Эволюция Big Data и Information Management. Reference Architecture.

Information Management
Reference Architecture
3rd Evolution
EMEA Enterprise Architecture

Contents
 Introduction
 Conceptual view
 Design Patterns
 IM Logical view and component outline
 Discovery Lab
 R/T Event Engine logical view
 Mapping to previous Reference Architecture release

Introduction
 This PPT documents the main architectural components of Oracle‟s
Information Management Reference Architecture.
 The architecture is intended to be practical and pragmatic, with many of the
ideas and experiences that inform the approach dating back almost 20 years in
Oracle and are based on real world customer experiences.
 We define Information Management to mean the following. Please note that
this definition embraces all types and forms of data as well as embracing
aspects such as Information Discovery and Governance:
“Information Management is the means by which an organisation maximises the efficiency
with which it plans, collects, organises, uses, controls, stores, disseminates, and disposes
of its Information, and through which it ensures that the value of that information is
identified and exploited to the maximum extent possible”
3rd Evolution of Oracle‟s Information Management Reference Architecture

Oracle’s Information Management Reference Architecture (3rd Edition)
 More relevant to Big Data oriented audience
 Better representation of pragmatic customer projects
 Includes Raw data store as part of the architecture
 Show effort / cost to store and interpret data that separates
schema-on-read and schema-on-write approaches
 Aligned to Analytics 3.0
 Consistent with Oracle‟s engineering efforts
What‟s changed?

Aligning analytical requirements and IM architecture
Enabling Analytics 3.0 with a pragmatic architecture
Analytics 2.0
Analytics 3.0
Analytics 1.0
• Reporting with limited use of
descriptive analytics
• Limited range of tabular data
• Batch oriented analysis
• Analysis bolted onto limited
set of business processes
• Firms “Competing on Analytics”
• Extended analytics to larger
and less structured datasets
• Emergence of Big Data into the
commercial world
• Recognition of Data Science
role in commercial orgs.
• Platform for monetisation
• Deeper analysis & more data
• Faster test-do-learn iterations
• Different types of data & wider
business process coverage
• Analysts focus on discovery and
driving business value
• “Agile” with operational elements
incorporated into design patterns
Adapted from Tom Davenport material

Oracle’s Information Management Reference Architecture (3rd Edition)
“All those layers and definitions in your
Reference Architecture, I just don’t get
it… and it looks complicated !”
Hadoop developer knee deep in complex Map:Reduce code
What‟s changed?
Business
Trends
Technology
Trends
Data
Trends

Actionable
Events
Event Engine Data
Reservoir
Data Factory Enterprise
Information Store
Reporting
Discovery Lab
Actionable
Information
Actionable
Insights
Input
Events
Execution
Innovation
Discovery
Output
Events
& Data
Conceptual View
Structured
Enterprise
Data
Other
Data

Component Outline
Data Engine Respond to R/T events in appropriate and/or optimised fashion
Data Reservoir Raw data Reservoir – typically event data at lowest grain
Data Factory Managed ETL onto, within and between platforms
Enterprise Data Data stores for Information Management
Reporting BI tools and infrastructure components
Discovery Lab Platform, data and tools to support discovery process
Execution – things you do every day
Innovation – innovation to drive tomorrows business
Line of Governance!
Discovery
Output
– Possible outputs include new knowledge, mining models / parameters, scored data…

Design Pattern: Discovery Lab
 Specific focus on identifying commercial value for exploitation
 Small group of highly skilled individuals (aka Data Scientists)
 Iterative development approach – data oriented NOT development oriented
 Wide range of tools and techniques applied
 Data provisioned through
Data Factory or own ETL
 Typically separate infrastructure
but could also be unified Reservoir
if resource managed effectively

Design Pattern : Information Platform
 Build the next generation Information Management platform
 Either Business Strategy driven or IT cost / capability driven initiative
 Initial project may be specifically linked to lower data grain or retention
BUT it is the platform as a whole that forms the solution required
 Platform for consolidating other IM assets onto
 Key issues related to differences in
procurement, development process,
governance and skills differences
 Discovery Lab may be implemented
as a pragmatic initial POV.

Design Pattern : Data Application
 Big Data technologies applied to a specific business problem
e.g. Genome sequence analysis using BLAST or log data from
pharmaceutical production plant and machinery required for traceabiliy
 Limited or no integration to broader Information Management estate
 Specific solution so Non-functional requirements have less impact
on solution quality or long term costs
 Platform costs and scalability are
important considerations

Design Pattern: Information Solution
 Specific solution based on Big Data technologies requiring broader
integration to the wider Information Management estate
e.g. ETL pre-processor for the DW or affordably store a lower level of grain
 Non-functional requirements more critical in this solution
 Scalable integration to IM estate
an important factor for success
 Analysis may take place in Reservoir
or Reservoir only used as an aggregator

Design Pattern: Real-Time Events
 May take place at multiple locations between place of data origination and the
Data Centre – requiring careful design and implementation
 May include Next-Best-Activity, declarative rules and Data Mining technologies
to optimise decisions. i.e. optimise across declarative, data mining, customer
preference & business-defined rules
 May include considerations for
personal preferences and privacy
(e.g. opt-out) for customer related
events
 Common component seen across
many industries & markets
e.g. connected vehicle
Real-Time optimisation of events

Design Pattern against component usage map
Design pattern Discovery Lab
Information
Platform
Data Application Information Solution R/T Events
Outline
Data science lab
Assess the value of
the data
Next Generation
information platform to
align IM capability with
business strategy
Addressing a specific data
problem in Hadoop with no
broader integration required.
Addressing a specific data
problem but requires broader
enterprise wide integrations. e.g.
ETL pre-processing, Event Store
at lower grain than existing DW
Execution platform to
respond to R/T events
Examples
Gov. Healthcare
Mobile operator
Spanish Bank (Business led)
UK Gov. Dept. (Tech. led)
Pharma Genome project
Pharma production archive
Investment Bank – trade risk
Mobile Operator – ETL processing
Mobile operator –
location based offers
Data Engine Possible Yes
Data Reservoir Yes Yes Yes
Data Factory Yes Yes Yes
Enterprise Data Yes
Reporting Yes
Discovery Lab Yes Implied Alternative approach
to Reservoir + Factory above

IM Logical View and
Components

Information Management – Logical View
Data Sources
Data Ingestion
Methods and process
to load data into our
managed data store
and manage data
quality
• Contemporary Information Management solutions must be able to ingest any type of data from any source in any format and
mechanism and at any frequency. e.g. Flat file loads, streaming…
• The data may be highly unstructured, mono-structured or highly poly-structured.
• Data will vary in volume and in Data Quality.
• Operational isolation should be considered to ensure operational applications will continue in the event of the loss of the
Information Management system
Data Engines &
Poly-structured
sources
Content
Docs Web & Social Media
SMS
Structured
Data
Sources
• Operational Data
• COTS Data
• Master & Ref. Data
• Streaming & BAM

Information Ingestion
Data Ingestion
Information Interpretation
Methods and process
to load data and
manage Data
Quality
Methods and
process needed to
access information
Managed Data
Load
All data under management
Query
• Data structure and processing required to load data into managed data stores
• Shape represents the work done on the data to load data and/or process between layers
• Layer may include file mechanism where required to facilitate loading
(e.g. Fuse fs or ZFS for operational isolation and file concat)
• Normal rules of micro-batch, taking all the data and KISS principles recommended
• DQ and loading stats presented through BI dashboards as a non-judgemental mechanism to improve DQ.
• Data may be landed in the Ingestion layer to facilitate loading but not typically stored for any length of time. e.g. Raw data loaded from web
logs but sessionised data then loaded to Raw. Another example is data used to manage CDC may be stored in this layer.

Data Interpretation
Data Ingestion
Methods and process
to load data and
manage Data
Quality
Methods and
process needed to
access information
Managed Data
Load
All data under management
Query
• Methods and processes required to access information in each of the stores
• Shape represents the cost of interpreting the data under management
• For schema-on-read the cost may include the AVRO, SerDe or reader class as well as the associated processing code to
select, filter and process the data.
• For schema-on-write the cost is represented by the complexity of the SQL required to access the data only – more complex
typically for 3NF than for a dimensional query.

Data Layers – cost, quality and concurrency trade off
Managed DataAccess & Performance Layer
Foundation Data Layer
Raw Data Reservoir
Immutable raw data reservoir
Raw data at rest is not interpreted
Immutable modelled data. Business
Process Neutral form. Abstracted
from business process changes
Past, current and future interpretation of
enterprise data. Structured to support
agile access & navigation
• Increasing enrichment
• Increasing data quality
• Reducing concurrency costs
• Data under management includes 3 key layers – Raw, Foundation and Access and Performance layers.
• Data normally loaded into Raw and Foundation layers BUT BI Apps loads data directly into APL and federated warehouses may
well also load data at aggregate level from federated operating companies.
• Data Factory is responsible for loading and then managing data between layers.
• Work is done to elevate the data between layers – typically further enriching and improving data quality.
• Work done in processing the data between the layers significantly reduce query costs. i.e. higher levels of concurrency can be
sustained for the same processing power.
• Increasing formalisation of definition

Data Layers – Analytical processing
Raw Data Reservoir
• Analytical processing capabilities of Hadoop and RDBMS used to elevate data between layers as previously described.
• These analytical capabilities can also be leveraged by tools that access the data directly.
Typically this would be by a Data Scientist for Discovery Lab operations or BI Tools and Services that are processing data using
a model previously defined by the Data Scientist.
OLAP
Data Mining
Statistics
OLAP
Text Mining
Other
Analytical
Processing
Data Mining
Text Mining
Image
Processing
• Increasing enrichment
• Increasing data quality
• Reducing concurrency costs
• Increasing formalisation of definition

Data Layers – Raw Data Reservoir
Raw Data Reservoir
• Immutable data store with data at lowest level of grain.
• Typically implemented in Hadoop or NoSQL for cost reasons but not always.
• May be:
• Queries directly,
• Used to derive base level data for Foundation Layer. Data may be represented logically in Foundation or physically as the
store is immutable BUT this effects ILM policy.
• or used to derive values or aggregates for Access and Performance layer. (e.g. propensity score or total monthly SMS‟s)

Data Layers – Foundation Data Layer
Raw Data Reservoir
• Immutable integrated and standardised store of enterprise class data. Stuff the business has agreed and organises around.
• Data at lowest level of grain of value for Enterprise data.
• Stored in business process neutral fashion to avoid data maintenance tasks to keep in step with current business interpretations.
• Typically close to 3NF. Special attention to modelling hierarchy, flexible entity attributions, customer / supplier etc.
• ONLY implemented in relational technology BUT this could be logical as previously noted in Raw Data Reservoir.
• May be queries directly by a select few individuals. Wider access to detail data provided through views in APL, often with VPD
implemented to prevent queries to antecedent data.
• Data in the Foundation Layer should be retained for as long as possible.
• Consideration should be given to retaining data in Raw Data Reservoir rather than archiving.

Data Layers – Access and Performance Layer
Raw Data Reservoir
• Layer facilitates access, navigation and performance of queries.
• Allows for multiple interpretations of data from Foundation or Raw data Reservoir.
• Most structures can be thrown away and re-built from scratch based on Foundation and Raw Reservoir.
• The exception is derived and aggregate data which may have to be retained if the underlying data/mechanism is archived.
• Most users presenting information in a standardised fashion on dashboards and reports will access this layer only.

Access and Performance Layer
Access & Performance Layer
Raw Data Reservoir
Data Engines &
Poly-structured
sources
Content
SMS
Structured
Data
Sources
• COTS Data
• BAM Data
• Data destined for Raw Data Reservoir may be loaded directly (e.g. through Flume) or may be stored temporarily in fs prior to
loading (e.g. Fuse fs)
• Relational data ingested in most appropriate mechanism before persisting in Foundation Data Layer (usual rules apply…)
• Ideally micro batch using simplest mechanism possible
• Only data of agreed quality loaded in FDL
• For efficient loading relationally data may be pre-staged in fs so a large number of small files can be concatenated
Data Factory Ingestion flow
Data Ingestion
Batch & Real-Time
ETL / ELT
CDC
Stream
File Ops.

Data Ingestion
Raw Data Reservoir
Flow shown:
1. Data to be formalised from HDFS store extracted and loaded into Foundation Data Layer.
e.g. where Flume/HDFS is being used as an ETL pre-processor for Enterprise Data
or where HDFS data is being logically modelled in the foundation layer
2. Data is re-structured and/or aggregated to facilitate access by users and business processes
3. Data may also be re-structured and/or aggregated from HDFS store where there are no specific
requirements to manage Enterprise Data in a more formal data store over time
1
2
3
Data Factory intra data processing flow

Information Provisioning – BI & Data Science Components
Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc BI Assets
Information
Services
Data Ingestion
Raw Data Reservoir
Virtualisation&
QueryFederation
• Data Virtualisation and the various components to access the data are as per our previous view on BI tools.
• Data Virtualisation is a key components that helps to deliver tools independence, services integration and a future state roadmap
• Big Data has focused considerable attention on Data Science
• Analytical capabilities delivered through analytical processing in the data layers and Advanced Analytical Tools used to drive capabilities
• Data Mining in particular often involves complex data processing to flatten data into a longitudinal form. This derived data and model results
are typically written to a project based sandbox.
• Agile discovery is often best served through a separate Discovery Lab infrastructure (see later details)
Data Science

Information Provisioning BI Flows
Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc BI Assets
Information
Services
Data Science
Data Ingestion
Raw Data Reservoir
2
3
1. Typical access mechanism for Enterprise data via Access and Performance layer structures
2. Access to Foundation Layer Data to specific functions, processes and users only
3. Data interpretation & DQ assured through encoded logic, Avro, SerDe, FileReader, HCat etc.
4. Diagonal flows shows how data can be joined between layers as well as accessed directly. e.g. Raw Data
can be queried directly through HIVE connector or joined to the RDBMS data and queried.
1
4
4

Data / Information Quality
Data Ingestion
Raw Data Reservoir
Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc BI Assets
Information
Services
Data Science
 Quality of data at rest assured by a number of factors in addition to the underlying quality of data at source
– File and event handling to ensure data is not missed (e.g. missing log files assured by log file sequence numbering)
– The processing of data between Raw and FDL / APL layers. This can be seen as a DQ firewall to ensure only data of known and
acceptable quality is loaded. Typically this involves an element of synchronisation as some data will need to be held off until required
reference data is available due to the micro-batch incremental loading approach.
 Quality of information presented to downstream tools and services determined by
– Model quality, understanding and performance of provisioning from modelled layers
– Consistency of definition, code quality and query performance when accessing Hadoop data (e.g. HR code, Avro definition…)

Data Reservoir & Enterprise Information Store
Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc
BI Assets
Information
Services
Data Ingestion
Raw Data Reservoir
Data
Science
Data Engines &
Poly-structured
sources
Content
SMS
Structured
Data
Sources
• COTS Data
• Streaming & BAM

Analysis Processing & Delivery
Discovery Lab & Data Science Tooling
Data Reservoir & Enterprise Data
Data
Science
(Primary
Toolset)
Statistics Tools
Data & Text Mining Tools
Faceted Query Tools
Programming & Scripting
Data Modeling Tools
Query & Search Tools
Pre-Built
Intelligence
Assets
Intelligence
Analysis
Tools
Ad Hoc Query
& Analysis Tools
OLAP Tools
Forecasting &
Simulation Tools
Reporting Tools
Data
Scientist
Virtualisation&
InformationServices
Data Factory
flow
1. Data Factory responsible for
access provisioning to data
or replication (all or sample)
to Sandbox in Discovery Lab.
2. Direct connection from Data
Science tools and analysis
sandbox. Data Science tools
read and write data from/to
project sandboxes.
3. Data Scientist can also
access standard dashboards,
reports and KPI‟s through
Data Virtualisation layer
Data Quality & Profiling
Graphical rendering tools
Dashboards & Reports
Scorecards
Charts & Graphs
Sandbox – Project 3
1
2
Data store
Analytical
Processing
3
Discovery Lab data flow

R/T event Engine – Logical
View and Components

Real-time
Data Engine
To Event Subscribers
(Events / Data)
Privacy Filter
Data Transform
Rules & Models
Mediation
Next Best Action
Real-Time
Data Store
From Input Events
Reference
Data
Models
& Rules
Privacy
Data
Analytics
Real-Time Data Engine – Logical View
Business Activity Monitoring
Real-Time event
monitoring

Real-Time Data Engine
 Message mediation service
 Privacy filter for event data. i.e. apply customer specified privacy
and preference filters to the data stream
 Transformation of the message data to outbound form
 Apply declarative rules and models to the data stream to detect
events for further downstream processing
 Next Best Activity (NBA) event detection and processing. NBA
typically also includes control group management and global
optimisation of rules
 Business Activity Monitoring
 Local data store – local persistence of rules and metadata
Components
Privacy Filter
Data Transform
Rules & Models
Mediation
Next Best Action
Real-Time Data
Store
BAM

Real-Time data engine flows
 Describe each of the data flows
Reference
Data
Models
& Rules
Privacy
Data
Event Analytics
From Input Events
To Event Subscribers
(Events / Data)
R/T Event Monitoring
To Do

Mapping from the previous
release of the architecture

Information Management Reference Architecture
Version 2.0 of the Architecture

Information Management Reference Architecture
Interpretation layer
shows the relative cost
of reading data
depending on its
location
Previous staging layer
now split into Data
Ingestion and Raw
store.
Ingestion layer
includes methods and
processes to load data
and manage Data
Quality. Shape
represents the relative
cost of these
processes. i.e. from
none for HDFS to lots
in APL.
Raw Reservoir is
typically at the lowest
level of grain. Often
lower than the
enterprise cares about
and so may not have
been included in
previous
representation.
Renamed from
Knowledge Discovery
to Discovery Lab but
otherwise unchanged.
The role of Discovery
Labs is becoming
more central though so
additional operational
guidance will be
added.
Discovery Lab
Still an immutable
store but may be
physically
implemented in
relational or non-
relational technologies
Key differences from 2.0 to 3.0 of the Architecture

Эволюция Big Data и Information Management. Reference Architecture.

Discovery Lab and
Governance considerations

Data discovery for the Enterprise
 Discovery phase
– Unbounded discovery
– Self-Service sandbox
– Wide toolset
– Agile methods
 Promotion to Exploitation
– Commercial exploitation
– Narrower toolset
– Integration to operations
– Non-functional requirements
– Code standardisation &
governance
Discovery and monetising steps have different requirements
Business
Value
Commercial
Exploitation
Time / Effort
Discovery phase
Understanding
of the data
Governance

To monetise fully you need to standardise
It‟s smart to standardise as part of Governance
 Discovery process requires
a broad toolset
 Standardisation is essential
for Commercial exploitation
 Sustainability depends on
standardisation / rationalisation
– Reduced training burden
– Reduced support costs
– Reduced license costs
– Ongoing agility & alignment
Data Discovery Toolset Data Exploitation Toolset
Rationalised
Components
• Cloudera CDH, Oracle, No-SQL
• Mammoth, Yarn, EM-plugin
• MR, Hive, Pig, Impala, Accum.
• Flume NG, Oozie
• …
• …
• …
Optional additions
• Oracle Connectors
• Additional corporate standard
components
Oraclestandard
deployment
Corporate
standard
Standardised Hadoop Zoo
Standardised deployment

Copyright © 2013, Oracle and/or its affiliates. All rights reserved.52
The kind of things we are looking to Discover
 Data science skills required
vary by the type of analysis
 Data Management skills vary
by the amount of data and its
structure
 So making data movement
and manipulation easy will
deliver a better result and
deliver it faster
Descriptive
Diagnostic
Predictive
Prescriptive
Business Impact
AnalyticalSkills
Insight
Foresight

Discovery is a Data process not a Development Process
Requirement
Analysis
High Level
Design
Low Level
Design
Coding Testing
Acceptance
Testing
Three Versions of the BI Development Process
Excel
Spreadsheet
Shared linked
spreadsheets
Local Access
Database
Shared Access
Server
SQL Server
Database
Oracle
Datawarehouse
Discovery & Profile Model Exploit
What IT thinks it should be
What normally happens
What Big Data is trying to achieve

Sandbox delivery options
• Separate Data Lab environment
• Delivered as part of Information
Management architecture
Self Service Sandboxes
• Self service provisioning of new
sandboxes for Discovery phase
• Automation of data access rights,
resources and tools provisioning
Data provision
• Quickly take on new data to
rapidly make available to Analysts
• Tools such as “Data Factory” can
fully automate data flows
Sandboxes facilitate “Agile”
Providing the technology platform for agile discovery

Monetise and Optimise steps are different
 New insights deployed into business process in some form
– Technical: e.g. Business rules, new customer segments
– Non-technical: e.g. Observations about behaviours
 Business Intelligence systems adapted to provide
monitoring, feedback and control optimisation
 The faster you iterate this cycle the greater the benefit BUT
 Big Data does not change the fundamental need
for accurate, consistent and integrated information
What happens when we want to exploit insights?
New
insights
Business Process

Rules of thumb for data
Organised information leads to better analyses
Information needs to be organised in order to analyse it
RDBMS are great when information is organised
Hadoop minimises the penalty for disorganisation
The closer you are to insight, the more complete and
organised information needs to be
Data needs to be organised to monetise it effectively

What that really means is…
 We need to apply structure
to data in order to analyse it
 Schema on read works well
for us in Discovery as we can
be agile about interpretation
 As we move into Discovery
schema on read can causes
Governance & quality issues
 Key lesson: The cost to store & manage is distinct from structural
considerations between Big Data and RDBMS technologies

Copyright © 2013, Oracle and/or its affiliates. All rights reserved. Insert Information Protection Policy Classification from Slide 1358
De-mystifying schema on read
DQ
Bus. Rules
Mapping
ETL
Data Reservoirs
 Traditional “Schema on Write”
– Data quality managed by formalised ETL
process
– Data persisted in tabular, agreed and
consistent form
– Data integration happens in ETL
– Structure must be decided before writing
 Big Data “Schema on Read”
– Interpretation of data captured in code for
each program accessing the data
– Data quality dependent on code quality
– Data integration happens in code

Underlying storage capabilities are different
0
1
2
3
4
5
Tooling maturity
Stringent Non-Functionals
ACID transactional
requirement
Security
Variety of data formats
Data sparsity
ETL simplicity
Cost effectively store low
value data
Ingestion rate
Straight Through
Processing (STP)
Hadoop
Relational
My Appllication

Analytics 3.0 platform include
both relational and
non-relational technologies
Ken Rudin* refers to this
as the genius of AND vs the
tyranny of OR
(see his TDWI „13 presentation)
Unified Reservoir simplifies
access to all data regardless of
characteristics & analysis
requirements
It’s smart to unify your data into a single Reservoir
Fully expose your data for discovery and monetisation
Ken Rudin is Director of Analytics at Facebook*
All
Data

Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc BI Assets
Information
Services
Advanced
Analytical
Tools
Information Provisioning Analysis Processing & Delivery
Data Ingestion
Information Access
Raw Data Reservoir
Methods and process
to load data and
manage Data
Quality
Methods and
process needed to
access information

Analytical processing and delivery
Virtualisation&
QueryFederation
Enterprise
Performance
Management
Pre-built &
Ad-hoc BI Assets
Information
Services
Advanced
Analytical
Tools
Data Ingestion
Information Access
Raw Data Reservoir
Structures and
processing required
to load data (batch
and Real-Time)
and manage
Data Quality
Structures required
to interpret the data
under management.
i.e. logical interpretation
• Data Virtualisation and the various components to access the data are as per our previous view on Bo tools.
• Data Virtualisation is a key components that helps to deliver tools independence, services integration and a future state roadmap
• What has changed is the focused on Analytics
• Analytical capabilities is delivered through analytical processing in the data layers and Advanced Analytical Tools used to drive capabilities
• Data Mining in particular often involves complex data processing to flatten data into a longitudinal form. This derived data and model results
are typically written to a project based sandbox.
• Agile discovery is often best served through a separate Discovery Lab infrastructure (described later)
OLAP
Data Mining
Statistics
OLAP
Text Mining
Other
Analytical
Processing
Data Mining
Text Mining

Эволюция Big Data и Information Management. Reference Architecture.

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Эволюция Big Data и Information Management. Reference Architecture. (20)

More from Andrey Akulov (20)

Recently uploaded (20)

Эволюция Big Data и Information Management. Reference Architecture.

Editor's Notes