists on the Customer table, customer_num
column. Even though an index exists on this column, and that index is used to
perform the join into table two, rows are still not being returned in sort order.
The first table to be processed is cust_calls via a sequential access method;
therefore, rows are read in random sequence.
– Why not read the Customer table first via customer_num to match the
requested SQL ORDER BY clause sort order and thus avoid a sort
operation? That is certainly an option. Returning to the acronym FJS, the
query optimizer is concerned first with filtering and joining records before
sorting.
– Not having any filters on any of the tables (no selectivity) meant that it
would be measurably more efficient to read at least one of the two tables
sequentially. The worst table first principle has the query optimizer read
cust_calls (Customer Calls) first, and use the unique index into Customer
to process the join. The amount of physical I/O to read Customer or
cust_calls (Customer Calls) via an index when we are returning every row
in the table is prohibitive relative to a sequential scan.


– After the optimal table order was determined, the SQL ORDER BY clause
could not be served via the same index used to process the query.

Note: Another specific query rewrite capability is transitive dependency.
Transitive dependency is an algebraic term. In effect,
IF A = B, AND B = C, THEN, BY DEFINITION A = C

From the example above, this means,
These two queries are logically equal, based on the
principle of transitive dependency.
SELECT * FROM t1, t2 WHERE t1.col = t2.col ORDER BY t1.col;
SELECT * FROM t1, t2 WHERE t1.col = t2.col ORDER BY t2.col;

This query was determined largely by Rule 3; the expected uniqueness of the
join columns made the Customer table more desirable as a second join table,
and the lack of filter criteria called for a sequential scan of the worst table first,
the cust_calls (Customer Calls) table.
A temporary file is required to support the SQL ORDER BY clause.

The third query in Example 10-7 on page 559 is contained between lines 0065
Query three references three tables, as seen on lines 0076 and 0077.
Customer is the dominant table in an outer join, line 0076.
The Orders table and Items table are both subordinate to Customer, line
0077. Although, tables Orders and Items have a normal join relationship
between themselves.
Based on Rule 1, the Customer table will be processed first, as seen on line
0091.
Because Customer has no filter criteria, this table will be read sequentially.
(Absence of filter criteria on Customer between lines 0078 through 0083.)
Order was processed next due to Rule 2; it is joined directly with Customer.
A temporary file is required to process the SQL ORDER BY clause, which is
on Customer.lname. There was no index on this column, and certainly not
one that was already in use to process Customer.

The fourth query contained in Example 10-7 on page 559 is contained between
lines 0108 and 0129. See the following:
This is a single table query, so table join order and join methods are not a
concern. The remaining element to determine is table access method.


There are two filters on this table, as shown on lines 0115 and 0117.
– There is no index to support the filter on Stock.description, so this filter will
not call for an indexed table access method.
– While there may have been an index on Stock.manu_code, this column is
evaluated as an inequality; return everyone from the phone book whose
name is not Mary Alice. That results in a lot of records and not a very
selective filter. If there were an index on Stock.manu_code, it would not be
used to process this filter based on lack of selectivity, (negative of index,
poor selectivity of a filter).
This leaves a table access method of sequential.
A temporary file is required to process the SQL ORDER BY clause.

10.5.4 Rules 4 and 5: Table size and table cardinality
Most often, Rule 3 in the query optimizer determines table order. Even if no filters
are present in the query, table joins are expected to be in the query. The
selectivity (cardinality) of the join relationships will determine table order.

Rules 4 and 5 of the query optimizer are rarely observed. See the following:
If Rules 1 through 3 cannot determine table order, then Rule 4 will be invoked.
(Remember the rules are applied sequentially, until table order is
determined.)
Rule 4 is table size. Assume a two table join with all other factors equal, one
of these two tables will be processed first based on the size of the two tables.
– Which table is read first is based on cost. It used to be true that the larger
table would be processed and have an index built on it to process the join.
Today, the smaller table is likely to be the one processed and have an
index built upon it; because the smaller table produces a smaller index, an
index which is more likely to fit in cache.
– The points above assume that neither table is supported via an index on a
join column, which is most often the case.

Rule 5 is table cardinality. The word cardinality is seemingly used in a variety of
ways in this IBM Redbook. Consider the following:
Technically the term cardinality is defined to mean the number of elements in
a set or group.
In this IBM Redbook, we also discuss the cardinality between two SQL tables;
that is, a one to many style relationship between a parent and then detail
table, and others. Cardinality in this context means the numerical relationship
between two data sets. If each state or province in a nation contains exactly
200 cities, then the cardinality between province and city is 1:200.


In Rule 5 of the query optimizer, the phrase table cardinality refers to the
table’s named order in the SQL SELECT FROM clause. Do not be overly
concerned that table name order is a determining factor in the creation of the
query plan. If a given query is determined by Rule 4 or 5, a number of other
errors exist in the run-time environment, and it is likely that all queries are
performing poorly.

The query optimizers from the various relational database software vendors
operate via a set of guidelines that are based on sound engineering principles.
The rules of a query optimizer are based on observed facts; facts such as
indexed retrieval is generally faster than sequential table access, unless (...).
Each vendor also uses algebraic formulas to assign costs to each operation,
decision, and so on. These exact algebraic formulas are not published, and can
be considered the proprietary and trade secret property of each vendor. Thus far,
we have discussed the elements of these costs and formulas, and the logic and
application behind these ideas. In the next section, we discuss various other
query optimizer features that involve the idea of query costing.

10.6 Other query optimizer technologies
Thus far in this chapter, we have completed the following:
Reviewed a reasonably complex and real world SQL SELECT example as a
means to learn about query optimizers by example. This review included the
most common query optimizer algorithms and tasks.
Then we reviewed the five rules of a query optimizer with a number of
simplistic SQL SELECT examples.
Through all of this, we considered the idea of costing, which is the underlying
foundation of all query optimizer behavior.

In this section, we are going to introduce (or review) various query optimizer
topics including; query rewrite, pipelined sorts, (detailed) query optimizer hints,
and other topics.

10.6.1 Query rewrite
During the query optimization phase, the query optimizer may change the syntax
of the given SQL SELECT to one which is logically equivalent; meaning a query
which returns the same data set, but with different syntax allowing for query
processing in a more efficient manner. Thus far in this chapter, two query rewrite
scenarios have been introduced. At this time, we review these query rewrites and
a few new ones.


Pseudo order by clause
Given a table, t1, with a composite index on column1 and column2, consider the
following:

Since every record returned by this query has a column1 value of 55, this query
could be rewritten as:

All queries in this block are logically equal:
SELECT * FROM t1 WHERE column1 = 55 ORDER BY “55”, column2
SELECT * FROM t1 WHERE column1 = 55 ORDER BY “X”, column2;

These queries are logically equivalent; meaning they each return the same data
set. The query optimizer will detect this condition and choose to use the
concatenated (single) index on columns; (column1, column2). Since the data is
read in collation (index sort order) sequence, no sorting needs to be performed
on these records. They are automatically presorted and cost less to process.

Knowing that the query optimizer observe this conditions, and rewrites the query
for better processing, should the application programmer manually write the
query to explicitly make use of pseudo order by? Ten years ago the answer to
this questions was yes. Today, perhaps nearly every query optimizer should
catch this condition. Manually rewriting queries for optimization is a good idea
certainly, unless their rewrite may make the syntax or purpose of the query
unclear.


Note: Generally, query processing has three logical phases; FJS, filter, then
join, then sort. If the same (single) index used in the filter and join phases can
also serve the group by and order by clause, then the query processor sort
package does not need to be invoked; records are already being processed in
sort order. This is referred to as a pipelined sort operation.

More commonly, queries do their filters and joins, and then have to be gated
(held) and sent to the query processor sort package. This condition is known
as a non-pipelined sort operation.

The first record produced in a pipelined query can be sent immediately to the
application, since records are presorted. A non-pipelined query has to hold all
records, and sort them before returning the first record. This is because the
last record produced by the query during filtering and joining may sort as the
first record to be returned to the application. Pipelined sorts can obviously give
the appearance of a faster query, because of their first row first capability.

Few group by or order by clauses can be pipelined, as the need to filter and
join rarely align themselves with the group by and order by. There are
strategies to increase the likelihood of observing a pipelined query; for
example, ensure that the columns in the order by clause match those of a
composite index used to filter and join. (Add columns to the index to match the
needs of filters, joins and sorts. The column order must match in both the
index reference and order by clause.)

Transitive dependency
Generally, every index contains columns from just one SQL table. Multi-table
indexes, (indexes containing columns from two or more tables) are available in
better relational database servers; these multi-table indexes are sometimes
referred to as star indexes. Multi-table indexes are rare, and are often read only.

Consider the following:

These two queries are logically equal:
SELECT * FROM t1, t2 WHERE t1.col1 = t2.col1 AND
t1.col2 = t2.col2
ORDER BY t1.col1, t2.col2;
SELECT * FROM t1,t2 WHERE t1.col1 = t2.col1 AND
t1.col2 = t2.col2
ORDER BY t1.col1, t1.col2;


The first query above lists columns from two or more tables in the ORDER BY
clause. But, t2.col2 equals t1.col1, as stated in the WHERE clause. Therefore,
the query can be rewritten as shown in the second example. The second query
more clearly displays that a single index can process the order by clause, and
hopefully, will match in an index used above for the join condition resulting in a
pipelined sort.

Be algebraic definition, transitive dependency is represented as,
IF A = B, AND B = C, THEN, BY DEFINITION A = C

In the application of queries, consider the following:

SELECT * FROM t1, t2
WHERE t1.col1 = t2.col1 AND t1.col2 = t2.col2
AND t1.col1 = 65 AND t2.col2 = 14;
WHERE t1.col1 = t2.col1 AND t1.col2 = t2.col2
AND t2.col1 = 65 AND t1.col2 = 14;

The message here is that the query optimizer can use transitive dependency to
consider alternate, but logically equal, query plans. If an index supports t2.col1
and t2.col2 but not both t1.col1 and t1.col2 (the index supports only t1.col1),
transitive dependency may uncover a better query plan.

Semantic equality
-- colPK is the primary key for table t1
SELECT DISTINCT colPK FROM t1;
SELECT colPK from t1;

Because the column list is known to produce a unique product set, either by
unique column constraints or check constraints, the call for the DISTINCT
keyword in the first SQL SELECT above is redundant and may be removed.

How would an example like the first SQL SELECT above ever occur? Very often
user graphical reporting tools allow for operations to be checked (called for),
when the given operations are unnecessary. (The user just checked the visual
control because it was present.)


Another example is shown. Consider the following.

SELECT * FROM t1 WHERE t1.col1 > ALL (SELECT col2 FROM t2);
SELECT * FROM t1 WHERE t1.col1 > (SELECT MAX(col2) FROM t2);

In the second SQL SELECT statement, only one comparison is done for every
record in table t1.

Decoding of SQL VIEWS
Consider the following,
CREATE VIEW v1 AS SELECT * FROM t1 WHERE code = “CB”;
SELECT * FROM v1 WHERE code = “CB”;

One of the tasks of the command parser or query optimizer is responsible for the
decoding of any SQL VIEWS which the given SQL command may reference. In
the example above, a redundant filter criteria is produced; the SQL VIEW already
restricts (filters) the resultant data set as is requested in the SQL SELECT.

Some time ago, early in the technology cycle, SQL VIEWS were executed and
had their results placed into a temporary table before the target SQL command
would be processed. Example as shown:
The following example is no longer current (true)
CREATE VIEW v1 AS SELECT state.*, COUNT(*) FROM states, cities
WHERE state.stateCode = city.stateCode
GROUP BY state.stateCode;

The query above returns the columns from the state table, along with a count of
the number of cities that state contains. Given a query equal to:
SELECT * FROM v1 WHERE stateCode = “IN”;

The query processor would first calculate the data set represented by the SQL
VIEW; that is, it would read, group, and calculate counts for all 50 states. Then
the query processor would read from this temporary table to process the SQL
SELECT, a read of just one state, not all 50.

The above is no longer an accurate representation as to how a modern relational
database server would process the example query. This example is mentioned
as another example of the operations of query rewrite, and the benefit of
decoding SQL VIEWS as part of parsing and query optimization.


Encoding of SQL VIEWS
Generally, SQL VIEWS are logical structures; they represent SQL SELECT
strings that front a number of physical SQL tables. In addition to providing a
macro, or shorthand manner to access a number of tables, SQL VIEWS can also
be used to provide unique security schemes. While SQL security allows the
database administrator to restrict access on the table or column level (vertical
security), SQL VIEWS that returns a subset of records within a table can be used
to provide horizontal security. For example, Bob can only see records from a
subset of states or provinces within a country, not all states and provinces.

Recently, SQL VIEWS have also become physical structures. Some relational
database vendors call these structures materialized views, materialized query
tables, or other similar names. The syntax of a given relational database vendor
for this capability may be SQL CREATE MATERIALIZED VIEW, SQL CREATE
TABLE {name} AS (SELECT ...), or another command. Below is an example of
this type of command, this one is an SQL CREATE TABLE,
Approximate syntax
CREATE TABLE bad_account AS
(SELECT customer_name, customer_id, a.balance cu
FROM account a, customers c
WHERE status IN ('delinquent','problematic', 'hot')
AND a.customer_id = c.customer_id)
DATA INITIALLY DEFERRED REFRESH DEFERRED ) BY DB2;

Since materialized views are real, physical storage structures, the query
optimizer query rewrite feature can use these definitions to reduce processing
load. If a given SQL SELECT calls to read from the native Account and
Customers tables, as listed above, the query rewrite capability redirects the
query to make use of the materialized view, thus saving the associated join
costs.

Rewrite nested queries as joins
Given the following nested query:
SELECT * FROM t1 WHERE t1.col IN
(SELECT colx FROM t2);

One of two conditions will be present in the query above:
The nested (inner) query will return zero or one records; a cardinality of M:1
(many to one).


The nested (inner) query will return more than one record; a cardinality of 1:M
(one to many).

Note: The case of an IN or EXISTS clause is referred to as a semi-join. To
process the (join), the record in the nested table needs merely to exist, it is not
actually attached to the record of the outer table.

If the query above has a cardinality of many to one, then the query can be
rewritten as a standard join. Example as shown:
SELECT t1.* FROM t1, t2 WHERE t1.col = t2.colx;

If the query above has a cardinality of one to many, then a join between these
tables can produce numerous t1 records where previously only one was
produced. The query optimizer will consider the choice of:
SELECT DISTINCT t1.* FROM t1, t2 WHERE t1.col = t2.colx;

Why consider the query rewrite at all? The data set produced inside the inner
query, table t2, may be massive and may force the production of a temporary
table. (Consider a more complex inner query with filters and possibly other table
joins, removing the option of just reaccessing table t2 for every record in table t1,
the outer table.)

Correlated subqueries
The query above was a nested query, (also referred to as a subquery). In short,
the inner query needs to be executed only one time, (before processing the outer
query). The inner query needs to be executed only one time, because the
resultant data set is unchanging; it is the same for every record in the outer
query.

A correlated subquery is a modification to a standard subquery. Correlated
subqueries are prohibitively expensive, although they are a legal operation.

Note: Correlated subqueries are so costly that some relational database
server products will return a special diagnostic code to the application when a
correlated subquery is received. (The thought is that the application may then
want to cancel the request for this operation.)

Listed below is an example of a correlated subquery:
SELECT * FROM t1 WHERE t1.col IN
(SELECT colj FROM t2 WHERE t2.col4 = t1.colz);


Because the inner query makes reference to a column value from the out portion
of the query, (in this case the inner query refers to column t1.colz), each record in
the out portions of the query causes the inner query to be entirely re-executed.

The query optimizer query rewrite capability will attempt to change a correlated
subquery to a standard join. However, in most cases, the correlated subquery
needs to be removed by a modification to the underlying data model (a
modification that we do not discuss here).

Due to query rewrite, below is an example of a correlated subquery which can be
rewritten as a (non-correlated) nested query,

These two queries are logically equal
SELECT * FROM t1 WHERE t1.col1 = 10 AND t1.col2 =
(SELECT MAX (t2.col2) FROM t2 WHERE t2.col1 = t1.col1);
SELECT * FROM t1 WHERE t1.col1 = 10 and t1.col2 =
(SELECT MAX(t2.col2) FROM t2 WHERE t2.col1 = 10);

The example above is also an example of transitive dependency.

OR Topped queries
AND topped query
SELECT * FROM t1 WHERE col1 = 4 AND col2 = 9;
OR topped query
SELECT * FROM t1 WHERE col1 = 4 OR col6 = 15;

The listing above shows two distinct queries; differing goals and differing
resultant data sets. The difference between AND topped queries and OR topped
queries is shown below:
These are both single table queries. The primary task of the query optimizer
is to determine the table access method.
In the case of the first query, a single concatenated index on both t1.col1 and
t1.col2 would allow for an indexed table access method. This is true because
of the AND condition.
In the case of the second query, the OR condition initially prevents the use of
an indexed table access method.
– A single concatenated index on both t1.col1 and t1.col6 will not serve the
query. Because of the OR condition, a record may have a t1.col1 value of


4 or a t1.col6 value of 15. The OR condition makes these two filter criteria
to be separate conditions.
– A single index can serve the t1.col1 filtering needs, or the t1.col6 filtering
needs, but not both.
– A single index can not serve both filter criteria because of the negation of
index, non-anchored composite key. In effect, the OR topped query says,
find everyone in the phone book with the last name of “Schwartz” or with
the first name of “Arnold”. The phone book is not sorted (anchored) on first
name.
One solution to an OR topped query is to perform an OR-union transfer.
Example as shown:
SELECT * FROM t1 WHERE col1 = 4
UNION
SELECT * FROM t1 WHERE col6 = 15;
A UNIONED SELECT allows for two or more distinct SQL SELECTS to be
executed as though they are just one simple SQL SELECT. These two or
more SQL SELECTS can be from the same or different tables. About the only
requirement is that each participating SQL SELECT must return the same
number of columns, with compatible data types (numeric to numeric, and
character to character).
An OR-union transfer should be considered when the following is observed:
– Each participant in the transfer makes use of an indexed table access
method.
– The final query runs faster.
By its syntax, a UNIONED SELECT will remove duplicates from the final data
set; that is, from the example above, a record may be both t1.col1 = 4 AND
t1.col2 = 15. However, that (single) record cannot be returned to the
application twice just because of a query rewrite operation.
A UNION ALL SQL SELECT will not call to remove duplicates from the
participating queries and, therefore, is more efficient than a standard
UNIONED SELECT. IF it is determined that each participating query
produces a distinct data set, they by all means issue the query as UNION
ALL.

Note: One question is whether or not a relational database supports a
multi-index scan. That is, can the relational database natively use two or
more indexes as a table access method for one table? And, under what
circumstances is it allowed? Multi-index scan needs to have the ability to
remove duplicate records from the (single) source table.


Redundant join elimination
SELECT * FROM t1, t2, t3 WHERE t1.col = t2.col
AND t2.col = t3.col AND t3.col = t1.col;

There are redundant join criteria in the example SQL SELECT above. Since t1
joins with t2, and t2 joins with t3, the query processor does not need to join table
t3 to table t1. The query rewrite capability of the query optimizer will consider the
table t3 to table t1 join path, but will remove at least one of the join paths since it
is unnecessary.

Shared aggregate calculation
The first query has 2 SUM() and 1 COUNT()
The second query, produced by the query optimizer
query rewrite, has 1 SUM() and 1 COUNT()

Remember that an AVG() is a SUM() / COUNT)()

Original query
SELECT SUM(col) AS s, AVG(col) AS s FROM t1;
Query rewritten by query optimizer
SELECT s, s / c FROM (SELECT SUM(col) s, COUNT(col) c
FROM t1);

The query optimizer query rewrite capability will detect the shared aggregation
opportunity and rewrite the first query above, as displayed in the second query
above. This example also displays an optimal application of dynamic table
reference in the FROM CLAUSE to a SQL SELECT. This technique can be used
extensively in business intelligence style queries.

Operation movement
SELECT DISTINCT * FROM t1, t2 WHERE t1.col = t2.col;

If the column t2.col is known to be unique (it has a unique index constraint placed
upon it), then the DISTINCT set operand can be performed on table t1 alone.


The DISTINCT set operand need not be performed on the larger (wider) data set
which is the product and tables t1 and t2 having been joined.

In this manner, DISTINCT set operands can be pushed lower into the execution
of the query plan, or higher; whichever results in a lower total query cost.

Predicate translation
Code Fragment 1
WHERE NOT (col1 = 10 OR col2 >3);
Code Fragment 2
WHERE col = YEAR(“1990-04-01”);

The first example can be rewritten as:
WHERE col1 <> 10 AND col2 <= 3;

This query optimizer query rewrite operation removes the OR topped nature of
the query and increases likelihood of an indexed table access method.

The second example is rewritten as:
WHERE col = 1994;

This rewrite removes the function invocation on every record that is processed;
further, the result of this given function was an integer data type which is more
efficient than the DECIMAL(16) or structure that often represents SQL
DATETIME column types.

SELECT * FROM t1 WHERE integerColumn = “123”;

The query optimizer rewrite capability translates a literal expression into an equal
data type whenever possible; “123” is a numeric data type and is converted to an
INTEGER to facilitate the join with greater efficiency. The query optimizer query
rewrite capability will encrypt literal value filter criteria to compare with encrypted
column values when the encryption key is set at the table level. This operation is
impossible for systems or applications that encrypt column values at the record
level.

Note: Most encrypted data types take 33% or more disk storage to record
than non-encrypted data. Omitting the password hint that accompanies
individual column values can eliminate as many as 50 bytes per record.


10.6.2 Multi-stage back-end command parser
Technically, in this entire chapter we discuss the query optimizer (which also
implies discussion of the query processor). The multi-stage back-end is headed
by a command parser, and there are certain technologies you can employ at the
parser phase that will also improve SQL performance.

Bypassing the command parser
Mixture of application program code and SQL
FOR i = 1 to 1000
INSERT INTO t1 VALUES (i);
END FOR

The program code fragment above contains a mixture of application program
code and embedded SQL commands, as many SQL based applications do. On a
given system with a given performance, imagine that the program loop above
executes in 60 seconds. Now consider the following:
Mixture of application program code and SQL
PREPARE s FROM “INSERT INTO t1 VALUES (?)”;
FOR i = 1 TO 1000
EXECUTE s USING i;
END FOR

While the first loop executes in 60 seconds, the second loop may execute in as
little as 12 or 20 seconds. Why this is occurring is detailed in the list below:
The application program will contain a mixture of program code and SQL
statements. As a given SQL statement is encountered, this SQL statement is
passed to the relational database server for processing via the agreed upon
communication protocol. Only SQL commands are passed to the relational
database server. The relational database server has no knowledge of
non-SQL program code. The relational database server has no idea, for
example, that the program has entered a looping construct and is about to
execute the very same command 1000 times in a row.
In the first loop listed above, each of 1000 SQL INSERT commands is sent to
the SQL command parser as though the database server has never spoken
to the application before. Each command has the following performed:
– The statement is parsed for syntax errors.


– Any referenced tables and columns are checked to see if they exist, and to
see if the given user (or role) has permission to perform the requested
operation.
– Any semantic (data) integrity constraints are retrieved, as well as any
triggers, check constraints, default column constraints, primary and
foreign key definitions, and others.
– And more.
Whether the loop moves from 60 seconds to execute to 12 seconds or 20, is
dependent on the amount of time that was spent in the command parser, or
was due to communication overhead.
The command parser can be bypassed second and subsequent times, by
registering SQL commands with the command parser for future use.
– The SQL PREPARE statement registers the given command with the
command parser.
– At this point, the original INSERT is invoked via the SQL EXECUTE
statement to the proper statement id.
– The location of host variables is preserved via the question mark place
holders, and then filled via the USING clause on execute.

Note: Most modern relational database server products offer SQL statement
cache; that is, a listing of the 100 or so most recently received SQL
commands. Given this feature, is it still necessary to use SQL PREPARE for
optimization? Yes. Any newly received SQL command has to be parsed and
compared to those which are resident in the cache. The SQL PREPARE
statement bypasses this operation also.

SQL INSERT cursors (buffered inserts)
Most application programmers encounter an SQL CURSOR the first time they
need to execute a SQL SELECT statement that returns more than one record
from the database server. In this case, the SQL CURSOR manages the transport
of the multi-row result set from the database server to the application client. The
SQL CURSOR is like a gondola transporting records from the bottom of the
mountain to the top. But, gondolas go both directions. SQL CURSORS can also
be used to send records from the application to the database server.

PREPARE s FROM “INSERT INTO t1 VALUES (?”);
DECLARE c CURSOR FOR s;
OPEN c;


FOR i = 1 to 1000
PUT c USING i;
END FOR
FLUSH c;
CLOSE c;

An SQL INSERT CURSOR is one that fronts a SQL INSERT statement, as
opposed to the standard (expected) SQL SELECT statement. An SQL INSERT
CURSOR uses the SQL command verbs, PREPARE, DECLARE, OPEN and
CLOSE, much like a standard SQL SELECT CURSOR. New commands include
PUT and FLUSH.

All SQL CURSORS are specified to be a given size, which determines how many
records it can contain. The SQL PUT verb places records in a buffer on the
application client side of this Client/Server topology. As the communication buffer
associated with this SQL CURSOR fills, the buffer is automatically flushed;
meaning, that all records are sent from the client to the server at once. This is
how a SQL INSERT CURSOR achieves its performance gain; in effect, the SQL
INSERT CURSOR reduces the number of transports between the client and the
server. Because this communication buffer may be left partially full after the
completion of the looping construct, the buffer is manually flushed to ensure that
the last, potentially partially full buffer is sent to the server.

The most difficult aspect of using SQL INSERT CURSORS is the error recovery.
When records are sent one at a time, error checking is rather simple. When
records can be sent variably and in larger numbers, error recovery is just a bit
harder, but still very manageable.

From the example above, the original 60 second SQL INSERT loop could be
moved to two or three seconds of execution time, an admirable gain.

Note: Not every relational database vendor supports SQL INSERT
CURSORS, or they may use different terminology and commands to
accomplish the same effect.

Multi-statement prepare
Similar to the first example of bypassing the command parser above, prepared
statements can actually contain numerous SQL commands. Consider the
following:
PREPARE s FROM “INSERT INTO t1 VALUES (?);
UPDATE t2 SET (col4) = (?) WHERE col1 = (?);”


EXECUTE s USING v1, v2, v3;

In the example above, two separate SQL commands are expected to be
executed together; that is, the application always needs these two distinct
statements to run one after another. A reduction in the number of
communications to the relational database server can be observed by preparing
a multi-statement.

More on SQL statement cache
SQL SELECT statements that return more than one record, must be managed
with a SQL CURSOR. The SQL SELECT associated with the SQL CURSOR is
then registered with the command parser, and second and subsequent uses of
this SQL CURSOR will bypass the command parser. SQL SELECT statements
that do not return more than one record would include:
Basically select only aggregates with no GROUP BY:
SELECT COUNT(*), MAX(col), MIN(col) FROM t1;
Select on equality to primary key column:
SELECT * FROM t1 WHERE colPK = 17;
SELECT * FROM t1 WHERE colPK = 19;

Note: Of the two SQL SELECTS shown above, use of the first SQL SELECT
is preferred. While the primary key on table t1 may be colPK today, that may
change. This often happens when two companies merge; the new primary key
in the database is then pre-appended with a code for company, origin, or
similar. While this may seem an unlikely event, the task of having to recode
(port) thousands or millions of lines of SQL program code is not desirable.

From the perspective of the SQL statement cache subsystem, the last two SQL
SELECT statements are unique, and do not allow for statement reuse (the ability
for the statement to be recalled from the cache). We recommend you instead use
host variables as shown in the following example:
Host variable H set before SQL invocation
SELECT * FROM t1 WHERE colPK = :H ;

10.6.3 Index negation
In general, index negation refers to the condition of having a perfectly useful and
on line (pre-existing) index data structure, but for some reason that index may
not be used to process a filter, join, or other task. Consider the following:
Negation of index, non-initial substring; given the example,


SELECT * FROM phoneBook WHERE lastName LIKE “%son”;
If an index exists on the phoneBook.lastName column it cannot be used
because of a non-initial substring condition. Imagine trying to find every
person in the phone book where the last name ends in “son”, such as
Hanson, Johnson, and Williamson. While the phone book is sorted by last
name, it is not sorted by the last three characters of the last name. You would
have to read the phone book sequentially and in its entirety to satisfy this
query.
Negation of index, non-anchored composite key. Consider the example:
SELECT * FROM phoneBook WHERE firstName= “Chuck”;
While the phone book is sorted by the concatenation of last name, then first
name, it is not sorted by first name alone. If you need to find every person in
the phone book by first name, you have to read the entire phone book.
The only exception to a non-anchored key condition is when all columns in
the column list are members of the concatenated index. The phone book
example fails to carry this point. Since an index is a vertical slice of a table, (it
contains a subset of columns and all rows), the non-anchored read of just first
name is better processed by sequentially scanning the (last name, first name)
index; better than sequentially the SQL table proper which is larger (wider).
Negation of index, (poor) selectivity of filter. Consider the examples:
SELECT * FROM phoneBook WHERE lastName > “Ballard”;
SELECT * FROM phoneBook WHERE lastName != “Comianos”;
Both of these queries will return 90-98% or more of the phone book. Better to
suffer through sequentially scanning the phone book, than to use the index
with the physical disk I/O that incurs, and still return 98% of the phone book.
The NOT often removes any selectivity of a filter of join criteria, by the nature
that it too negates.

Partial index negation
A query similar to that shown below will cause some vendor relational databases
to use only the leading column of a concatenated index. For example:
SQL CREATE INDEX of concatenated (multi-column) index:
CREATE INDEX i1 ON t1 (col1, col2);

SELECT *
FROM t1
WHERE col1 > 50


AND col2 = 4;

Because of the range operand on the leading column in the concatenated index,
the index would be used to process the filter criteria on column col1 (if selectivity
of this filter calls for this operation). Then the data page from the SQL table
proper is fetched. Then the filter criteria for column col2 was processed from the
data page. This behavior can be detected by reviewing the index and filter
utilization in the text representation of the query plan. Basically not having the
capability was just a lack of functionality in the relational database software
product of a given vendor.

10.6.4 Query optimizer directives (hints)
Most modern relational database vendors offer the ability to influence the query
plan that the query optimizer produces. Some vendors offer the ability to promote
(call for, or force) certain elements of the query plan, or just demote (prevent)
certain elements, or the ability to both promote and demote. Most queries that
you will encounter which require manual tuning (creation of optimizer directives)
will need help because the query plan is making a poor choice. We recommend
that you use query optimizer directives to demote poor query plan paths. Do not
fall into the practice of writing query optimizer directives that promote the
(currently) more optimal path. As the run-time environment changes (improves),
new more optimal query plan choices may arrive.

Note: In the query optimizer case study in 10.2, “Query optimizer by example”
on page 503, we reviewed a canned application system; that is, one where the
source code to the application may not be accessible to us (the database
administrators). Some relational database vendors offer the ability to set query
optimizer directives in newly created files (or system catalogs) which are
external to any canned application system. Needless to say, this is a very
useful feature when supporting canned applications.

You can also embed query optimizer directives inside SQL VIEW definitions;
perhaps leave the base table in place, and create a SQL VIEW entitled,
“v_faster_table”, which accesses a given table while demoting poor query plan
choices.

Query optimizer directives can be organized into categories which, not
surprisingly, equal the basic categories of operations that a query optimizer
considers. These types of optimizer directives include:
Table join order
Table access method
Table join method
Optimization goal


Query rewrite directives

Generally, query optimizer directives are embedded inside an SQL command
statement as comments; in other words, the query optimizer directives are not
part of the declarative portion of the SQL command. Consider the example:
SELECT --+ This is a single line comment, begins with “--+”
* FROM t1;
SELECT { This is a multi line comment,
extending for multiple lines and surrounded by
curly braces }
* from t1;
SELECT /* This is also a multi line comment,
surrounded by Java and C language style markers */
* FROM t1;

While different relational database vendors may use differing and specific query
optimizer directives syntax, below is a representative offering of the standard
tags and their associated effect.

ORDERED, INDEX, AVOID_INDEX, FULL, AVOID_FULL
A call to “ORDERED” states that the table join order should equal that as stated
in the SQL FROM clause of the SQL SELECT statement. “FULL” states that the
named tables, enclosed in parenthesis should use a table access method of
sequential scan, while “AVOID_FULL” states the opposite. “INDEX” and
“AVOID_INDEX” do the same for indexes. Example as shown:
SELECT --+ ORDERED, AVOID_FULL(e), INDEX(i2)
*
FROM employee e, department d
WHERE e.deptno = d.deptno AND e.salary > 100000;

The query optimizer directive above states that the SQL tables should be joined
in order of Employee to Department. Also a call is made to use the index entitled
“i2”, whichever table maintains that index. If the named index does not exist, that
specific query optimizer directive will be ignored.

Multiple entries inside the parenthesis are separated with commas. If necessary,
the table names should be prefaced with the owner or schema name.


Combination of both “AVOID_FULL” and “AVOID_INDEX” on a given SQL table
can be used to encourage a hash join. (Remember we prefer to demote bad
choices, not explicitly promote the currently best choice.)

INDEX_ALL
“INDEX_ALL” is the query optimizer directive to invoke multi-index scan. For
example:
OR topped query, would benefit from multi-index scan
SELECT --+ INDEX_ALL (ix_on_col1, ix_on_col2)
*
FROM t1
WHERE col1 = 5 OR col2 = 15;

USE_NL, AVOID_NL, USE_HASH, AVOID_HASH
“NL” is an abbreviation for the nested loop (table) join method. Each of these
query optimizer directives are for table access methods. (The current vendor we
are reviewing did not document query optimizer directives for a sort-merge table
join method, or a push down semi-hash join. This vendor did offer other controls
to promote or demote these table join methods, but they are not documented
here.)

In the case of a hash table join method, an SQL table must be identified as the
SQL table upon which the hash index is built, or should it be the table to access
the hash index execute to join? The “BUILD” label marks the table upon which
the hash index is built. The “PROBE” label marks the table that reads the hash
index. For example:
Hash index will be built on Dept table
SELECT --+ USE_HASH ( dept /BUILD )
name, title, salary, dname
FROM emp, dept, job
WHERE loc = “Almaden”
AND emp.deptno = dept.deptno
AND emp.job = job.job;

This vendor also has a “/BROADCAST” label for a hash join used in a clustered
database environment. A call to “/BROADCAST” will copy tables built with hash
indexes under 100 KB to all participating nodes in query. This is a useful feature
to place (copy) smaller dimension tables to all nodes, thus encouraging
co-located joins. (Co-located join is a clustered database concept. In effect,


co-located joins are those where both sides of the join pair are local to one
(physical) node. This eliminates or reduces internodal traffic for joins.)

“USE_NL” or “AVOID_NL” call for the nested loop table join method to be used,
or not used, with regards to the named tables.

FIRST_ROWS, ALL_ROWS
Operations such as creating a hash index to access a table, or sorting a given
table to allow for a sort-merge table join method, require an amount of up front
query processing that is time spent before any records may be returned to the
user application. A “FIRST_ROWS” query optimizer directive, will avoid query
plans with these types of up front costs, and instead give preference to query
plans that return records more quickly, if only initially (first record, versus total
data set performance).

From the discussion on pipelined sorts versus non-pipelined sorts, held earlier,
this category of query optimizer influencer will not be viewable to the user
application if records have to be gated (held), to complete a sort operation.
Under these conditions, the query optimizer may negate this specific query
optimizer directive.

Following are the results a query received with no query optimizer directives:
This output has been edited
WHERE t1.col = t2.col
Estimated Cost: 125
Estimated # of Rows Returned: 488
1) Abigail.t2: SEQUENTIAL SCAN
DYNAMIC HASH JOIN
Dynamic Hash Filters: Abigail.t2.col = Abigail.t1.col

This same query is then executed with a query optimizer directive of
“FIRST_ROWS”. The following are the results:
This output has been edited
SELECT --+ FIRST_ROWS
* FROM t1, t2
WHERE t1.col = t2.col
Estimated Cost: 145


2) Abigail.t2: INDEX PATH
(1) Index Keys: col
Lower Index Filter: Abigail.t2.col = Abigail.t1.col
NESTED LOOP JOIN

We see that the cost is higher for the second query, because of the (CPU)
intensive nested loop (table) join method. While the cost is higher on the second
query, and the total data set performance may take more time, the first record
returned will occur measurably faster than that in the first query, which must first
build a hash index.

NESTED
By default, the query optimizer will attempt to eliminate nested queries
(subqueries) and correlated subqueries. The query optimizer directive “NESTED”
disables that capability. “NESTED” is one query optimizer query rewrite directive.

Optimizer directives not shown
While query optimizer directives are measurably complete, there are still a few
query plan-related decisions that are not set via query optimizer directives. As
examples:
Whether a sort operation happens entirely in memory or overflows to disk is a
function of the amount of shared or program memory that is allocated to the
query. This, generally, is not set via query optimizer directives.
The degree of parallelism is determined automatically by most vendors based
on:
– Whether records are locked as part of this query. When locking, most
vendors will demote parallelism to prevent locking too many records
concurrently, and hurting overall system throughout and concurrency.
– The number of source and target drives. If you have 14 CPUs, but all of
the tables are being read from one hard disk, how much parallelism do
you actually require? If you are performing an unload of data to a single
ASCII text file, how much parallelism do you actually require?
– And as implied above, if you only have one CPU, how much parallelism do
you require beyond parallel table scans?
– The presence of SQL SELECT TRIGGERS can prevent not only
parallelism, but also the transfer of numerous records per each buffer
transport inside the SQL CURSOR.


– The number of database disk drives designated (allowed) for sort
operations; SQL SELECT DISTINCT, SQL SELECT ... UNION ..., SQL
GROUP BY, SQL ORDER BY, hash index creation, and others.

Example query plan with query optimizer directives
This file has been edited for clarity:
SELECT { +ORDERED,
INDEX ( emp ix1 ),
FULL ( job ),
USE_HASH ( job /BUILD ),
USE_HASH ( dept /BUILD ),
INDEX ( dept ix3 )
}
j.*, d.*
FROM emp e, job j, dept d
WHERE e.location = 1
AND e.jobno = j.jobno
AND e.deptno = d.deptno
AND d.location = “MILWAUKEE”

DIRECTIVES FOLLOWED
ORDERED
INDEX ( emp ix1 )
FULL ( job )
USE_HASH ( job /BUILD )
USE_HASH ( dept /BUILD )
INDEX ( dept ix3 )
DIRECTIVES NOT FOLLOWED:

Estimated Cost: 963


1) Abigail.emp: INDEX PATH
Filters: Abigail.emp.location = “Milwaukee”
(1) Index Keys: empno jobno deptno location (Key-Only)

2) Abigail.job: SEQUENTIAL SCAN

DYNAMIC HASH JOIN
Dynamic Hash Filters: Abigail.emp.jobno = Abigail.job.jobno

3) Abigail.dept INDEX PATH

(1) Index Keys: location
Lower Index Filter: Abigail.delp.location = “Milwaukee”

DYNAMIC HASH JOIN
Dynamic Hash Filters: Abigail.emp.deptno = Abigail.dept.deptno

Additional comments:
With the ORDERED query optimizer directive, table order is equal to that
listed in the SQL FROM clause; Emp, then Job, then Dept.
Hash indexes are built on Job and Dept as called for in the USE_HASH()
directive.
With no filter on table Job, the hash index is built on every record in the table;
therefore, the sequential table access method. (This does not qualify as a
second table being sequentially scanned inside of one SQL SELECT, since
this table access method is used only to build the hash index.)
An index is used to build the hash index on table Dept, as called for in the
“INDEX ( dept ix3 )” directive.
The table access method on table Emp is indexed, for which it was called.
This access method is also key only, which the query optimizer detected the
opportunity to do so. (All necessary columns from table Emp were part of a
usable index, the index used to access that table.)


10.6.5 Data distributions
Data distributions was presented in “Data distributions” on page 524. Here we
present additional information.

Filter (predicate) correlation also known as predicate inference
Query one, distinctiveness of filters can be multiplied
SELECT * FROM autos WHERE make = “CarMaker” and model =
“CarModel”;
Query two, distinctiveness can not be multiplied
SELECT * FROM addresses WHERE cityName “Smithtown”
AND streetName = Main”;

In an Autos table with 100,000 records, 10% of which are CarMakers, and only
CarMaker makes the CarModel (one of six models made by CarMaker), then the
filter on table Autos can be expected to return 1/10 times 1/6 of 100,000 records.
These two filters have a strong correlation.

In an Addresses table with 100,000 records, and 10% of the cities are named
“Smithtown”, and 1/6 of the streets are named “Main”, one should not make any
assumption about the correlation of these two filters; nearly every town in
America has a street named “Main” (or “High” street if you prefer).

This is an area where data distributions can help. Data distributions can be
gathered on indexed or non-indexed columns, individual columns, or THEN
groupings of columns. Generally, the query optimizer will assume that it can
multiply the selectivity of predicates, as shown in the first SELECT example
above with Autos. A data distribution on both the cityName and streetName
columns in Addresses would correctly inform the query optimizer that it should
not consider adding the selectivity of these two filters.

Other applications of data distributions
By default, most query optimizers know the number of records in a table, and the
number of distinct values in an index. Given a table with 1000 records, and a
unique index, that index contains 1000 distinct values (by definition). The
optimizer also knows the number of distinct values in a duplicate index. (The
optimizer also knows the minimum and maximum values of each index, as well
as the second minimum and second maximum.)

Most query optimizers, by default, assume an even distribution of key values
within a given range. That is, it assumes that an equal number of persons exist
on every continent (land mass) on planet earth. Given the following:


SELECT * FROM worldPopulation
WHERE continent IN (“Antarctica”, Australia”);
SELECT * FROM worldPopulation
WHERE continent in (“Asia”, “Europe”);

The query optimizer would, by default, consider the selectivity of these two
queries to be equal. Since the second query returns a very large percentage of
the entire contents of the Continent table, this query would be better served by a
sequential scan table access method (negation of index, poor selectivity of filter).
The first table should use the index.

Data distributions inform the query optimizer of the selectivity of specific values in
an indexed or non-indexed column. This additional information helps the query
optimizer choose better query plans.

Another application of data distributions is for concatenated indexes. By default,
the query optimizer knows the uniqueness of the entire index (or perhaps only a
small number of leading columns). Not having this data was one contributor to
the error in the query plan reviewed in 10.2, “Query optimizer by example” on
page 503.

Note: Some relational database software vendors count every distinct value in
a column, or samples therein. Some vendors organize this counting into
distinct ranges, sometimes called quantiles.

The selectivity of filters in general, is calculated using guidelines similar to those
below (selectivity is represented by the value (F), and EXPRESSIONS are
represented by the bulleted condition):
indexed_column = LITERAL_VALUE
indexed_column IS NULL (which is a a literal value)
F = 1 / (number of distinct keys in index)
t1.indexed_olumn = t1.indexed_column
F = 1 / (number of distinct keys in the larger index)
indexed_column > LITERAL_VALUE
F = (2nd-max - LITERAL_VALUE) / (2nd-max - 2nd-min)
indexed_column < LITERAL_VALUES
F = (LITERAL_VALUE - 2nd-min) / (2nd-max / wnd-min)
non_indexed_column = LITERAL_VALUE
F = 1 / 10


non_indexed_column > LITERAL_VALUE
F=1/3
non_indexed_column LIKE {STRING EXPRESSION}
F=1/5
EXISTS {subquery}
F = 1, if subquery estimated to return more than zero rows, else 0
NOT {EXPRESSION}
F = 1 - F {EXPRESSION}
EXPRESSION1 AND EXPRESSION2
F = F ( EXPRESSION1 ) union F ( EXPRESSION2 )
EXPRESSION1 OR EXPRESSION2
F = F ( EXPRESSION1 ) + F ( EXPRESSION2 ) -
( F ( EXPRESSION1 ) union F ( EXPRESSION2 ) )
non_indexed_column IN LIST
Treated as an OR, OR, OR for each item

The primary purpose of offering the list above is to demonstrate the coarseness
of query optimizer statistics regarding selectivity of filter and join criteria without
data distributions.

10.6.6 Fragment elimination, multidimensional clustering
For all that SQL provides, it does little in the area of the physical storage of data;
in other words, how tables and indexes are physically organized on the hard disk.
SQL tables and indexes are logical structures, that, at some point, become
physical bytes on a collection of disk pages. How are these pages to be
organized?

The earliest relational database server products placed both (normal) table data
and index data inside the same physical disk space allocation. Modern relational
database server products place normal table in a physical allocation by itself, and
then each individual index in its own separate physical disk allocation. This
newer convention allows for like data to be contiguous; that is, index data from a
given index is contiguous, allowing for read ahead scans. The data from the table
proper is also contiguous, also allowing for read ahead scans. And all of the
behavior increases the physical I/O cache rates.

Records in the SQL table proper are allowed to be in random sequence. While
not explicitly stated, this is a base assumption to the governing of SQL tables.


Most vendors offer a clustered index operation; that is, generally in a one time
only operation, the data in the table is physically sorted (written) in the collation
sequence of one of the indexes found to exist on that table. (A given list of data
can only be truly sorted in one manner, based on one hierarchy.)

Splitting a tall table
Given a table with three years of data, but having the majority of queries
executed against data created in the last 90 days, creates an interesting
opportunity. Many years ago, the database administrator would implement a
design pattern, referred to as splitting a tall table (tall, meaning having many
data records).

When splitting a tall table, one table contains the data from the last 90 days. As
these records age, they are deleted from the 90 days table and inserted into the
history table. Queries targeting the last 90 days of data read from just one table.
Queries that need data from the full date range execute a SQL UNION ALL
query. For example:
UNION automatically removes duplicate records, but UNION ALL does
not. Since records are either in one table below but not both, we can
execute a UNION ALL query which is more efficient:
SELECT * FROM recentTable
UNION ALL
SELECT * FROM historyTable;

You can instruct modern relational database servers to perform this type of
activity automatically.

Table partitioning
The original phrase given to the practice table partitioning was fragmenting a
table. Generally, and in the area of disk storage, fragmentation is a term with
negative connotations. Still, once a phrase is released it is often hard to
recapture. The relational database industry calls this idea both table
partitioning and fragmenting tables.
With table partitioning, a table can be physically organized by the value of a
column or set of columns. These columns are expected to be those which are of
importance to critical queries. For example:
Sample SQL CREATE TABLE and SELECT:
CREATE TABLE t1
(
col1 INTEGER,


col2 INTEGER
)
FRAGMENY BY EXPRESSION
col1 < 100 IN driveSpace1,
col1 >= 100 IN driveSpace2,
col1 >= 200 IN driveSpace3;


A “driveSpace” is a non-SQL object. In this case, driveSpace is the concept of a
distinct disk storage location. Relational database vendors call these entities
chunks, volumes, containers, and perhaps other terms.

Given the SQL CREATE TABLE statement above, the relational database server
automatically places records in a specific driveSpace determined by the value in
col1. When the query is received as shown above, the relational database server
(the query optimizer) can perform fragment (partition) elimination; that is, we do
not need to access any records in “driveSpace1” or driveSpace3”, because they
contain no records with a col1 value of 150.

These are not indexes. This is similar to a free and omnipresent natural index.
You can automatically reduce the amount (table size) of the records to
processed based on how the table is organized on the hard disk.

Most vendors allow data in a table to be partitioned, as well as partitioning each
index. See the following:
Partitioning is allowed on numerous columns, not just one column.
There are expression operands to determine driveSpace, as well as hash and
others.
Specifying expression operands can be wordy, but they serve both filter
equalities and ranges. Hash value partitioning does not serve range filters
well at all since the column values are essentially rotated through the
driveSpaces.
Under the best circumstances, these partitions can be added and removed
from the table in full or nearly full multi-user mode. Since unique indexes have
to provide a unique data integrity constraint, often these are the culprit as to
whether these partition attach operations are zero or low cost.

The query optimizer automatically calculates which table partitions have to be
examined or not, and these costs are considered in the determination of the final
query plan. This basic operation is referred to as fragment elimination.


Given table partitioning and an SQL CREATE table as shown below, consider
the following:
CREATE TABLE t1
(
col1 INTEGER,
col2 INTEGER
)
FRAGMENY BY EXPRESSION
col1 < 100 AND col2 < 100 IN driveSpace1,
col1 < 100 AND col2 >= 100 IN driveSpace1b,
col1 >= 100 AND col2 < 100 IN driveSpace2,
col1 >= 100 AND col2 >= 100 IN driveSpace2b,
col1 >= 200 AND IN driveSpace3;


Given the above run-time environment, the query optimizer must call to examine
driveSpaces 1b, 2b, and 3, to find records with a col2 value of 150.

Multidimensional clusters
One vendor in particular has a different take on table partitioning. It was stated
earlier that any given data set can have only one concurrent sort order. This is
true. Multidimensional clusters are like a multi-level table partitioning
implementation. Multidimensional clustering uses specific column values to
automatically store records in distinct drive areas. A multidimensional cluster is
created with syntax similar to:
CREATE TABLE t1
(
col1 INTEGER,
col2 INTEGER
)
ORGANIZE BY (col1, col2);

The relational database server will automatically create and enforce a distinct
disk space allocation for each unique combination of col1 and col2 values. That
means that each of the data pairs (1,1), (1, 2), and (2, 2) go into their own distinct
disk space allocation. (Three separate disk space allocations.)


Note: The term dimensional in the phrase multidimensional clusters has no
relationship to the phrase dimensional data modelling.

Standard table partitioning is viewed by some to be too much work, with its
precise syntax determining record location by value expression or hash.
Multidimensional clustering involves much less work, but better serves columns
with few distinct values, since each distinct value causes a new disk space
allocation.

Table partitioning, SQL VIEWS, and query rewrite
Consider the following SQL VIEW:
CREATE VIEW v1 (col1, col2, col3, col4)
AS
SELECT a, b, c, d
FROM t1
WHERE ...
UNION ALL
SELECT e, f, g, h,
FROM t2
WHERE ... ;

Given any data integrity check constraints or table partitioning expressions that
may be in place on these underlying tables, the query optimizer query rewrite
feature may remove specific filter criteria from any SQL SELECT accessing this
view, since these filter criteria are known to be redundant (unnecessary).

10.6.7 Query optimizer histograms
Query optimizer histograms are the idea that the query optimizer could form a
query plan with its associated cost, and record this plan and expected cost for
later reference. If the given query plan did not perform as expected, then the
query optimizer might later experiment and try a different query plan to see if
query performance can improve. This is not artificial intelligence since no new
algorithms are being created. This is merely the query optimizer recording costs
to add to that data which it considers in the creation of a query plan.

While query histograms might someday be delivered, thus far they have proven
too costly to include in current relational database server products. Having also
to consider histogram data in the calculation of a current query plan takes


additional time and resource. Without histograms, the current query optimizer
makes the most optimal choice 98% or more of the time, it seems. If not, you still
has query optimizer directives to rely upon.

10.7 Summary
In this chapter we have presented significant detailed information about the topic
of query optimizers. This can help you understand how to create better data
models and how to better format and store your physical data. You are working
to “get your data in” to the data warehouse, but doing so while you keep in mind
that it must be done in such a way that enables you to maximize you capability to
efficiently and effectively “get your data out” of the data warehouse to support
your BI environment.

To get started on that path, this is an excellent time to continue on with
Chapter 11, “Query optimizer applied” on page 599. Here you will find help how
to put the knowledge from this chapter into action. So, on you go.


11

Chapter 11. Query optimizer applied
In Chapter 10, we reviewed a business application system to analyze the SQL
statements used. In that application, there was a query with an SQL SELECT
statement that executed in five minutes, when the requirement was for
sub-second response time.

This is a situation typically faced by application developers. In this chapter we
offer direct and specific techniques that you can employ to help you avoid issues
such as this.

The query and the SQL statement, were analyzed and the problems were
recognized and changed to meet the application requirement. This example is
from an actual application system analysis, where it took the application analyst
two hours to solve this problem. The following are comments about some of the
issues encountered:
It took two hours to solve this problem, in part because we had never seen
this exact problem before. As it turns out, we saw this same condition weeks
later, and were able to solve the problem in less than thirty minutes.
– While we solved this problem relatively quickly the second time we saw it.
But, there are always new problems.
– In Chapter 10, we stated that a database administrator with six months of
experience can performance tune a single (non-clustered, non-distributed)
relational database server system with 500 GB of disk storage and 8
CPUs in well under two hours. Performance tuning relational database


server systems is a generally well understood process and is typically well
documented.
– The problem in the example is not one of relational database server
system tuning, it is one of SQL statement tuning. While it takes two hours
to tune a server, it regularly takes two hours to tune just one poorly
performing SQL statement. In this example, there were dozens of SQL
statements that were performing poorly. That means dozens times two
hours.
In our example, we needed to change the indexes on the table to solve the
problem. Many times, however, the structures of the tables have to be
changed to solve the problem. And, changing table structures often means
application program code also has to change, unit test and then system tests
have to be performed, and so on. The net result is that solving SQL statement
tuning problems is very costly and takes a lot of time.

The techniques we describe in this chapter should better enable you to deliver
performant SQL-based business application systems on time and at lower cost.
They will also improve your ability to maintain previously developed SQL-based
business application systems.


11.1 Software development life cycle
A software development life cycle (SDLC) is defined as the methodology you
uses to develop, support, deliver, and maintain computer applications. As a
phrase, SDLC is logical; it defines a categorization (of physical methodologies).
One specific (and perhaps the most common), SDLC methodology is the
Waterfall Method. Figure 11-1 displays a Waterfall Method SDLC.

Data Admin Developers

Incorrect
400 Table Mappings of
Data Model Legacy Data

Coding and
1. Discovery 2. Data model 3. debugging 4. System test 5. Cut over

Missing Columns,
Errors in Schema
Errors in Performance
Dependencies Issues

Figure 11-1 Example Waterfall method (SDLC)

While any specific representation of a Waterfall Method SDLC might display five,
seven, or another number of phases, each representation shares these basic
ideas:
There are iterative steps that are performed in each phase. An error or other
conditions can cause you to revisit, or at least partially revisit, prior phases.
Generally, you move from:
– User discovery: Gathering the requirements of the application:
• Generally, the product of the user discovery phase is a requirements,
costing, and staffing document of some form.
• Persons involved include business area leaders (persons who
understand and can articulate the business goal to be achieved with
the new system), project managers, data analysts, operating system
administrators, network administrators, and perhaps others.

Chapter 11. Query optimizer applied 601

– Data modeling: Includes both logical and physical data modelling. Here,
we assume we are creating a business application system that sits atop a
relational database.
• Generally the product from this phase includes the Structured Query
Language Data Definition Language (SQL DDL) statements that
enable creation of the SQL tables, SQL indexes, SQL data integrity
constraints, and all that is needed to support this application on a
relational database server.
• Persons involved in this phase include product managers and data
modelers.
– Coding and debugging: In this phase the application program code is
created along with unit test, early quality assurance, and program
documentation.
• Persons involved in this phase include the application programmers.
• The application programmers write program code in whatever
language they choose, such as Java, Visual Basic®, C/C++, and
COBOL. Each of these languages requires a specific programming
language skill set, in addition to expertise in delivering useful user
interfaces and systems.
– System test: In this phase, the unit deliverables are assembled to create
a working system. Generally, the product of this phase is correction and
proof.
• Persons involved in this phase include the application programmers.
– Cutover: Here the application is put into production. To support the new
system, data is migrated from the existing system to the newly created
one.
• After a final trial period, the application is considered completed, and
enters the maintenance phase (not displayed).

11.1.1 Issues with the life cycle
The question discussed in this subsection is, “Are there any specific strengths or
capabilities you can make use of when developing application that sits atop a
relational database server?” For example, “are there any specific optimization
best practices provided?” Or, “Is there some other opportunity that is being
overlooked?” In addition to the phases in the execution of a Waterfall Method
SDLC, Figure 11-1 on page 601 also depicts some commonly observed sources
of contention. For example:
Unit development and unit test are often performed on test or staged
data. If unit development and unit test for example with the month end


summary report, are performed with 1000 records in the SQL table, when the
production system will have millions of records, this enables performance
issues and application errors to remain hidden until late in the development
and delivery cycle.

Note: Fixing any issue (such as program logic errors and performance) later in
the software development life cycle costs measurably more time and energy to
repair, when compared to uncovering and correcting these issues earlier in the
software development life cycle.

Application developers generally write their own SQL DML.
– Structured Query Language Data Manipulation Language (SQL DML)
statements are those statements that are located throughout the body of
the application. They lay out (define, one time only) the database. But they
run every day, thousands of times, throughout the life of the application.
– With all of the difficulty of staying current on a given application language,
and having skill to design and deliver capable user interfaces, and working
with business goals and objectives, frankly we are typically satisfied if the
application programmer is expert in their given application programming
language. It is less likely that we will expect the application programmer to
be an expert at reading 400 table SQL schemas than for the application
programmer to know how to program highly performant (query optimizer
expert) SQL DML statements or to develop multi-user SQL record
concurrency models.
– Creation of these SQL DML statements often uncovers errors in the data
model. With thousands of columns and dependencies, there typically are
errors in the data model, even if only a few. Discovering these errors
sooner avoids such activities as unit redevelopment, unit retesting, and
system retesting.
Final cut over to the new system is a risky procedure.
– At this point of the software development life cycle, we are moving as
much data into the new system as it took the existing system to acquire
over a number of years. And, we need to perform this cutover quickly and
without error.
– Cutover of the existing data will also typically uncover forgotten usages of
existing fields and produce additional application requirements.

11.1.2 Process modeling within a life cycle
Figure 11-2 on page 604 depicts a modified Waterfall Method SDLC. The
primary objective in this modified SDLC is to leverage the natural strengths of a


relational database server, and better manage the commonly observed
application development life cycles issues.

Data cut over

Process model
No schema
interpretation

End user- Coding and Coding and
Data model 3. debugging 4. debugging
System test System test
discovery

Author SQL API
Testing with
real data
Performance test

Figure 11-2 Waterfall Method SDLC, optimized for SQL development

Related to Figure 11-2, consider the following:
Immediate existing data cutover following data model.
– Immediately after the creation of the physical data model, a cutover of the
existing data is loaded into the relational database server.
– The existing data model was created at the start of this software
development effort. The data migration scripts can be developed by
engineers who are not on the critical path, since this work requires no
expertise (skills) which will benefit the new system.
– Loading existing data into the new system allows for discovery of new
application requirements. For example, receiving character data to
numeric conversion and existing fields which were supposed to be of one
data type, are now discovered to have new meaning and use.
– Successful loading of the existing data serves as one validation of the
newly created data model.

Note: Besides careful study and engineering, how does your current
SDLC truly verify the accuracy and performant nature of the data
model?

– This procedure also allow for early measurement of the difficulty of
performing the migration, an event most persons only become intimately
familiar with when they actually perform the (final) cut over.


– All development is done on production-sized and actual data. This
condition increases the accuracy and likelihood of performant unit and
system deliverables.
Create a process model.
– The data modeling team gathers great expertise with the business rules
and objectives as they design and deliver their data model. If the modelers
truly understand the business, a requirement for accurate data modeling,
then they understand every business process and every state that records
occupy as data moves through the application (and data model).
– What we are doing here is leveraging the already existing expertise the
data modelers have, and producing a labor savings.
– The final product of the process model is a set of application programming
interfaces (API). In effect, every SQL data manipulation language (DML)
statement that is expected to be run against (supported by) this data
model is produced here.
– If an 18 person development team has two data modelers, 10 application
developers, and assorted other persons, we move one of the application
developers over to the data modeling side. This resource will author every
SQL DML that the application system is expected to run.
– Figure 11-3 on page 606 displays the Model-View-Controller (MVC)
design pattern. MVC came into public prominence in 1995.
– Today, MVC is the industry de facto application development design
pattern. In short, MVC calls for a separation of the application
programming for the (data) model, view (the user interface), and controller
(the business logic).
– A programmer with special skills creates the process model (the SQL
API). These skills include query optimizer and SQL statement tunings,
SQL concurrency models and programming, enterprise-wide SQL error
recovery, and others.


Create
Customer

Modify
Customer

Ship
Order

Report
Weekly Sales

Create
New Site

......

Model Controller View

Figure 11-3 Model-View-Controller (MVC) with SQL API

Execute an early performance trial.
– SQL has a unique capability. Running the SQL SELECT statement for the
previously mentioned month end summary report effectively models the
performance load of that entire program. Most of that program load is the
processing of the relational database SQL statements. Whether the output
uses a given font or font size, paginates in some specific manner, or
similar, does little to the performance load. It is the SQL data manipulation
language statements that form 90% or more of the system load.
– Execute the SQL API routines with the anticipated frequency, concurrency
and spread. Model for the expected number of users on average, at peak
load, and with consideration for future system growth.
– This operation allows us to determine months before final delivery if this
system will perform adequately when finished.
– If the system does not perform adequately, take appropriate action now.
This may entail adjusting the data model, changing the design of data
loads and summations, individual SQL statement tuning and more.
The point is, we can make these changes now with little or no cost,
because few dependencies exist. No application program code has been
authored, no user documentation has been written, and we are still weeks
or months from delivery.


Note: It is at this point, months before delivery, that we know the system will
perform in a given manner, either meeting expectations (needs) or not. We
can then proceed to alter the data model or system design, at little or no
additional cost. Successful execution of the process model is final validation of
the accuracy and performant nature of the data model.

Following the process model, with its performance trial, all unit and system
development and testing is done on real data and at the expected volume. The
application programmers are given defined points to interact with the relational
database server. The application programmers are not expected to know SQL,
and certainly not how to create and model highly performant SQL. And finally,
the data model can be code frozen, which is always important as changes to the
data model during unit and system test is one of the most costly changes that
can occur.

11.2 Artifacts created from a process model
To this point, all attention has been given to application development. The
artifacts created as part of a process model benefit not only development, they
also greatly aid the maintenance phase of an SDLC.

The SQL API is that set of routines that run inside a given application, and
execute (generally, SQL DML) statements against the relational database
server. The SQL API contains a number of function calls to perform tasks such
as run month end summary report, and load new location dimension data. In a
400 table data model, you might have 2000 or more distinct operations that need
to operate against that data model.

Note: The SQL API is being created. The question is, do we organize this
work as a distinct delivery within the entire application development effort?
And, do we use a specific (human) resource with specialized skills to create
this strategic deliverable?

If creation of the SQL API is not done as part of a distinct effort, then the SQL API
is created by every application programmer on the project, and the functions of
the SQL API are scattered throughout the entire (tens of thousands of lines of)
application program code.


11.2.1 Create the SQL API
After the creation of the physical data model, we recommend you load the
existing system data into the database server. The SQL API is then created by
the data modelers, who have knowledge of the requirements of the application,
and one additional (human) resource. This new resource, and the data modelers
have skills that include:
Knowledge of the data model, and the design of those SQL DML statements
that need to run against this data model. These SQL DML statements, along
with the anticipated frequency, concurrency, and spread that these
statements will run in production, form the logical process model. If the data
modelers do not already know every entity attribute and every program state
that each entity goes through in support of the application, then they could not
have modeled that application accurately. Creating the logical process model
is a record of information the data modelers already have.
Ability to create the Model portion of Model-View-Controller, and the ability to
create functions in the application programming language of choice, one for
each single or set of SQL DML routines above. This is a skill set that is
outside of the normal data modeler world, and why a new resource is
included in this activity. These program code artifacts form the physical
process model, or referred to more simply as the SQL API.
Knowledge of that information provided in Chapter 10. This group of
engineers know how to tune SQL statements and work with the query
optimizer.

Note: There is not an industry standard name for the (human) resource
who can write SQL DML and SQL API in an application language, work
with the query optimizer in-depth, and demonstrate all of these skills.
However, it can be a team of persons. If we gave a name to this single
resource, it would be a name such as an Enterprise Data Architect.

Knowledge about executing a performance trial against the relational
database server of choice. Here we run the functions within the SQL API.
There is relational database server tuning, SQL statement tuning, and other
work performed here.
This team authors the SQL API, and executes a performance trial with the
elements of the SQL API with the anticipated frequency, concurrency, and
spread that those statements are expected to run once the application is put
into production. Inadequacies in performance may affect the data model
(including column definitions, table relationships, index plans, and others), the
application program code syntax, and every other tuning topic covered thus
far.


11.2.2 Record query plan documents
For each function in the SQL API the query optimizer query plan is recorded, as
well as the various execution times for a given function such as peak
performance time, best, second best, worst, and average.

This performance trial is automated via whatever technology is available, such
as a load trial tool or shell(C) scripts. As an automated process, this test can be
repeated and in a controlled environment. One of these automated test profiles
should be non-intrusive from a load perspective, allowing it to run on a
production system. We can then run these tests weeks or months after the
system has gone into production as part of our available maintenance and test
routines. If an user complains that a given routine now takes two hours to run
when it used to take 15 minutes, we have greater options to consider. We have
baselines on which to compare this claim, and tools to authenticate and correct
the condition. Figure 11-4 depicts a coarse data model of the information
managed in this activity.

Set query plan

SQL API statement Version
Standard indexes Run time settings

Multi-table indexes

Trial detail
Tables

Query plan
Trial

Figure 11-4 Coarse model of data managed in this activity

From Figure 11-4, consider the following:
– The data model proper produces SQL tables. Generally, an SQL index
specifies a single or set of columns from one SQL table. Some relational
database vendors support multi-table SQL indexes.
– Since the normal update and maintenance of indexes have costs, each
index is validated (known to exist) for a given purpose. The purposes
include support of primary key or unique data integrity constraint, foreign
key constraints, and then support for a given SQL API statement (routine,
function). A given SQL API function may require use of zero or more SQL
indexes.


– Each SQL API function is associated with a single or set of query
optimizer query plans. These various plans for each SQL API function
may be recorded as best case or acceptable case, or with given run-time
settings.
– Performance trials are also recorded, which have a grouping of SQL API
functions which are run (with a frequency, concurrency, and spread). At
this level too, we observe query statistics (times, resource consumed),
and query plans, as these may vary under load and other variables.

11.3 Example of process modeling
Example 11-1 displays a sample data model. This sample SQL data definition
language code is referenced throughout the remainder of this section.

Example 11-1 Data model example

create table cust_calls
(
customer_num integer,
call_dtime datetime year to minute,
user_id char(32) default user,
call_code char(1),
call_descr char(240),
res_dtime datetime year to minute,
res_descr char(240),
--
primary key (customer_num, call_dtime),
foreign key (customer_num)
references customer (customer_num),
foreign key (call_code)
references call_type (call_code)
);

create table catalog
(
catalog_num serial(10001),
stock_num smallint not null,
manu_code char(3) not null,
cat_descr text,
cat_picture byte,
cat_advert varchar(255, 65),
--
primary key (catalog_num),
foreign key (stock_num, manu_code)
references stock constraint aa


);

create table customer
(
customer_num serial(101),
fname char(15),
lname char(15),
company char(20),
address1 char(20),
address2 char(20),
city char(15),
state char(2),
zipcode char(5),
phone char(18),
--
primary key (customer_num)
);
create index zip_ix on customer (zipcode);

create table items
(
item_num smallint,
order_num integer,
quantity smallint check (quantity >= 1),
total_price money(8),
--
primary key (item_num, order_num),
foreign key (order_num)
references orders (order_num),
references stock (stock_num, manu_code)
);

create table manufact
(
manu_code char(3),
manu_name char(15),
lead_time interval day(3) to day,
--
primary key (manu_code)
);

create table orders
(
order_num serial(1001),
order_date date,
customer_num integer not null,


ship_instruct char(40),
backlog char(1),
po_num char(10),
ship_date date,
ship_weight decimal(8,2),
ship_charge money(6),
paid_date date,
--
primary key (order_num),
foreign key (customer_num) references customer (customer_num)
);

create procedure read_address (lastname char(15))
returning char(15) as pfname, char(15) as plname,
char(20) as paddress1, char(15) as pcity,
char(2) as pstate, char(5) as pzipcode;

define p_fname, p_city char(15);
define p_add char(20);
define p_state char(2);
define p_zip char(5);

select fname, address1, city, state, zipcode
into p_fname, p_add, p_city, p_state, p_zip
from customer
where lname = lastname;

return p_fname, lastname, p_add, p_city, p_state, p_zip;
end procedure;

create table state
(
code char(2),
sname char(15),
--
primary key (code)
);

create table stock
(
stock_num smallint,
manu_code char(3),
description char(15),
unit_price money(6),
unit char(4),
unit_descr char(15),
--
primary key (stock_num, manu_code),
foreign key (manu_code) references manufact


);

create table sports
(
catalog_no serial unique,
stock_no smallint,
mfg_code char(5),
mfg_name char(20),
phone char(18),
descript varchar(255)
);

create table call_type
(
call_code char(1),
code_descr char(30),
--
primary key (call_code)
);

create table catalog
(
catalog_num serial(10001),
cat_descr text,
cat_picture byte,
cat_advert varchar(255, 65),
--
primary key (catalog_num),
references stock constraint aa
);

CREATE TABLE msgs
(
lang char(32),
number integer,
message nchar(255)
);

SET NO COLLATION;

CREATE INDEX idxmsgs_enus on msgs (message);

SET COLLATION ‘fr_fr.8859-1’; -- Create an index using fr_fr.8859-1.

CREATE INDEX idxmsgs_frfr on msgs (message);


SET NO COLLATION;

create table log_record
(
item_num smallint,
ord_num integer,
username character(32),
update_time datetime year to minute,
old_qty smallint,
new_qty smallint
);

create trigger upqty_i
update of quantity on items
referencing old as pre_upd
new as post_upd
for each row
(
insert into log_record
values (pre_upd.item_num, pre_upd.order_num, user, current,
pre_upd.quantity, post_upd.quantity)
);

create view custview (firstname, lastname, company, city)
as
select fname, lname, company, city
from customer
where city = ‘Redwood City’
with check option;

create view someorders (custnum,ocustnum,newprice)
as
select orders.order_num,items.order_num,items.total_price*1.5
from orders, items
where orders.order_num = items.order_num
and items.total_price > 100.00;

11.3.1 Explanation of the process example
With regards to Example 11-1 on page 610, refer to the following:
This data model contains:
– Thirteen SQL tables.
– Two SQL views. The SQL view Custview has a new integrity constraint
built within it (a filter constraint) but references only one table. The SQL
view Someorders joins two tables and has a single filter within it.


– One SQL stored procedure, which does a single table read with a filter.
– One SQL trigger. One SQL update of Items. Quantity, a new record, is
SQL inserted into another SQL table.
– Ten SQL primary keys, and eight SQL foreign keys.
– There are three additional (non-primary key, non-foreign key) indexes
which are made, and they are Customer.Zipcode, and two on
Msgs.Message.
The two indexes on Msgs.Message are present to demonstrate indexes on
varying language sets. As indexes on a 255 multi-byte column in a 291 total
byte record length table, these indexes will surely be removed (again, these
are present for demonstration purposes only).
Additionally, we will not bother to cost the need for the primary key and
foreign key indexes since their use (presence) is more feature functionality
than pure support for query processing.

Example 11-1 on page 610 provides us the product of our logical and/or physical
data model. At some point before, during, or after the creation of the data model,
we considered the concept of security and roles. For example:
During user discovery (phase one of the Waterfall Method SDLC), we may
have discovered that there were two categorizations of users accessing this
system; operators and senior operators. Operators are given certain SQL
permissions (such as SELECT and UPDATE) on given tables and columns,
and so are senior operators (presumably with more permissions).
From the perspective of the defined user roles, we also document the
expected use of the system:
– An operator is found to execute, for example, SQL-SELECT-001, then
002, then 003, in sequence, at least 22 times per hour, with a think time
(pause) of five seconds per iteration.
– The senior operator is also defined to follow given operations and
sequences.
– Acceptable service levels (timings) are determined for all of the above.
The creation of this package of application definition, knowledge, and
requirements forms the logical process model.

11.4 An SQL DML example
Also from the demonstration database (the source of the SQL DDL in
Example 11-1 on page 610), we capture a number of SQL DMLs. These are
shown in Example 11-2 on page 616.


Example 11-2 A contributor to the logical process model
0000
0002
0003 delete from stock
0004 where stock_num = 102;
0005
0006 insert into sports
0007 values (0,18,’PARKR’, ‘Parker Products’, ‘503-555-1212’,
0008 ‘Heavy-weight cotton canvas gi, designed for aikido or
0009 judo but suitable for karate. Quilted top with side ties,
0010 drawstring waist on pants. White with white belt. Pre-washed
0011 for minimum shrinkage. Sizes 3-6.’);
0012
0013 select
0014 max (ship_charge),
0015 min (ship_charge)
0016 from orders;
0017
0018 select
0019 o.order_num,
0020 sum (i.total_price) price,
0021 paid_date - order_date span
0022 from orders o, items i
0023 where o.order_date > ‘01/01/89’
0024 and o.customer_num > 110
0025 and o.order_num = i.order_num
0026 group by 1, 3
0027 having count (*) < 5
0028 order by 3
0029 into temp temptab1;
0030
0031 SELECT message
0032 FROM msgs
0033 WHERE lang = ‘en_us’
0034 ORDER BY message;
0035
0036 select
0037 order_num,
0038 count(*) number,
0039 avg (total_price) average
0040 from items
0041 group by order_num
0042 having count(*) > 2;
0043
0044 select
0045 c.customer_num, c.lname, c.company,
0046 c.phone, u.call_dtime, u.call_descr


0047 from customer c, cust_calls u
0048 where c.customer_num = u.customer_num
0049 order by 1;
0050
0051 select
0052 c.customer_num, c.lname, c.company,
0053 c.phone, u.call_dtime, u.call_descr
0054 from customer c, outer cust_calls u
0055 where c.customer_num = u.customer_num
0056 order by 1;
0057
0058 select
0059 c.customer_num, c.lname, o.order_num,
0060 i.stock_num, i.manu_code, i.quantity
0061 from customer c, outer (orders o, items i)
0062 where c.customer_num = o.customer_num
0064 and manu_code IN (‘KAR’, ‘SHM’)
0065 order by lname;
0066
0067 select
0068 c.customer_num, lname, o.order_num,
0069 stock_num, manu_code, quantity
0070 from customer c, outer (orders o, outer items i)
0073 and manu_code IN (‘KAR’, ‘SHM’)
0074 order by lname;
0075
0076 select
0077 c.customer_num, lname, o.order_num,
0078 order_date, call_dtime
0079 from customer c, outer orders o, outer cust_calls x
0081 and c.customer_num = x.customer_num
0082 order by 1
0083 into temp service;
0084
0085 select *
0086 from stock
0087 where description like ‘%bicycle%’
0088 and manu_code not like ‘PRC’
0089 order by description, manu_code;
0090
0091 select
0092 order_num, total_price
0093 from items a
0094 where 10 >
0095 (


0096 select count (*)
0097 from items b
0098 where b.total_price < a.total_price
0099 )
0100 order by total_price;
0101
0102 select distinct stock_num, manu_code
0103 from stock
0104 where unit_price < 25.00
0105 union
0106 select stock_num, manu_code
0107 from items
0108 where quantity > 3;
0109
0110 update sports
0111 set phone = ‘808-555-1212’
0112 where mfg_code = ‘PARKR’;

11.4.1 Explanation of DML example
With regards to Example 11-2 on page 616, refer to the following:
Entries such as the SQL DELETE on lines 0003-0004 are less than ideal in a
production applications, but serve our needs here.
Each of the distinct SQL statements in Example 11-2 on page 616 will be
found within function calls of the given application development language,
such as Java or Visual Basic. Following the design pattern of
Model-View-Controller, each of these (Model component) functions will offer
little in the way of program logic. Each will perform their given SQL statement,
and contain the minimum amount of error processing and recovery code.
Each of the individual SQL DML statements found inside Example 11-2 on
page 616 will be further discussed as follows:
– Each single statement will be reviewed as though it were a member of a
distinct SQL API function call.
– Example 11-2 on page 616 contains one SQL DELETE, one SQL
INSERT, one SQL UPDATE, and twelve SQL SELECTs.

Lines 003-0004, the SQL DELETE
The SQL DELETE on lines 003-0004 exist in the real world as something similar
to:
DELETE FROM stock WHERE stock_num = ?;

And a host variable will be supplied on execution. See the following:


From the SQL schema in Example 11-1 on page 610, the Stock table acts as
our inventory table, where Manufact acts like the list of valid stock providers
(Manufacturers). We can determine this from the relationships in these two
SQL tables, as well as from the sample SQL SELECTs in Example 11-2 on
page 616.
The primary key on the Stock table is based on the composite key
(Stock_num, Manu_code).
Since the SQL DELETE in this example references a subset of the composite
key (the leading columns) we will be concerned about to the cardinality of this
key value if we were performing a two table join under this relationship (see
example in “Query optimizer by example” on page 503).
As it stands, the filter criteria in this SQL WHERE clause is supported via an
index, the primary key to this table.
The resultant query plan will be reviewed to confirm use of the index, and the
observed service level (timing) of this statement will be checked against
necessary targets.

Lines 0006-0011, the SQL INSERT
Again, this code is taken from a sample database. The real-world SQL INSERT
will resemble SQL closer to:
INSERT INTO sports VALUES (?, ?, ?, ?, ?, ?);

Host variables will supply the column values on execution. We check for the
presence of SQL TRIGGERS on the event of this SQL INSERT. If SQL
TRIGGERS cause subsequent SQL SELECTS, UPDATES, or DELETES, then
these statements are costed. In this our actual case, no SQL TRIGGERs are in
effect for this SQL INSERT. The observed service level (timings) of this
statement are checked against necessary targets.

Lines 0013-0016, SQL SELECT # 1
This SQL SELECT returns two aggregate columns from a single SQL table with
no joins or filters. Since Orders.Ship_charge is not indexed, this SQL SELECT
will cause a sequential scan of this entire table. If a higher observed service level
(timings) was needed for this specific statement, tuning will have to be performed
to this database server or data model. (Does the relational database server
provide a technology to know the maximum and minimum of this column without
requiring an index, other?)

If the required service level is met, then we are fine. If the service level is not met,
then we need to measure the anticipated frequency of the execution of this
statement, and the importance that it execute as requested. Solutions to improve
the performance of this statement will include:


Indexing this column.
Creating a materialized view or materialized query table to support this query.
Creating SQL TRIGGERS on the SQL INSERT, UPDATE, or DELETE of this
table and keeping these statistics as calculated values in a second,
supporting SQL table.
Others, as allowed for by the given relational database vendor.

Lines 0018-0029, SQL SELECT #2
This SQL SELECT reads from two SQL tables where the tables are joined and
the Orders table has two range operand filter criteria (lines 0023 and 0024).
There is a SQL GROUP BY to support one of the aggregate expressions, (line
0020). Also, there is a filter on an additional aggregate expression, (line 0027).

Since Items.Order_num is indexed, as shown in Example 11-1 on page 610 (it is
an anchored primary key contributor within Items), the table order is likely to be
Orders then Items, with an indexed table access method on Items. The
aggregate calculations will be expensive and timings could be reduced by
maintaining these columns within Orders using SQL TRIGGERs. The redundant
columns we will keep in Order include Price (the calculated order total price) and
Item Count (the number of line items in a given order). Span will not need to be
precalculated since its resultant value is the product of two columns on the same
row (entity instance) from Orders.

This SQL SELECT reads from one SQL table, and has one filter criteria. From
the SQL DDL in Example 11-1 on page 610, we see that Msgs.Lang is not
indexed, but that the SQL ORDER BY Column Msgs.Message is. Earlier we
commented that the index on Msgs.Message is not likely to be performant, since
it occupies 255 of 291 bytes (column length versus total record length in the
table).

This SQL SELECT will be processed via a sequential scan on this table to satisfy
the filter criteria, and results will be sent to the relational database server sort
package to satisfy the SQLORDER BY clause. If observed service levels are not
being met, an index on Msgs.Lang will help. If Msgs.Message were a smaller
field, then a composite index on Msgs (Lang, Message) would be ideal, since this
single SQL index will service both the filter criteria and the SQL ORDER BY
clause. See also pseudo by clause in 10.6.1, “Query rewrite” on page 568.

This is a very interesting SQL SELECT. In simple language, we explain these
lines:


In the data model, the Orders table is the order header. That is, the columns
within a customer order that do not repeat (Order Date, Customer Number,
and other). The Items table serves as the Order Detail table for the Order Line
Items.
The SQL SELECT essentially reports on Order header conditions; the count
of line items in the order, and the average price.

Note: This is a very interesting SQL SELECT because it exemplifies the
structure and cost of SQL SELECTs received when you do not process the
model.

As the SQL SELECT currently exists, there is a SQL GROUP BY on
Items.Order_num, which is an anchored column in a foreign key. We could
receive index support for the processing of this query, but that fact needs to
be confirmed by reading the query plan. While there are no filters on this SQL
SELECT, the presence of the SQL GROUP BY calls for records to be read in
sorted fashion (therefore, desire for an index).
If the current SQL SELECT does not meet requested service levels (timings),
the ideal index in this table will be on Items (Order_num, Total_price), allowing
for a key only indexed retrieval (also assuming we do not keep these
aggregate columns in the Order header table).
As we review this query, we can discover one other area for optimization:
– The SQL table, Items, has three indexes, that include a primary key on
(Item_num, Order_num), foreign key on (Order_num), and foreign key on
(Stock_num, Manu_code).
– As a column, the item number (that is the sequencing of line items within a
customer order), is probably insignificant. “Show me all third line items on
a multi-item customer order” is not a typical query, and probably not a
query we need to support.
– If the primary key were made to be on (Order_num, Item_num), then the
foreign key for (Order_num) could also be served from this one index. The
net result is one less index on this table; always a good thing.
As implied above, it might be better to model the item count (per order) at the
order header level, and average (Items.Total_price) at the Items level, as a
calculated column. All of this depends of the needs of the user application; the
observed frequency, concurrency and spread of (SQL) statements run, and
the requested service levels (timings).


This SQL SELECT returns all records from two SQL tables where the tables are
joined. From Example 11-1 on page 610, we can see that both
Customer.Customer_num and Cust_calls.Customer_num are indexed.

Similar to SQL SELECT #4 above, and from Example 11-1 on page 610, we see
that Cust_calls.Customer_num could lose one index. That column is anchored in
both a primary key and foreign key in that table, which is largely redundant.

The only difference between SQL SELECT #5 and #6 is the presence of the SQL
OUTER keyword on line 0054. Because SQL Table Cust_calls in subservient to
Customer, the Customer table will be processed first in SQL SELECT #6. If this
query criteria causes less than requested service levels (timings), we might seek
to model around this condition.

This SQL SELECT reads from three SQL tables where all tables are joined, and
the SQL table Items has a filter, (line 0064). Customer is the dominant SQL table,
(line 0061), while SQL Tables Orders and Items are placed in a SQL OUTER
clause.

From Example 11-1 on page 610, we can determine that all of the join columns
are indexed. The filter on Items.Manu_code is not anchored via an index, and is
not a member of the index that will be needed to perform the join into SQL table
Items.

If requested service levels for this SQL SELECT were not being met, you might
consider possibly removing the SQL OUTER clause, replacing it with a
non-OUTER SQL SELECT, and then UNION ALL customers with no orders (the
net result of the SQL OUTER clause). Also, reviewing the indexes in place on
Items to allow for key only processing, or single index processing of both the join
and the filter criteria.

The only difference between SQL SELECT #7 and #8 is the presence of another
SQL OUTER clause, as shown on line 0070. The same tuning guidelines will be
in effect as for SQL SELECT #7. That is, removal (engineering around) the SQL
OUTER clause, and including the Items.Manu_code filter in any participating
indexes.


While SQL SELECT #9 references different table and join columns than SQL
SELECT #7 and #8, the same tuning guidelines apply here as well. All of the join
columns are supported via indexes.

This SQL SELECT reads from one SQL table with two filter criteria. Both filter
criteria offer poor selectivity and are not likely to use indexes (negation of index,
non-initial substring, line 0087, and negation of index, poor selectivity filter, line
0088, the inequality).

Note: This is the only SQL SELECT in the grouping that makes use of a
“SELECT *”; that is, the wildcard to select all columns in the list. We want to
remove this syntax, since adding or dropping columns from the SQL table may
break our application program code. In effect, we will later attempt to SELECT
10+1 columns into a list of 10 variables.

If this were a critical query, we might need to model a Keywords type table; as in,
the description of a given Stock unit is recognized by the keywords as recorded
in a primary key table of symbols.

This SQL SELECT provides a correlated subquery, as shown on line 0098 and
the reference to an outer table column, A.Total_price, inside the inner SELECT.

Note: This is another interesting example. An experienced engineer would
detect that this query could be rewritten, if the query optimizer query rewrite
capability does not. This specific example supports the argument for a
process model versus having the application programmers author their own
SQL DML statements.

The query optimizer query rewrite capability could detect that the inner SQL
SELECT can be executed once, to retrieve the effect of,
SELECT total_price, COUNT(*) FROM items GROUP BY 1;

And then join the outer SQL SELECT to that result. The correlated subquery can
be rewritten as a join, since the cardinality of the outer to inner SQL SELECTs is
many to one.


This is a unioned SQL SELECT. The presence of the UNION keyword, without a
UNION ALL, will call to remove duplicates from the resultant data set. This
condition makes the DISTINCT keyword in the first SQL SELECT redundant (not
necessary).

Table access methods for both SQL SELECTs are sequential, since neither filter
criteria is supported by an index.

In simple terms, this SQL SELECT seems to report the effect of; show me every
item I have sold for fewer than $25 and of which I have more than three. If this is
an important query, we could data model it as such.

Lines 0110-0112, the SQL UPDATE
The Sports table has no index, and the table access method for this query will be
sequential. If a greater service level (timings) is requested, an index on
Sports.Manu_code will work.

Results: SQL objects not used
By the review of the SQL API above, we see that the two SQL VIEWS are
currently unused, and the only additional index, that on Customer.Zipcode, is
also unused. The index, at least, could be removed since maintaining it adds
cost to the application.

Based on the necessary timings of the above SQL API routines, there may have
been a call to change the data model. This is primarily to maintain redundant or
precalculated (expression) columns.

11.5 Conclusions
In this chapter we have discussed how you can apply the query optimization
techniques presented in Chapter 10. The discussions are heavily oriented to
application development, but that is the means by which we access and analyze
the data stored in the data warehouse to feed the business intelligence
environment.

Whether by stored application or by ad hoc query, the basic objective of business
intelligence is to get the data out of the data warehouse, analyze it, and enable
more informed business decisions to improve management and control of the
business enterprise.

The base of your data structure is the data model. And to support the
requirements of BI, we have focussed specifically on dimensional modeling.


However, having a data structure is just part of a solution. The key is to have a
data structure that supports understanding and use of the data, ease of access
for analysis, and high performance of the applications and queries.

Knowing that your data model is highly performant before ever writing any
applications or queries is of significant value. It can enable improved client
satisfaction by delivering highly performant queries from the initial introduction of
the system. It can also save time and money by eliminating the need to rebuild
poorly designed data structures and applications. In short, it can enable you to
do it right the first time.


Glossary

Access Control List (ACL). The list of principals BLOB. Binary Large Object, a block of bytes of data
that have explicit permission (to publish, to (for example, the body of a message) that has no
subscribe to, and to request persistent delivery of a discernible meaning, but is treated as one solid
publication message) against a topic in the topic entity that cannot be interpreted.
tree. The ACLs define the implementation of
topic-based security. Business subject area. A particular function or
area within an enterprise whose processes and
Additive Measure. Measure of a fact that can be activities can be described by a defined set of data
added across all dimensions. elements.

Aggregate. Pre-calculated and pre-stored Candidate key. One of multiple possible keys for
summaries, kept in the data warehouse to improve an entity.
query performance
Cardinality. The number of elements in a
Aggregation. An attribute level transformation that mathematical set or group.
reduces the level of detail of available data. For
example, having a Total Quantity by Category of Commit. An operation that applies all the changes
Items rather than the individual quantity of each item made during the current unit of recovery or unit of
in the category. work. After the operation is complete, a new unit of
recovery or unit of work begins.
Analytic. An application or capability that performs
some analysis on a set of data. Compensation. The ability of DB2 to process SQL
that is not supported by a data source on the data
Application Programming Interface. An from that data source.
interface provided by a software product that
enables programs to request services. Composite Key. A key in a fact table that is the
concatenation of the foreign keys in the dimension
Associative entity. An entity created to resolve a tables.
many-to-many relationship into two one-to-many
relationships. Computer. A device that accepts information (in
the form of digitalized data) and manipulates it for
Asynchronous Messaging. A method of some result based on a program or sequence of
communication between programs in which a instructions on how the data is to be processed.
program places a message on a message queue,
then proceeds with its own processing without Configuration. The collection of brokers, their
waiting for a reply to its message. execution groups, the message flows and sets that
are assigned to them, and the topics and associated
Attribute. A characteristic of an entity, such as a access control specifications.
field in a dimension table.
Connector. See Message processing node
connector.


Cube. Another term for a fact table. It can Data Model. A representation of data, its definition,
represent “n” dimensions, rather than just three (as characteristics, and relationships.
may be implied by the name).
Data Partition. A segment of a database that can
DDL (Data Definition Language). a SQL statement be accessed and operated on independently even
that creates or modifies the structure of a table or though it is part of a larger data structure.
database. For example, CREATE TABLE, DROP
TABLE, ALTER TABLE, CREATE DATABASE. Data Refresh. A data loading technique where all
the data in a database is completely replaced with a
DML (Data Manipulation Language). An INSERT, new set of data.
UPDATE, DELETE, or SELECT SQL statement.
Data silo. A standalone set of data in a particular
Data Append. A data loading technique where department or organization used for analysis, but
new data is added to the database leaving the typically not shared with other departments or
existing data unaltered. organizations in the enterprise.

Data Append. A data loading technique where Data Warehouse. A specialized data environment
new data is added to the database leaving the developed, structured, shared, and used specifically
existing data unaltered. for decision support and informational (analytic)
applications. It is subject oriented rather than
Data Cleansing. A process of data manipulation application oriented, and is integrated, non-volatile,
and transformation to eliminate variations and and time variant.
inconsistencies in data content. This is typically to
improve the quality, consistency, and usability of the Database Instance. A specific independent
data. implementation of a DBMS in a specific
environment. For example, there might be an
Data Federation. The process of enabling data independent DB2 DBMS implementation on a
from multiple heterogeneous data sources to appear Linux® server in Boston supporting the Eastern
as if it is contained in a single relational database. offices, and another separate and independent DB2
Can also be referred to “distributed access”. DBMS on the same Linux server supporting the
western offices. They would represent two instances
Data mart. An implementation of a data of DB2.
warehouse, typically with a smaller and more tightly
restricted scope - such as for a department, Database Partition. Part of a database that
workgroup, or subject area. It could be independent, consists of its own data, indexes, configuration files,
or derived from another data warehouse and transaction logs.
environment (dependent).
DataBlades. These are program modules that
Data mart - Dependent. A data mart that is provide extended capabilities for Informix
consistent with, and extracts its data from, a data databases, and are tightly integrated with the DBMS.
warehouse.
DB Connect. Enables connection to several
Data mart - Independent. A data mart that is relational database systems and the transfer of data
standalone, and does not conform with any other from these database systems into the SAP®
data mart or data warehouse. Business Information Warehouse.

Data Mining. A mode of data analysis that has a Debugger. A facility on the Message Flows view in
focus on the discovery of new information, such as the Control Center that enables message flows to be
unknown facts, data relationships, or data patterns. visually debugged.


Deploy. Make operational the configuration and Foreign Key. An attribute or set of attributes that
topology of the broker domain. refer to the primary key of another entity.

Dimension. Data that further qualifies and/or Gateway. A means to access a heterogeneous
describes a measure, such as amounts or durations. data source. It can use native access or ODBC
technology.
Distributed Application In message queuing, a
set of application programs that can each be Grain. The fundamental atomic level of data
connected to a different queue manager, but that represented in a fact table. As examples, typical
collectively constitute a single application. grains that could be used, when considering time,
would be day, week, month, year, and so forth.
Drill-down. Iterative analysis, exploring facts at
more detailed levels of the dimension hierarchies. Granularity. The level of summarization of the
data elements. It refers to the level of detail
Dynamic SQL. SQL that is interpreted during available in the data elements. The more detail data
execution of the statement. that is available, the lower the level of granularity.
Conversely, the less detail that is available, the
Engine. A program that performs a core or higher the level of granularity (or level of
essential function for other programs. A database summarization of the data elements).
engine performs database functions on behalf of the
database user programs. Instance. A particular realization of a computer
process. Relative to database, the realization of a
Enrichment. The creation of derived data. An complete database environment.
attribute level transformation performed by some
type of algorithm to create one or more new Java Database Connectivity. An application
(derived) attributes. programming interface that has the same
characteristics as ODBC but is specifically designed
Entity. A person, place, thing, or event of interest to for use by Java database applications.
the enterprise. Each entity in a data model is unique.
Java Development Kit. Software package used to
Extenders. These are program modules that write, compile, debug and run Java applets and
provide extended capabilities for DB2, and are applications.
tightly integrated with DB2.
Java Message Service. An application
FACTS. A collection of measures, and the programming interface that provides Java language
information to interpret those measures in a given functions for handling messages.
context.
Java Runtime Environment. A subset of the Java
Federated data. A set of physically separate data Development Kit that allows you to run Java applets
structures that are logically linked together by some and applications.
mechanism, for analysis, but which remain
physically in place. Key. An attribute of set of attributes that uniquely
identifies an entity.
Federated Server. Any DB2 server where the
WebSphere Information Integrator is installed. Materialized Query Table. A table where the
results of a query are stored, for later reuse.
Federation. Providing a unified interface to diverse
data. Measure. A data item that measures the
performance or behavior of business processes.

Glossary 629

Message domain. The value that determines how ODS. (1) Operational data store: A relational table
the message is interpreted (parsed). for holding clean data to load into InfoCubes, and
can support some query activity. (2) Online Dynamic
Message flow. A directed graph that represents Server - an older name for IDS.
the set of activities performed on a message or event
as it passes through a broker. A message flow OLAP. OnLine Analytical Processing.
consists of a set of message processing nodes and Multidimensional data analysis, performed in real
message processing connectors. time. Not dependent on underlying data schema.

Message parser. A program that interprets the bit Open Database Connectivity. A standard
stream of an incoming message and creates an application programming interface for accessing
internal representation of the message in a tree data in both relational and non-relational database
structure. A parser is also responsible to generate a management systems. Using this API, database
bit stream for an outgoing message from the internal applications can access data stored in database
representation. management systems on a variety of computers
even if each database management system uses a
Meta Data. Typically called data (or information) different data storage format and programming
about data. It describes or defines data elements. interface. ODBC is based on the call level interface
(CLI) specification of the X/Open SQL Access
MOLAP. Multidimensional OLAP. Can be called Group.
MD-OLAP. It is OLAP that uses a multidimensional
database as the underlying data structure. Optimization. The capability to enable a process
to execute and perform in such a way as to maximize
Multidimensional analysis. Analysis of data performance, minimize resource utilization, and
along several dimensions. For example, analyzing minimize the process execution response time
revenue by product, store, and date. delivered to the user.

Multi-Tasking. Operating system capability which Partition. Part of a database that consists of its
allows multiple tasks to run concurrently, taking turns own data, indexes, configuration files, and
using the resources of the computer. transaction logs.

Multi-Threading. Operating system capability that Pass-through. The act of passing the SQL for an
enables multiple concurrent users to use the same operation directly to the data source without being
program. This saves the overhead of initiating the changed by the federation server.
program multiple times.
Pivoting. Analysis operation where user takes a
Nickname. An identifier that is used to reference different viewpoint of the results. For example, by
the object located at the data source that you want changing the way the dimensions are arranged.
to access.
Primary Key. Field in a table record that is
Node Group. Group of one or more database uniquely different for each record in the table.
partitions.
Process. An instance of a program running in a
Node. See Message processing node and Plug-in computer.
node.
Program. A specific set of ordered operations for a
Non-Additive Measure. Measure of a fact that computer to perform.
cannot be added across any of its dimensions, such
as a percentage.


Pushdown. The act of optimizing a data operation Synchronous Messaging. A method of
by pushing the SQL down to the lowest point in the communication between programs in which a
federated architecture where that operation can be program places a message on a message queue
executed. More simply, a pushdown operation is one and then waits for a reply before resuming its own
that is executed at a remote server. processing.

Relationship. The business rule that associates Task. The basic unit of programming that an
entities. operating system controls. Also see Multi-Tasking.

ROLAP. Relational OLAP. Multidimensional Thread. The placeholder information associated
analysis using a multidimensional view of relational with a single use of a program that can handle
data. A relational database is used as the underlying multiple concurrent users. Also see Multi-Threading.
data structure.
Type Mapping. The mapping of a specific data
Roll-up. Iterative analysis, exploring facts at a source type to a DB2 UDB data type
higher level of summarization.
Unit of Work. A recoverable sequence of
Semi-additive Measure. Measure of a fact that operations performed by an application between two
can be added across only some of its dimensions, points of consistency.
such as a balance.
User Mapping. An association made between the
Server. A computer program that provides federated server user ID and password and the data
services to other computer programs (and their source (to be accessed) used ID and password.
users) in the same or other computers. However, the
computer that a server program runs in is also Virtual Database. A federation of multiple
frequently referred to as a server. heterogeneous relational databases.

Shared nothing. A data management architecture Warehouse Catalog. A subsystem that stores and
where nothing is shared between processes. Each manages all the system meta data.
process has its own processor, memory, and disk
space. Wrapper. The means by which a data federation
engine interacts with heterogeneous sources of
Spreadmart. A standalone, non-conforming, data. Wrappers take the SQL that the federation
non-integrated set of data, such as a spreadsheet, engine uses and maps it to the API of the data
used for analysis by a particular person, department, source to be accessed. For example, they take DB2
or organization. SQL and transform it to the language understood by
the data source to be accessed.
Static SQL. SQL that has been compiled prior to
execution. Typically provides best performance. xtree. A query-tree tool that allows you to monitor
the query plan execution of individual queries in a
Static SQL. SQL that has been compiled prior to graphical environment.
execution. Typically provides best performance.

Subject Area. A logical grouping of data by
categories, such as customers or items.

Glossary 631

Abbreviations and acronyms
ACS access control system DCE Distributed Computing
ADK Archive Development Kit Environment

AIX® Advanced Interactive DCM Dynamic Coserver
eXecutive from IBM Management

API Application Programming DCOM Distributed Component
Interface Object Model

AQR automatic query re-write DDL Data Definition Language - a
SQL statement that creates or
AR access register modifies the structure of a
ARM automatic restart manager table or database. For
example, CREATE TABLE,
ART access register translation
DROP TABLE.
ASCII American Standard Code for
DES Data Encryption Standard
Information Interchange
DIMID Dimension Identifier
AST Application Summary Table
DLL Dynamically Linked Library
BLOB Binary Large OBject
DMDL Dimensional Modeling Design
BW Business Information
Life Cycle
Warehouse (SAP)
DML Data Manipulation Language -
CCMS Computing Center
an INSERT, UPDATE,
Management System
DELETE, or SELECT SQL
CFG Configuration statement.
CLI Call Level Interface DMS Database Managed Space
CLOB Character Large OBject DPF Data Partitioning Facility
CLP Command Line Processor DRDA® Distributed Relational
CORBA Common Object Request Database Architecture™
Broker Architecture DSA Dynamic Scalable
CPU Central Processing Unit Architecture
CS Cursor Stability DSN Data Source Name
DAS DB2 Administration Server DSS Decision Support System
DB Database EAI Enterprise Application
Integration
DB2 Database 2™
EAR Entity, Attribute, Relationship
DB2 UDB DB2 Universal DataBase
data model. Also denoted by
DBA Database Administrator E/R.
DBM DataBase Manager EBCDIC Extended Binary Coded
DBMS DataBase Management Decimal Interchange Code
System EDA Enterprise Data Architecture


EDU Engine Dispatchable Unit ISM Informix Storage Manager
EDW Enterprise Data Warehouse ISV Independent Software Vendor
EGM Enterprise Gateway Manager IT Information Technology
EJB™ Enterprise Java Beans ITR Internal Throughput Rate
E/R Enterprise Replication ITSO International Technical
E/R Entity, Attribute, Relationship Support Organization
data model. Also denoted by IX Index
EAR. J2EE Java 2 Platform Enterprise
ERP Enterprise Resource Planning Edition
ESE Enterprise Server Edition JAR Java Archive
ETL Extract, Transform, and Load JDBC™ Java DataBase Connectivity
ETTL Extract, Transform/Transport, JDK™ Java Development Kit
and Load JE Java Edition
FJS Filter, then Join, then Sort JMS Java Message Service
FP Fix Pack JRE™ Java Runtime Environment
FTP File Transfer Protocol JVM™ Java Virtual Machine
Gb Giga bits KB Kilobyte (1024 bytes)
GB Giga Bytes LDAP Lightweight Directory Access
GUI Graphical User Interface Protocol
HADR High Availability Disaster LPAR Logical Partition
Recovery LV Logical Volume
HDR High availability Data Mb Mega bits
Replication
MB Mega Bytes
HPL High Performance Loader
MDC Multidimensional Clustering
I/O Input/Output
MPP Massively Parallel Processing
IBM International Business
Machines Corporation MQI Message Queuing Interface

ID Identifier MQT Materialized Query Table

IDE Integrated Development MRM Message Repository
Environment Manager

IDS Informix Dynamic Server MTK DB2 Migration ToolKit for
Informix
II Information Integrator
NPI Non-Partitioning Index
IMG Integrated Implementation
Guide (for SAP) ODBC Open DataBase Connectivity

IMS™ Information Management ODS Operational Data Store
System OLAP OnLine Analytical Processing
ISAM Indexed Sequential Access OLE Object Linking and
Method Embedding


OLTP OnLine Transaction VG Volume Group (Raid disk
Processing terminology).
ORDBMS Object Relational DataBase VLDB Very Large DataBase
Management System VP Virtual Processor
OS Operating System VSAM Virtual Sequential Access
O/S Operating System Method
PDS Partitioned Data Set VTI Virtual Table Interface
PIB Parallel Index Build WSDL Web Services Definition
PSA Persistent Staging Area Language

RBA Relative Byte Address WWW World Wide Web

RBW Red Brick™ Warehouse XBSA X-Open Backup and Restore
APIs
RDBMS Relational DataBase
Management System XML eXtensible Markup Language

RID Record Identifier XPS Informix eXtended Parallel
Server
RR Repeatable Read
RS Read Stability
SCB Session Control Block
SDK Software Developers Kit
SDLC Software Development Life
Cycle
SID Surrogage Identifier
SMIT Systems Management
Interface Tool
SMP Symmetric MultiProcessing
SMS System Managed Space
SOA Service Oriented Architecture
SOAP Simple Object Access
Protocol
SPL Stored Procedure Language
SQL Structured Query
TCB Thread Control Block
TMU Table Management Utility
TS Tablespace
UDB Universal DataBase
UDF User Defined Function
UDR User Defined Routine
URL Uniform Resource Locator

Abbreviations and acronyms 635

Related publications

The publications listed in this section are considered particularly suitable for a
more detailed discussion of the topics covered in this redbook.

IBM Redbooks
For information on ordering these publications, see “How to get IBM Redbooks”
on page 638. Note that some of the documents referenced here may be available
in softcopy only.
Data Modeling Techniques for Data Warehousing, SG24-2238
Business Performance Management . . .Meets Business Intelligence,
SG24-6349
Preparing for DB2 Near-Realtime Business Intelligence, SG24-6071
SG24-6653

Other publications
These publications are also relevant as further information sources:
“IBM Informix Database Design and Implementation Guide,” G215-2271.
Ralph Kimball, The Data Warehouse Toolkit: Practical Techniques for Building
Dimensional Data Warehouses, John Wiley & Sons, 1996.
Ralph Kimball, Laura Reeves, Margy Ross, and Warren Thornthwaite, The
Data Warehouse Lifecycle Toolkit: Expert Methods for Designing, Developing,
and Deploying Data Warehouses, John Wiley & Sons, 1998.
Ralph Kimball and Margy Ross, The Data Warehouse Toolkit: The Complete
Guide to Dimensional Modeling, 2nd Ed., John Wiley & Sons, 2002.
Flemming, C., Handbook of Relational Database Design, Addison-Wesley,
1989.
Yourdon, E., Modern Structured Analysis, Yourdon Press, 1989.
Federal Information Processing Standards Publication 184, “Integration
Definition for Information Modeling (IDEF1X),” December 1993.


Claudia Imhoff, Nicholas Galemmo, Jonathan G. Geiger, Mastering Data
Warehouse Design: Relational and Dimensional Techniques, John Wiley &
Sons, 2003,
Toby J. Teorey, Database Modeling & Design, The Morgan Kaufmann Series
in Data Management Systems, Academic Press, 1999.
W. H. Inmon, Building the Data Warehouse, John Wiley & Sons, 2005.
W. H. Inmon, Claudia Imhoff, Ryan Sousa, Corporate Information Factory,
2nd Edition, John Wiley & Sons, 2000.
Melissa A. Cook, Building Enterprise Information Architectures, Prentice-Hall,
Inc., 1996.
Lou Agosta, The Essential Guide to Data Warehousing, Prentice-Hall, Inc.,
1999.

How to get IBM Redbooks
You can search for, view, or download Redbooks, Redpapers, Hints and Tips,
draft publications and Additional materials, as well as order hardcopy Redbooks
or CD-ROMs, at this Web site:
ibm.com/redbooks

Help from IBM
IBM Support and downloads
ibm.com/support

IBM Global Services
ibm.com/services


Index
Bitmap index 315, 538
Numerics boolean data type 461
3NF 1, 62, 79, 266, 334, 342
bounded box index 539
also see third normal form
BPM 31–32, 35, 99
3NF data warehouse 81
BPM - also see business performance management
5W-1H rule 229
BPM framework 33
bridge table 185, 397
A Brio 98
accumulating fact table 127 B-Tree index 315
additive 53 B-tree index 538
Additive fact 299 B-tree+ index 523
aggregate expression 562 Business analysis 29
aggregate navigation 312, 422 business intelligence xiii, 4, 16, 21, 32, 47
aggregate navigator 312 also see BI xiii
aggregates 308 Business measurements 30–31
aggregation approaches 309 business meta data 451
alerts 38 Business Objects Universe 496
analysis 5 business performance management 5, 22, 31
analytic applications xiii, 35, 38, 72 also see BPM
analytic structures 42 functional components 33
analytical area 69, 73 business process analysis
Apache SOAP server 28 summary 244
architecting a data mart 41 business process assessment factors 232
arithmetic modulus operand 538 business process management 99
artifacts 607 business processes xiii, 120
Ascential Software 482 business requirements
ASCII 2 changing 226
atomic 70 business rules 35, 456
atomic dimensional model 342 Business users 84
atomic grain 122–123 BusinessObjects 98
atomic level data 308 Business-wide enterprise data warehouse 230
attributes 51, 70, 270
adding new 225
auto-index 542
C
calculated fact 299
Auto-index (table) access 540
cardinality 56, 70, 289, 316, 526, 540, 567
Casual users 84
B change handling strategy 156
balanced hierarchy 144, 278 check constraints 580
how to implement 144 chunk 595
balanced time hierarchy 144 classification of users 83
band range value 166 Clickstream analysis 29
BI 2, 7, 23, 35 closed-loop analytics 36
also see business intelligence closed-loop processing 39


clustered index 315, 318 data modeling life cycle 47, 66
code reuse 25 data modeling techniques 47
Cognos Impromptu 98 data object naming 462
Cognos Impromptu Catalog 496 data partitioning 319
co-located join 586 data standards 457
command parser 502, 572, 579–580 data warehouse 4–5, 34, 47, 341
Common Object Request Broker Architecture 26 architectural approaches 57
also see CORBA data warehouse architecture 41
communication buffer 581 data warehouse keys 459
composite primary key 203, 301–302 data warehousing xiii
concatenated index 520, 530, 532, 575 DataStage 484
concatenated primary key 202, 301 Content Browser 485
conceptual design 69 DataStage Manager 496
conditional join 552 date
conformed dimensions 61, 236, 250, 259, 268, 437 data type 460
conformed facts 61, 236, 250, 298, 414 dimension or fact 281
container 595 date and time dimensions 108, 117, 127, 139, 279,
content denormalization 468 399, 435
continuous loading 37 international time zones 142
CORBA 26 International times zones 282
also see Common Object Request Broker Archi- date attributes 281
tecture date dimension 280, 282
correlated subquery 574, 588 DB2 Command Center 223
cost based query optimizer 18, 551 DB2 Cube Views 212, 313
CRM system 337 DB2 Explain facility 215
cube meta data 313 DB2 OLAP Server 97
Customer Relationship Management 337 DB2 optimizer 217, 313
Customer Service 29 DB2 SQL compiler 215
DB2 UDB 212, 215, 540
DCOM (Distributed Common Object Model) 26
D DDL 75
dashboard 35, 82, 94, 98
also see Data Definition Language
data analysis techniques 77
Decision Support 13
data analysis tools 83
default column constraints 580
data architecture xiii
degenerate dimension 135, 137, 192, 302, 316,
data cubes 82
435
Data Definition Language 67, 602
degenerate dimensions 134, 266, 268
also see DDL
denormalization 50
Data Dictionary 481
denormalize 113
data distributions 591
denormalized 55, 116
data domains 458
denormalized E/R models 48
data driven 101
denormalized tables 335
data federation xiii
dependent data mart 40, 47, 61
data lineage 486
architecture 65
Data Manipulation Language 603
dependent data mart architecture 65
data mart 5, 34, 40
Dependent data warehouse 230
data mart consolidation 5, 22, 42, 230
derived fact 299
data mining 29, 100
Development Life Cycle 18
data model xiii, 4
dicing 88


Digital dashboard 29 disadvantage 52
dimension EAR - see Entity, Attribute, Relationship
also see conformed dimensions 268 Eclipse 28
fact valule band 316 encryption key 578
grain 259 Enterprise Data Architect 608
hierarchies 143 enterprise data warehouse 47, 62, 68
meta data 323 architecture 62
dimension keys 459 components 62
dimension meta data 322 Enterprise portal 29
dimension table 53, 259, 270 Enterprise Resource System 337
attributes 259, 269 Enterprise users 84
characteristics 55 entities 51
dimensional attributes 269 Entity Relationship Modeling - see E/R modeling
dimensional data marts 82 Entity, Attribute, Relationship 4
dimensional design life cycle 17 also see E/R
dimensional model 15, 48, 52, 71, 266, 285 ETL 39, 60, 68, 72, 75, 123, 305, 454
handling changes 224, 285 ETL process 41
types 55 event-based fact tables 205
dimensional model design life cycle 227 Explorer 495
also see DMDL expression operands 595
phases 227–228 Extensible Markup Language 482
dimensional model development 333
dimensional model life cycle 6
dimensional modeling 4, 10, 12, 16, 229
F
fact 15, 191, 293
dimensions
Additive 134, 299
activities for identifying 257
aggregation rules 300
preliminary 255
also see conformed facts 298
DMDL 6, 118, 133, 227, 320, 330, 333, 351, 472
Derived 134, 299
also see dimensional model design life cycle
derived 300
the phases 228
Factless 299
dominant table 553, 566
identification activities 293
drill-across 92
Non-additive 191
drill-down 87, 90–91
non-additive 134, 192, 296, 299
drilling 90
preliminary 255
driveSpace 595
Pseudo 134, 299
duplicate index 531, 558
Semi-additive 134
semi-additive 193, 299
E Textual 134, 191, 299
E/R data model 8, 10, 12, 15 types 298
also see E/R model year-to-date 134, 300
E/R diagram 48, 343 fact table 15, 52–53, 191, 275, 287, 296
E/R model 48, 62, 64, 71, 117, 232, 238, 334 accumulating 124, 251
also see E/R data model adding new facts 225
converting to dimensional 108, 297 calculate growth 304
non-transaction based tables 109 characteristics 53
transaction based tables 109 comparison of types 251
E/R modeling 1, 10, 49 criteria for multiple 249
advantages 51 event 301

Index 641

growth 303 granularity 246, 254
meta data 326 date and time dimensions 279
multiple grains 120, 250 fact table 247
periodic 124, 251 impact on storage space 124
size 303 multiple fact tables 249
transaction 124, 250 trade-offs 253
types 124, 250
fact table granularity 247
guidelines for choosing 248
H
hash index 18, 538, 541
fact tables
hash index (table) join 542
event 205
hash table join 586
fact value band 316
heterogeneous dimension 291, 407
factless fact table 208, 211, 299
hierarchies 269
Factors 123
Balanced 143, 278
fast changing dimensions 164, 286, 400
multiple 278
activities 286
Ragged 154, 278
approaches for handling 286
Unbalanced 145, 278
Federated Database 42
hierarchy 213
filter columns 559
histograms 597
filter correlation 591
hot swappable dimension 188, 292, 408
filter selectivity 518
implementing 188
Finance 29
HTML 7
first normal form 71
also see Hyper Text Markup Language
floors of data 78
HTTP 27
foreign dimension keys 303
hub and spoke 59, 492
foreign key 53, 74, 109, 168, 202, 274, 286–287,
Human Resources 29
301, 315, 622
Hyper Text Markup Language 7
foreign key definitions 580
also see HTML
fragment elimination 512, 593
Hyperion Essbase 98
functional dependencies 71
future business requirements 124
I
IBM Alfablox 97
G Identify aggregates 422
garbage dimension 176, 291, 408
Identify business process requirements 227
Geo-spatial analysis 29
Identify the business process 229, 422
getting data in 37, 39, 497
identify the business process
getting data out 37, 39, 498
multiple grains 249
Globally Unique Identifiers 265
Identify the dimensions 120, 227, 257, 323
grain 120, 123, 185, 209, 245, 272
activities 257
characteristics 245
Identify the facts 120, 227, 326, 411
factors to consider 123
Identify the grain 120, 227, 423
having multiple grains 249
IIOP (see Internet Inter-ORB Protocol)
grain atomicity 252, 434
impact analysis 487
grain definition
independent data mart 40, 47, 58, 63
changing 225
issues 60
grain definition report 323
Independent data warehouse 230
grain definitions 246, 293–295, 302
index leaf page data 538
grain identification 245
index maintenance 315


index negation 18, 582 L
index types 314, 317 left handed join 552
join index 317 list of nouns 70
MDC indexes 318 load trial tool 609
multidimensional indexes 318 logical data model 69
selective index 318 design activity 69
virtual index 317 logical data modeling 66
Index utilization guidelines 519 logical model
Indexed (table) access 538 validation 70
indexed retrieval 537 logical process model 608
indexed star schema 220
indexed table access method 514
indexing 314, 422 M
many-to-many relationships 105
indexing dimension tables 316
Marketing 30
indexing techniques - vendor specific 317
Materialized Query Tables 42, 212, 313, 573
information dashboards 38
materialized views 573
also see dashboard
messaging systems 39
Information Integration 40
meta data 18, 25, 41, 321, 325, 328, 423, 447
information pyramid 16, 23, 77–78
cube 313
information technology 1
non-technical 321
Information visualization 29
techincal 321
Informix Software 7
meta data design 453
Ingres Software 7
meta data directory 490
inner query 574
meta data history 455
interconnected data mart 61
meta data management 227, 320, 422
Internet Inter-ORB Protocol 27
meta data management system 341
intersect entity 106
meta data model 454
IT 24
meta data standards 457
IT users 84
meta data strategy 454
meta data types
J business 451
Java RMI (Remote Method Invocation) 26 Design 483
Java Web Application Server 28 Operational 484
Java/J2EE 3 operational 453
join column key 553 Physical 483
join cost 513 reference 452
join criteria 552 structured 452–453
junk dimension 176 technical 452
unstructured 453
MetaArchitect 493
K
key performance indicators 32, 38 MetaBroker 492
also see KPI MetaBrokers 483
key-only retrieval 514, 531 metadata
KPI categories 491
also see key performance indicators class hierarchy 490
relationship hierarchy 490
metadata repository 455
metadata types 451

Index 643

MetaStage 482 non-technical meta data 321
Microsoft Excel 486 non-unique index 314
Microstrategy 97 Normal Forms 15
mini-dimension table 287 normalization 50, 72
mini-dimensions 166, 225, 286, 288 normalization of dimension tables 173
Model-View-Controller 605, 608, 618 normalization of entities 69
Monitor 320 normalized database 342
MQSeries queues 39 normalized E/R models 48–49
mulltistar model 55 Not Applicable scenario 274, 282
multidimensional analysis 77, 86, 100 Not Applicable scenario records 260
MultiDimensional Clustering 18, 42, 593 noun list 70
Multidimensional Cube Analysis 29 null keys 274
multidimensional structures 69, 97
multidimensional techniques 87
multi-index scan 576, 586
O
Object Management Group 482
multilingual information 471
Object Request Broker 27
multiple granularities 434
ODS. See Operational Data Store
multiple hierarchies 278
OLAP 42, 420, 481
multi-stage back end command parser 579
also see Online Analytical Processsing
multi-threaded process architecture 500
OLAP databases 34
multi-user concurrency 538
OLTP 2, 6–7, 16, 24, 49–50, 135, 230
multi-valued dimension 183, 290, 406
also see on-line transaction processing
OnLine Analytical Processing 42
N also see OLAP
naming data 462 on-line transaction processing 1, 6
abbreviations 463 also see OLTP
acronyms 464 operational data 79
general considerations 462 Operational Data Store 5, 42
glossary 464 operational meta data 453
historical considerations 462 Operations 29
near real-time BI 39 optimizer directives 511
near real-time business intelligence 36–37 Oracle Software 7
negation of index 536 outer join 552
Negation of index, (poor) selectivity of filter 583
Negation of index, non-anchored composite key
583
P
Parallel ETL Engines 39
Negation of index, non-initial substring 582
parallel operations 512
nested loop (table) join 541, 544
Parser 501
nested loop join 514, 586
partial index negation 583
nested queries 573
partitioning 319, 422
non-additive fact 191, 299
performance design and tuning 75
non-anchored composite key 536
performance metrics 32
non-anchored key 583
periodic fact table 126
non-blocking architecture 500
physical data modeling 67, 73
non-indexed star schema 220
Physical design considerations 227
non-key columns 438
aggregations 308
non-outer join 556
meta data 329
non-pipelined sort 570, 587
physical model


design factors 74 query rewrite 18, 548, 568, 575
physical modeling query rewrite directive 588
activities 74
pipelined parallel architecture 500
pipelined sort 570, 587
R
ragged hierarchy 154, 278
Pivoting 87, 90
Rational Data Architect 333
Platinum ERwin 484
read ahead scans 537–538
Power users 84
real-time 36
pre-aggregated data 215
real-time business intelligence 5, 22, 40
predicate inference 591
real-time information xiii
pre-fetching 537
Recover 320
preliminary dimensional schema 262
recursive loop 544
preliminary dimensions 410
Redbook Vineyard 336, 400
preliminary facts 256, 295, 410, 412
business process description 344
primary key 70, 74, 106, 156, 202, 259, 263, 301,
the project 334
315, 622
Redbooks Web site 638
composite 301
Contact us xvi
concatenated 301
redundant join 577
uniqueness 303
reference meta data 452, 467
primary key definitions 580
referential integrity 260, 274
primary key resolution 72
relational database
primary query rules 551
administrative interfaces 508
process management xiii
character based interface 508
Process MetaBroker 494
fragment elimination 512
project plan 9
indexed access 514
pseudo fact 209, 299
join cost 513
pseudo order-by clause 565
key-only retrieval 514
Publish and Subscribe 493
nested loop join 514
push down semi-hash join 546
outer cartesian product 510
page corners 508
Q parallel operations 512
QualityStage 492 sequential scan 514
quantile 592 table join methods 514
query 5 threads 512
Query analysis and reporting 85 relational database tuning 505
Query and reporting 29 example activities 505
query and reporting applications 97 relationships 51
query and reporting tools 94 remote path (table) access 539
Query Builder 488 Remote Procedure Call 27
Query definition 85 Reorganize 320
query optimizer 18, 497, 548, 568, 575, 605 repeating groups 71
directives 511 Replication 39
query optimizer directives 18, 536, 584 reporting 5
query optimizer hints 534 reporting environment 82
query optimizer histograms 597 reporting tool architectures 83
query optimizer statistics 541 requirements
query plan 515 changing 226
query processor 501, 503 plan for the future 124

Index 645

requirements analysis 242 snowflaking 168, 172, 288, 404
entities and measures 242 SOA 6
requirements gathering 240 also see service oriented architecture
maintaining history 242 SOAP (also see Simple Object Access Protocol)
questions 241 software development life cycle 8, 601
requirements gathering approach 237 also see SDLC
source-driven 238 software server components 499
user-driven 239 disk
requirements gathering report 323 memory
Restructure 320 process
right handed join 552 sort merge (table) join 542, 544–545
right time data delivery 79 sort order 543
RMI - Remote Method Invocation 27 source systems 62, 65, 79, 135
ROLAP 92 source systems query 80
role-playing 279 splitting a tall table 594
role-playing dimension 179, 290, 406 spreadsheet 94, 97
roll-down 93 SQL 5, 7
roll-up 87, 93, 140 also see Structured Query Language
ROWID (table) access 540 SQL API 607
RPC (also see Remote Procedure Call) SQL CURSOR 580
R-tree index 539 SQL error recovery 605
rules based query optimizer 551 SQL query optimizer 18
also see query optimizer
SQL statement tuning 505, 600
S SQL VIEWS 573
scorecard 98
staging area 62, 65
SDLC 601
standard join 556
second normal form 71
star index 539, 570
semantic data integrity constraints 580
star model 55
semi-additive 53
star schema 52, 123, 174, 208, 223, 271, 298, 313
semi-additive fact 193, 199, 299
statement cache 580
semi-join 574
stored procedures 95
Sequential (table) access 538
structured meta data 452–453
sequential scan disk access method 514
Structured Query Language 5
sequential scans 537
also see SQL
Sequentially scan 320
subordinate table 553
service oriented architecture 6
summarized data 81
- also see SOA
summary area 68, 73, 342
shared aggregation 577
summary tables 42, 212
Sherpa & Sid Corporation 426
surrogate keys 107, 113, 259, 263–264, 436
Simple Object Access Protocol 26
reasons for use 263
Slice and Dice 87, 221
system of record 68, 340, 342
Slicing 88
system test phase 9
slowly changing dimensions 155, 283, 400
activities 285
types 283 T
snowflake model 55 table access methods 502, 512
snowflake schema 174 table join methods 18, 514
snowflaked dimension table 175 table join order 18


table partitioning 18, 319, 594 handling history 306
tactical business intelligence 35 volume 595
tall table 594
technical meta data 321, 452
temporal data 10
W
W3C (also see World Wide Web Consortium)
temporary index 537
Waterfall Method 8, 601
textual fact 299
Web services 25, 28, 40
third normal form 1, 49–50, 71
also see XML Web services
also see 3NF
architecture 27
threshold values 38
Web Services Description Language 26
time 283
WebSphere
a dimension not a fact 283
queues 39
a dimension or a fact 139
WebSphere Application Server 28
as a fact 283
WebSphere Information Integrator 34, 313
data type 461
WebSphere MQ 27
in dimensional modeling 139
WebSphere Studio 28
time dimension 282
weighting factor 184
topped query 575
World Wide Web 16, 25
transaction fact table 125
World Wide Web Consortium 26
transitive dependencies 71, 548, 566, 571
WSDL 27
triggers 580
also see Web Service Description Language
Type 1 slowly changing dimension 284
Type 2 slowly changing dimension 284
Type 3 slowly changing dimension 284 X
Type-1 slowly changing dimension 156 XML 27
Type-2 slowly changing dimension 158 XML Web services 25
Type-3 slowly changing dimension 161 also see Web services
types of dimensional models 55
Y
U year-to-date facts 420
UDDI 26
also see Universal Desctiption Discovery and In-
tegration
UDDI registry 28
unbalanced hierarchy 145, 278
UNION ALL 576
UNION ALL SQL SELECT 576
UNIONED SELECT 576
unique index 314, 531, 558
Universal Description, Discovery, and Integration
(see UDDI)
unstructured meta data 453
user classification 83

V
value chain 33
Verify the model 227
business requirements 305

Index 647

Dimensional Modeling: In a Business
Intelligence Environment
(1.0” spine)
0.875”<->1.498”
460 <-> 788 pages

Back cover ®

Dimensional Modeling:
In a Business Intelligence
Environment
Dimensional In this IBM Redbook we describe and demonstrate
modeling for easier dimensional modeling techniques and technology, INTERNATIONAL
data access and specifically focused on business intelligence and data TECHNICAL
analysis warehousing. It is to help the reader understand how to SUPPORT
design, maintain, and use a dimensional model for data ORGANIZATION
warehousing that can provide the data access and
Maintaining
performance required for business intelligence. Business
flexibility for growth
intelligence is comprised of a data warehousing
and change infrastructure, and a query, analysis, and reporting, BUILDING TECHNICAL
environment. Here we focus on the data warehousing INFORMATION BASED ON
Optimizing for query PRACTICAL EXPERIENCE
infrastructure. But only a specific element of it, the
performance dimensional model, or, more precisely, the topic of
dimensional modeling and its impact on the business and IBM Redbooks are developed by
business applications. The objective is not to provide a the IBM International Technical
treatise on dimensional modeling techniques, but to focus at Support Organization. Experts
a more practical level. There is technical content for designing from IBM, Customers and
Partners from around the world
and maintaining such an environment, but also business create timely technical
content. For example, we use case studies to demonstrate information based on realistic
how dimensional modeling can impact the business scenarios. Specific
intelligence requirements. In addition, we provide a detailed recommendations are provided
discussion on the query aspects of BI and data modeling. For to help you implement IT
solutions more effectively in
example, we discuss query optimization and how to evalueate your environment.
the performance of the data model prior to implementation.
You need a solid base for your data warehousing
infrastructure . . . . a solid dimensional model.
For more information:
ibm.com/redbooks

SG24-7138-00 ISBN 0738496448

Dimensional modeling in a bi environment

More Related Content

What's hot (18)

Similar to Dimensional modeling in a bi environment (20)

More from divjeev (17)

Recently uploaded (20)

Dimensional modeling in a bi environment