Marketing Pipeline Intelligence: A Dimensional Model
2 likes3,041 views
The document outlines the business intelligence needs of a sports product manufacturing firm's marketing team to analyze data on customer interactions and revenue generation through various marketing touch points. It details the creation of a dimensional database schema, focusing on ad-hoc queries to understand consumer progression through a marketing pipeline, alongside proposed architectures and the use of accumulating snapshot fact tables. The goal is to facilitate in-depth analysis of marketing activities and consumer behaviors to optimize marketing efforts and increase sales.
Marketing Pipeline Intelligence: A Dimensional Model
- 1. DecisionLab.Net business intelligence is business performance October, 2010 ___________________________________________________________________________________________________________________ ________________________________________________________________ Marketing Pipeline Intelligence: A Dimensional Model by Daniel Upton ____________________________________________________________________________________________________________________________________________________________________________________ DecisionLab http://www.decisionlab.net dupton@decisionlab.net Direct: 760 525 3268 http://blog.decisionlab.net Carlsbad, California, USA
- 2. Marketing Pipeline Intelligence: A Dimensional Database Schema Daniel Upton Principal / Business Intelligence Developer www.DecisionLab.Net blog.DecisionLab.Net LinkedIn.com/in/DanielUpton __________________________________________________________________________________________________________________________________________________________________________________ Page 2 of 27
- 3. Business Requirement Summary: The marketing organization within a sports product manufacturing firm needs to acquire more insight from their data about their interactions with prospective and existing customers. Specifically, they want to know the exact relationships -- in terms of head counts, time lags and, of course, transaction-type details (ultimately including sales revenue) – among each of their many consumer “touch points”. For any given customer, these touch points include at least one of the following activities – ideally, but in truth rarely, occurring in the following idealized sequence: (1) Receipt of a prospective customer’s (herein “Lead’s”) email address (and little or nothing else) (2) Receipt of more identifying info, such that a lead may now also be classified as a known “Consumer” (3) Participation by a Consumer in a Promotion (eg. sweepstakes) (4) Initiation by a Consumer of one or more of the many available print or emailed newsletter subscriptions (5) Distribution Channel: Initial web-based (e-commerce) purchases by (what, by definition, is now) a Customer based on website- purchase referral from a variety of E-Commerce Sales Channels. (6) Distribution Channel: Initial web-based Registration of Product purchases (whether purchased via brick-n-mortar reseller or direct e- commerce) (7) Repeat customer interactions on any of the above touchpoints, obviously including repeat purchases and registrations. Although the business requirement alludes to “..all relationships of consumer counts, time lags and revenue…”, the following examples provide some idea of the wide spectrum of requirements. Required Ad-hoc Query Types: (A) How many unique (Count) leads, consumers, promotion participants, newsletter subscribers, web-purchasers and reseller- purchasers do we have, and how does this Count vary by lead source, sweepstake participation, consumer geography, newsletter subscription activities, (multiple) product categorization hierarchy, sales channel and, of course, time? (B) For the same above criteria, how many (Count) of them progress from each lower-value activity (eg. new lead email) to a higher value group (eg. large, recent repeat-purchases)? (C) What can we learn about who purchased what, where and when? (D) Which of our Website-purchasing- vs. Reseller-purchasing Customers, from which Geographies have purchased (or registered) which of our Products, according to a variety of Product Categorization hierarchies (product line, brand, website categorization, manufacturing categorization), over what time periods and for what amount of actual or (MSRP for reseller-customers) estimated revenue? (E) What product-mix relationships exist between and among website-purchasing customer vs. reseller-purchasing customers? Note: Website products categorizations are currently handled inconsistently between these two distribution channels. __________________________________________________________________________________________________________________________________________________________________________________ Page 3 of 27
- 4. (F) What are the actual sequences, as well as time-spans in days (herein “Time Lags”) between Consumers’ progressions in and out of (ie. newsletter ‘unsubscribes’) the aforementioned activities, including progression into becoming purchasing Customers, and subsequent to initial purchase? (G) Among all touch points, what are the sequences / paths that have led to customer purchases that are (select any number of the following) more frequent, consistent, long-standing, seasonally-dependent / independent or, of course, largest? (H) What are the counts, time lags and interaction (transaction) details with which our existing -- and our known prospective -- customers have actually followed any portion of our idealized “low-value to high-value” marketing pipeline sequence? (I) Conversely, what do we know about customers who leveraged few, or none, of our available non-purchase-related touch points before, or after, their actual product purchases or product registrations? (J) What has been the subsequent revenue from customers before or after interaction in any form within any of the touch points? (K) Catch-All: Except where the granularity of the business data logically prevents it, all quantitative facts (measures) must be able to be broken out by Distribution Channel, Product, E-Commerce Sales Channel, Consumers demographics, skill-level, sport-activity-level, known attitudes, geography and timing. Scary! (L) An odd request: What consumer demographics can we find (other than “geography” (the obvious one) that most dramatically differentiates our Sweepstake Participants, both according to individual Sweepstake and Country from which a specific Sweepstake Participation occurred? (M) Determine which of our newsletters (subscriptions) or sweepstake events coincide with best and worst revenue performance and with what time relationships (ie. between subscription (or sweepstake) and purchases.) (N) Determine the ECommerceSalesChannels (ie. referring e-commerce websites) that coincide with most and fewest newsletter subscriptions, and with what time relationships between subscriptions and purchases. (O) Determine which exact Products (purchased or registered) are associated with the most, and fewest, newsletter subscriptions, and with what time relationships. … (Z) Overall: In essence, the business side is excited about ad-hoc multi-dimensional analyses, is poised with an impressive OLAP server, and will be thrilled to see, in the schema we have in mind, virtually every fact table slice-able by virtually every dimension table, except insofar as it would actually violate a known business rule. Rather than advising them otherwise, we will do our utmost to deliver all of it in a single SSAS cube. The Classic ‘Accumulating Snapshot’ Fact Table (herein ‘AccumSnap’): AccumSnap fact tables are completely different from either Periodic Snapshot Facts or Transaction facts. Unlike the other two archtypes, AccumSnap’s allow us to aggregate counts and time lags between related yet distinct processes, which may occur in unpredictable sequences. Sets of processes that combine into a pipeline-like scenario lend themselves to the AccumSnap. One classic example is the college admissions pipeline, wherein many related processes occur along the way between a student making initial contact with a college and a student arriving for class on day one. Another classic AccumSnap is a customer pipeline, wherein many touchpoints (processes) occur between an organization and a prospective and/or existing customer. If each of the processes are __________________________________________________________________________________________________________________________________________________________________________________ Page 4 of 27
- 5. simple enough to be incorporated into a dimension table, then the classic AccumSnap will suffice. To read Ralph Kimball’s Design Tip #37, an authoritative description of the archtype (single fact table, and just the essentials), click the following link: http://www.ralphkimball.com/html/designtipsPDF/DesignTips2002/KimballDT37ModelingPipeline.pdf “Schema Hub”: Extending the Accumulating Snapshot-centric Star Schema with Transaction Facts When some of the processes along a pipeline are complex enough, or iterate as transactions, we would like to accurately model them as transaction-grained fact tables. In our schema, we will indeed use an AccumSnap, but add multiple Transaction fact tables that are within what I’ll call the “Schema Hub”, directly related to the AccumSnap fact tables (vs. the schema periphery), forcing the AccumSnap (being higher/coarser grained than the Transaction facts) to now serve double-duty as a join table between each of these other fact tables themselves, as well as a join to a few dimensions. Conversely, the Transaction Fact Tables, being finer-grained than the AccumSnap, will serve double-duty as M2M join tables between AccumSnap and dimensions directly related to the transactions. **No need to visualize these specific relationships yet. Screenshots and detailed discussion on both of those point will follow. The advantage of relating most of these fact tables together so directly at the schema’s hub, is that their position does not inherently limit their relationship to any other table, be it a fact or dimension table. This would not be the case if fact tables were more isolated and thereby related directly to only a few dimension tables. After all, in order to slice and dice facts by dimension values, the table relationships must exist. Moreoever, this hub approach is my method to maximize the range of available ad-hoc queries from a single cube spanning multiple, closely related processes. Expert feedback is welcome on this point. Options for Downstream Consumption of Schema To the extent that the schema on the following page pulls from a medium or large dataset with, say, ten million to one billion fact rows in any of the core fact tables, I consider that an OLAP Cube, such as SQL Server Analysis Services (SSAS), is probably required, as an middle aggregation layer, which is then consumed by either a dashboard or set of reports as the front-end, rather than having the front- end pull directly from the schema. Since, as readers will see, all of the core fact tables will serve double duty as either ‘Intermediate’ dimensions, or as many-to-many (herein M2M) join tables, query performance from the schema would be very slow without OLAP, perhaps even too slow for single-user (non-simultaneous) queries. For smaller datasets, the architect will have to decide on whether OLAP benefits are worth the time and costs. Going forward here, our assumption is that a single, sophisticated SSAS OLAP cube will be built. __________________________________________________________________________________________________________________________________________________________________________________ Page 5 of 27
- 6. Reference 1: MarketingPipelineIntelligence Schema Here is a view (table names only) of the entire MarketingPipelineIntelligence schema. I suggest that readers print out this page for reference during subsequent reading even if other pages are not printed. Reference 3, near the end of the document, shows all fields, but in very small font! __________________________________________________________________________________________________________________________________________________________________________________ Page 6 of 27
- 7. Going forward, I will sequentially display and discuss sub-groups of the above tables, as topic sets that merit discussion, while also occasionally adding in other tables that add business value but merit little description. Along the way, some tables will appear multiple screenshots, since they participate in multiple relationships. Before delving into specific tables, let’s review the schema in general. To begin with, it is fundamentally based on the Kimball-style Star Schema (not 3NF / Inmon style / CIF) insofar as… (a) Fact tables, with quantitative measures, are largely distinct from dimension tables, with qualitative attributes, as subsequent table details will show. Importantly, each measure/fact shares a common granularity with others in the same fact table, although some may aggregate differently (Sum, Avg, Max, Last, etc.) Note 3 below, however, describes one departure from this classic approach, wherein some fact tables serve double-duty. (b) In general, facts rows relate ‘many-to-one’ to dimension rows. It is never the reverse of this and, on occasion, may be one-to-one for join purposes. (c) Dimension tables, even those with multiple independent hierarchies, are generally de-normalized unless very large and/or very sparse. (d) Role-playing dimensions are used extensively (DimDate and DimProduct, in this case) However, the schema does have its unique features, not typically seen in star schemas. The following notes apply… Note 1: For non-SSAS readers, the terms ‘Fact Table’ and ‘Measure Group’ are used to describe essentially the same thing, and the terms ‘Fact’ and ‘Measure’ are also equivalents. Note 2: In SSAS, these atypical relationships will be handled with combinations of dimension relationships that are either ‘Many-to- Many’ (M2M), ‘Cascading M2M’, ‘Referenced’ …or Combined M2M-and-Referenced Relationships! Details will follow. Note 3: Data Modeling Style: When it’s logical, I like to relate multiple fact tables DIRECTLY together with few or no other dimension tables in between in order to eliminate any artificial limits on fact-dimension relationships and thus on available queries. Goal: Except insofar as the actual business logic prohibits, I try to allow all or most fact tables to be slice-able by all or most dimensions. Depending on the relative granularity between fact tables, this means that, in SSAS, some fact tables (just using their PK and FK’s) serve double-duty as Intermediate Dimensions (joins) in Referenced cube-dim relationships. This is also the reason that my fact tables usually contain single-field, surrogate PK’s, instead of composite PK’s using multiple FK’s. It’s frequently performed well for me in the my pursuit of “…fewer, faster, more comprehensive” cubes. Here, and throughout this document, I am actively seeking expert feedback. _______________________________________________________________________________ __________________________________________________________________________________________________________________________________________________________________________________ Page 7 of 27
- 8. Reference 2: SSAS “Cube Dimension Usage” Grid As with the above screenshot, the next two (which go together), serve as a recurring reference and will be convenient if printed out. A brief discussion follows. __________________________________________________________________________________________________________________________________________________________________________________ Page 8 of 27
- 9. A continuation of the above screenshot… The above two-part grid describes the proposed configuration for SSAS Cube Dimension Usage. Much of this documents subsequent discussion involves describing each of these fact-dimension relationship. Wherever possible, the above color codes will be used for quick referencing. __________________________________________________________________________________________________________________________________________________________________________________ Page 9 of 27
- 10. Set 1: FactPipelineAccumSnap, the schema’s core fact table. It is shown below in two side-by-side screenshots so all fields are readable) Let’s simultaneously review the distinguishing features of Kimball’s Accumulating Snapshot (herein ‘…AccumSnap’) fact-table archtype, and the features of the above table that are an expansion on, or adaptation of, the arch-type. Specifically… 1. AccumSnap’s consist exclusively of five types of fields: a. Primary key (PK) field -- obviously not nullible. In other cases, PK is simply a composite on selected FK fields. b. Date Foreign keys fields. __________________________________________________________________________________________________________________________________________________________________________________ Page 10 of 27
- 11. i. Classically, for all dimensions (directly) related to an AccumSnap fact table, a Date Foreign Key is included to track the time of the occurrence of that process. Moreover, each dimension touching the AccumSnap usually has an associated Date dimension, and the role-playing Date dimension is well suited here. ii. In our case, a Date Foreign Key field exists here whenever a business process is captured simply in a dimension and lacks an associated ‘…Trans’ fact table (which we’ll explore soon). Since DimLead and DimConsumer fit this description, Date Foreign Keys exist to track the timing of those processes. Whenever an associated ‘…Trans’ fact table exists, the Date Foreign Keys in not in ‘…AccumSnap’, but rather in the ‘…Trans’ fact table itself, in order to track the process with adequate granularity to cover multiple iterations of a given process (eg. multiple subscriptions) for a given consumer. c. Non-date Foreign key (FK) fields. Standard stuff. Not nullible. d. Count Fields: Here, the allowed values are { 0,1 }. Not nullible. i. Arch-type AccumSnap: Count each process by itself, not in relation to (lead/consumer’s) progression to another process. ii. This AccumSnap: Also counts progress of leads/consumers from lower-value processes (eg. receipt of lead email) towards higher-value process (eg. product registration). All fields with both of words ‘Into’ AND ‘Count’ are of this kind. e. Arch-type: Time-lag fields, which measure elapsed time (days, in our case) between progression of leads/consumers from one process to another. This field is nullible, so we can distinguish between ‘Null’ lag days -- meaning that a consumer has not made the specific progression – and ‘0’ lag days, meaning that a consumer’s specific progression occurred in less than 1 day. It is also worth mentioning that special attenting must be paid to ensure that these “…LagDays” field values coincide exactly with slicing these facts by date dimension attributes. Not a trivial matter for ETL and QA. Please take a moment now to note the potential business value of each of the above-listed fields. From the end-users’ perspective, the Accumulating Snapshot fact table is a sensible way to capture information about processes which occur in pipeline-like environments that are either predictable (eg. the required processes which college hopefuls must progress toward to become enrollees) as well as unpredictable (such as ours, wherein the conversion of a portion of leads into known consumers, then sweepstake participants, subscribers and hopefully, paying customers does occur, but with new participants entering the pipeline in various places, and in unpredictable sequences, such that we get some new customers whom we had never heard of prior to purchase). Having said that, if our real-world (allegedly pipeline-like) environment actually includes processes with iterative, quantitative details, as ours does, the Accumulating Snapshot fact table arch-type, by itself, lacks the fine, transaction-granularity to store those details. To accommodate this requirement, we therefore will add a collection of ‘…Trans’ fact tables. Once we begin describing the ‘…Trans’ fact tables, which do include some fields that could be used to derive values for Count and LagDays fields already shown in ‘…AccumSnap’ table, some readers who build cube and/or relational reports will quickly note that they could eliminate the need for those Count and LagDays fields in ‘…AccumSnap’ writing expressions from ‘…Trans’ fields to calculate them on the fly. While this is true, these fields, I believe, are best calculated during ETL and stored in the star schema itself, because the expressions tend to be rather complex and will thereby hinder query-response performance. As you consider this, take a moment once again and consider the complexity of deriving a few of the ‘…AccumSnap’ tables more complex Count and LagDays fields. Do we want that complexity completed during off-hours ETL and cube processing time, or during end-user sessions? If we go with the ‘…AccumSnap’ as is, we can consider it to be a specialized Aggregate Table, which admittedly makes it a rarity in the SSAS ‘05/’08 OLAP space. On this point, expert feedback is requested. In subsequent pages, we will cover how to implement these atypical relationships for use in business intelligence (especially MSAS cubes). However, before diving deeply into that, let’s describe each of the schema’s tables and, first of all, their more routine relationships. __________________________________________________________________________________________________________________________________________________________________________________ Page 11 of 27
- 12. _______________________________________________________________________________ Set 2: DimLead,DimConsumer. Simple. Leads are email addresses, sometimes also containing additional information as the above table shows. Consumers are people for whom we’ve gathered sufficient information to positively identify a person. In our case, we choose to require a “login” (primary) email address, last name, firstname, gender and birthyear, with all other fields being desired but not required. As a note, other processes are designed, in part, to complete these additional DimConsumer fields. Please take another moment to note each of these fields and consider their potential business value. _______________________________________________________________________________ __________________________________________________________________________________________________________________________________________________________________________________ Page 12 of 27
- 13. Set 3: FactSubscriptionActionTrans and DimSubscription Notes: (A) FactSubscriptionActionTrans historically tracks consumers’ subscription, unsubscription, and re-subscription to each of our variety of newletters. One row equates to one consumer’s action with regard to one subscription. Non-key fields include: a. SubscriptionActionDescription allowed values are {Initial subscribe, Unsubscribe, Repeat Subscribe} b. SubscriptionActionCount allowed values are always {1}, and thus serve only as a filterable row-count. (B) DimSubscription: Here, one row equates to one newletter, whether print- or emailed-format. It contains a categorization field, but is otherwise simple. (C) Lastly, DimSubscription must relate to many other tables, too; which we will fully address in a subsequent section. _______________________________________________________________________________ __________________________________________________________________________________________________________________________________________________________________________________ Page 13 of 27
- 14. Screenshot for… Set 4: FactECommerceCartItemTrans, FactProductRegistrationTrans and DimProduct, and… Set 5: DimECommerceSalesChannel and FactECommerceCartItemTrans __________________________________________________________________________________________________________________________________________________________________________________ Page 14 of 27
- 15. Set 4 Notes: (A) FactECommerceCartItemTrans a. Transaction-grained fact table. One row equates to a single e-commerce cart line item (which may include a quantity > 1). b. ‘…Spend’ field is for all ordered quantities of that (line-item) product (or product set). Unit price is not stored, but derived downstream (with MDX or SQL) LeadConsumerSurrogateID_FK has an enforced many-to-one relationship with ‘…AccumSnap’.LeadConsumerSurrogateID_PK c. All fields beginning with ‘Is…’ are degenerate (fact) dimensions, with Boolean (0=No, 1=Yes). (B) FactProductRegistrationTrans (fields,etc.) a. Transaction-grained fact table. One row equates to a single web-based product registration line item (which may include quantities > 1). Thus one customer’s product-registration session involving multiple products will create multiple line items here. b. Same principle as above applies to obtain unit price (this time not as actual, but instead as MSRP) c. Same many-to-one relation to “…AccumSnap’ as above. d. As above, fields beginning with ‘Is…’ are degenerate (fact) dimensions, with Boolean (0=No, 1=Yes). (C) DimProduct a. Dimension Key: WebupdateProductIDSurrogate_PK i. One row equates to one product. ii. Type 2 slowly-changing dimension (SCD 2), with StartDate, EndDate to capture Product change history iii. Common to all dimension hierarchies and fields. iv. Surrogate PK is the integration key between the e-commerce and web-based product registration source systems, since they systems, se disparate product keys and hierarchies. b. IsInferred field supports minimal dimension entry for (early arriving) facts mistakenly arriving before corresponding dimension updates, which will then populate the other, temporarily null, fields in the row. c. Multiple independent hierarchies (‘…ProductLevel…’, ‘…ProductLevelAlternate…’) in a single, denormalized dimension table. d. A Role-Playing Dimension will be used (either in cube dimensions, or as relational views for relational reporting), for the following two reasons: i. This Product dimension is a conformed (standardized master) Product table, and serves as the analytic product-related integration point between the two otherwise disparate sales channels. ii. Noting that the two fact tables represent different processes, queries that drill into or filter one fact table by product must not be forced to filter the other one identically, even if fields from both fact tables appear in displayed output. (D) As with DimSubscription, the DimProduct table must relate to many other tables, too; which we will fully address in a subsequent section. Set 5 Notes: (A) FactWebCartItemTrans: Already described on previous page. (B) DimWebSalesChannel is a hierarchized, denormalized dimension describing the various websites referring online purchasing customers directly to firm’s e-commerce cart. _______________________________________________________________________________ Set 6: FactSweepstakeParticipationTrans, DimSweepstake, FactM2MBridgeSweepstakeCountry, DimCountry __________________________________________________________________________________________________________________________________________________________________________________ Page 15 of 27
- 16. *** Preliminary M2M Note: The schema’s first simple (non-cascading) M2M relationship is described here. (A) FactSweepstakeParticipationTrans: Transaction-grained fact table a. One row equates to one customer participating in one sweepstake b. The only allowed value in ‘…Count’ field is {1}, and it serves as a filterable row-count. (B) Dimension tables with (non-cascading) M2M Relationship: Three left-most tables above. a. Why M2M? The Business requirement for the M2M relationship between DimCountry and DimSweepstake is that (1) sweepstakes can span multiple countries, (b) multiple sweepstakes can be available to participants in a single country, and most importan tly, some countries prohibit sweepstake participation, and we can follow that with the DimCountry.AllowSweepstakes field. DimCountry could be converted to a Type 2 SCD if we needed to track those changes historically. b. For a quick review of how-to, using MS Analysis Services (2005 or later), implement the simple (non-cascading) M2M relationship between DimCountry attributes and FactSweepstakeParticipationTrans measures, interested readers can do the following. Others (SSAS Seniors) should skip forward. i. Create dimensions for DimCountry and DimSweepstake ii. Build a cube named Marketing Pipeline Intelligence, adding both dimensions and measure groups (fact tables) as shown above. 1. Note: The only dimension-fact relationship type that will not automatically be correctly established during cube construction is DimCountry-to-FactSweepstakeParticipationTrans. iii. To relate DimCountry to FactSweepstakeParticipationTrans: (1) In BI Dev Studio (herein ‘BIDS’) open Cube from Solution Explorer; (2) go to Dimension Usage tab; (3) locate the grid-intersection position of DimCountry Dimension and FactSweepstakeParticipationTrans Measure Group; (4) click elipse button; (5) in ‘Select relationship type’ choose ‘Many-to- __________________________________________________________________________________________________________________________________________________________________________________ Page 16 of 27
- 17. Many’ instead of the more common “Direct”; (6) Dimension: select ‘DimCountry; (7) Intermediate measure group: select ‘FactM2MBridgeSweepstakeCountry’; (8) Click ‘OK’. c. The challenge of the additional required M2M relationships in this schema, including “Cascading Many-To-Many” relationships involving the above four tables, as well as ‘M2M plus Referenced Relationships’ will be described together in a subsequent section. _______________________________________________________________________________ __________________________________________________________________________________________________________________________________________________________________________________ Page 17 of 27
- 18. Set 7: Relationships of Core Fact Tables And Two Selected Dimensions: FactMarketingPipelineAccumSnap, FactSweepstakeParticipationTrans, FactSubscriptionActionTrans, FactEEcommerceCartItemTrans, FactProductRegistrationTrans, DimLeadEmail, and DimConsumer __________________________________________________________________________________________________________________________________________________________________________________ Page 18 of 27
- 19. Note: (A) ‘…AccumSnap’ table serves double duty here, not only as a fact table per se, but also as a join / intermediate dimension between the four ‘…Trans’ facts (on far left above) and two outrigger / Referenced dimensions (on the far right in the above screenshot): a. In order for the ‘DimLeadEmail’ and DimConsumer’ dimensions to relate to each of the ‘…Trans’ measure groups, the ‘…AccumSnap’ measure group (with its granularity being coarser (higher) than the ‘…Trans’ measure groups, yet finer than the two dimensions, we will also need to use the ‘…Accumsnap’s one PK field and both of it’s (non-date-related) ‘…_FK’ fields to form a join / intermediate dimension. _______________________________________________________________________________ Set 8: Core M2M: Most of the Core Many-to-Many (M2M) Relationships __________________________________________________________________________________________________________________________________________________________________________________ Page 19 of 27
- 20. Slicing the Above ‘…AccumSnap’ Facts with The Three Left-Most Dimensions Above: As the Business Requirement Summary dictates, measures in the ‘…AccumSnap’ fact table (uppermost on this page) must be sliced by each of the three dimensions shown here (left-most). Since none of them relate directly to ‘…AccumSnap’, and since each of the in-between ‘…Trans’ fact tables has a finer granularity than either the ‘…AccumSnap’ or the corresponding dimension, the relationship here is M2M, with the respective ‘…Trans’ fact tables serving as M2M bridges. In SSAS, these are referred to within the M2M relationship as ‘Intermediate Measure Groups’. As a parting note on this Set, I acknowledge that it is atypical of an M2M relationship insofar as no dimension table exists between ‘…AccumSnap’ and the ‘..Trans’ fact table. However, the essential M2M relationship is the same, and testing demonstrates that it works correctly. This paradigm applies identically to the next set as well. Feedback on this from expert reviewers is certainly appreciated! Slicing of Each of The Above Three Sets of ‘…Trans’ Facts with Each of The Above Three Left-Most Dimensions Above: This is more challenging since, for two of the three above dimensions, their relationship with two of the fact tables is far from direct. Let’s break it down. Slicing FactProductRegistrationTrans by DimSubscription: Why?: Business Requirement Item (M): Determine which of our newsletters (subscriptions) or sweepstakes coincide with best and worst revenue performance and with what time relationships (ie. between subscription (or sweepstake) and purchases.) How to (in MSAS)? The Fact-Dimension relationship here is M2M, with FactSubscriptionActionTrans (it’s adjacent fact table) serving as the Intermediate Measure Group. Slicing FactECommerceCaretItemTrans by DimSubscription: Why?: Same as above. How (in MSAS)?: The Fact-Dimension relationship here identical as above, except for the ‘destination’ fact table. Slicing FactSubscriptionActionTrans by DimECommerceSalesChannel: Why?: Business Requirement Item (N): Determine the ECommerceSalesChannels (ie. referring e-commerce websites) that coincide with the most, and fewest, newsletter subscriptions (and what time relationships). Relationships here only differ from the above ones by, obviously, different ‘destination’ and ‘Intermediate’ (adjacent to dimension) measure groups. Slicing FactSubscriptionActionTrans by DimProduct (in both of it’s two dimension roles): Why?: Business Requirement Item (O): Determine which exact Products (both online-purchased or registered) that coincide with most and fewest subscriptions to specific newsletter (and time relationships). How to (in MSAS)?: Relationships here only differ from the above ones by, obviously, different ‘destination’ and ‘Intermediate’ (adjacent to dimension) measure groups. Notably, this process will be duplicated, since DimProducts will play two roles in our cube, with each role using it’s own adjacent ‘Intermediate Measure Group’. _______________________________________________________________________________ __________________________________________________________________________________________________________________________________________________________________________________ Page 20 of 27
- 21. Set 9: More Complex M2M Relationships. Discussed below in two sub-sets First SubSet (Unique M2M): Recall that the business also requires that ‘…AccumSnap’ measures also be sliced by both Sweepstake and (Sweepstake) Country, so now we not only have another simple (one-bridge) M2M relationship but, in fact, a Cascading M2M (more than one M2M bridge-joins in a series), to get from DimCountry to ‘…AccumSnap’. In MSAS, this complex relationship must be built AFTER the non-cascading M2M for DimSweepstake to ‘…AccumSnap’, so it can be built on top of it. Once that is done, cube designers are reminded that, for the aforementioned Cascading M2M relationship, the Intermediate Measure Group should be FactM2MBridgeSweepstakeCountry (adjacent to DimCountry), not ‘…Trans’ (adjacent to ‘…AccumSnap’). Thank you, Marco Russo! (For insight into M2M design fundamentals, see http://www.sqlbi.eu ,then /Projects/Many-to-Many Dimensional Modeling). Also unique here is the fact that this particular relationship is atypical even with the Cascading M2M category. Specifically, this setup does not one M2M bridge-joins, not two, but rather one and one- half (count them). As you’ll see in the next screenshot of prototype cube-browse results, it does indeed provide accurate end results. Expert feedback is requested on this. Second SubSet (Still More Unique M2M): ‘Combined M2M…Referenced Relationships’ Remember that the business made the odd request I listed as Item ‘L’ under ‘Ad-Hoc Query Types’ in this paper’s introduction, which is to allow for answers to the question: “What lead demographics can we find that most clearly differentiates our Sweepstake Participants, according to individual Sweepstake and by Country from which a specific Sweepstake Participation occurred?” To answer this question, we must now relate DimLeadEmail to ‘FactM2MBridge…’ (specifically, the ‘FactM2Mbridge…’measure is ‘SweepstakeCountryUniqueCombinationsCount’. To accomplish this, we will establish a still-more complex fact- dimension relationship. This time, of course, the fact table is ‘FactM2MBridge…’. This is a truly unique M2M relationship. Specifically, it is a (non-cascading) Combined M2M… Relationship (from ‘FactM2MBridge… ‘ to ‘FactSweep…Trans’, which will be built using the __________________________________________________________________________________________________________________________________________________________________________________ Page 21 of 27
- 22. ‘Intermediate Dimension’ from an existing …Referenced Relationship (from DimLeadEmail (a referenced dim) to ‘FactSweep…Trans’ with (in SSAS OLAP) ‘InterDim_Fact …AccumSnap’ as an intermediate dimension). The following other dimensions can use an identical relationship setup to relate to “FactM2MBridge…”: DimConsumer, (SSAS Role) DimDateLeadNewReceived, and (SSAS Role) DimDateConsumerNewInfoReceived). Since you, the reader, obviously cannot test the results in the above screenshot yourself against my un-published source data, you’ll have to trust me for now that the result is correct. The following notes on source data rows may help: (1) The ‘CanadaSpecial’ Sweepstake was available only in Canada; (2) the ‘GlobalDrive’ Sweepstake was available both in Canada and the USA. (3) ‘Grace…’ participated in both; (4) ‘Tom…” participated only in ‘GlobalDrive’. As always, browsing valid M2M results causes unusual, non-additive displayed results, especially depending on placement of dimensions. However, I have verified that they are indeed correct, which demonstrates that even the unusual “Combined M2M…Referenced” fact-dimension relationships in this schema, such as this one between ‘Lead Email’ and ‘LeadsIntoInitialSweepPartipLagDays AvgMDX’, can produce accurate results. Specifically, ‘Lead Email’ is able to accurately slice ‘Fact M2M Bridge Sweepstake Country Count’, even though the relationship between the two is this “Combined M2M…Referenced” type. __________________________________________________________________________________________________________________________________________________________________________________ Page 22 of 27
- 23. The screenshot immediately above, from the same prototype SSAS cube’s ‘Dimension Usage’ tab, illustrates some points to discuss (some on which readers will have to trust my test results) 1. Prior to building the referenced relationship between ‘FactSweep… Trans (using ‘InterDim_Fact..AccumSnap’ as ‘Intermediate Dimension’) to DimLeadEmail’ (hint: it turns out to be needed first), I setup an M2M relationship between “FactM2MBridge…” and ‘DimLeadEmail’. Browsing demonstrated that it produced erroneous results on ‘Tom…’. At that stage, no other ‘Intermediate’ measure group was available. 2. Then, after building the referenced relationship just mentioned above, the ‘FactSweep…Trans’ becomes available as an ‘Interme diate Measure Group’ to our M2M Relationship, and so it was used, with browsing demonstrating the correct values shown in the browser screenshot just before the above screenshot. So, it seems that we have demonstrated correct values from a Combined M2M …Referenced Relationship, which is great because it tends to validate the overall approach of placing to many fact tables so closely related to each other. *** Expert feedback is very much desired on this point. Anyone seen this methodology before? *** Slicing of Each of The Three Sets of ‘…Trans’ Facts (displayed not above, but in previous screenshot) by DimSweepstake AND DimCountry: Here is another challenging set of non-direct table relationships. As before, let’s break these down. Slicing FactSubscriptionActionTrans by DimSweepstake: Why?: Business Requirement Item (M): Determine which of our newsletters (subscriptions) or sweepstake events coincide with best and worst revenue performance and with what time relationships (ie. between subscription (or sweepstake) and purchases.) How to (in MSAS)? For each of the three aforementioned ‘…Trans’ fact tables, the Fact-Dimension relationship here is identical to many aforementioned M2M relationships, simply with differing ‘Intermediate’ (adjacent to dimension) and ‘Destination’ Measure Groups. Slicing FactSubscriptionActionTrans by DimCountry: Why?: Same requirement as above How to (in MSAS)? For each of the three aforementioned ‘…Trans’ fact tables, the Fact-Dimension relationship here is our first set of “two hop” cascading M2M relationships. These can only be built after each of the respective ‘Intermediate’ (single-step) M2M relationships are built. In each of these cases, the ‘Intermediate’ Measure Group’ is always ‘FactM2MBridgeSweepstakeCountry’. Expert feedback requested here, especially with regard to experience with query and/or cube processing performance. Who can provide feedback from experience with Cascading M2M scalability? _______________________________________________________________________________ Set 10: DimDate (A) Role-Playing Dimensions Concept: The role-playing dimension concept is well-known and not really a design challenge, per se, for this schema, so I left it for last in this discussion. Having said that, it is used extensively and requires many of the complex fact-dimension relationships that are identical to aforementioned ones. Thus, no additional discussion on them is provided here. For individual role names and relationships, see the embedded “Cube Dimension Usage” table in this document. Lastly on this point, within each role, I will usually append the role name to each field, such as ‘Week_LeadNewReceived’. (B) Business Value: The six role-playing dimensions enable users to slice all fact tables by the dates associated with any of the other fact tables, which is one of ways in which this schema provides anwers to a huge array of questions. __________________________________________________________________________________________________________________________________________________________________________________ Page 23 of 27
- 24. (C) Note on MDX Time-Series Calculations Utility Dimensions: The six role-playing date dimensions, once in the cube, can each support an MDX Time Utility Dimension. My approach here is, generally, to make all six of them identical in terms of calculated values. Although beyond the scope of this paper, at last two web resources can help those wanting to learn more about the technique. a. ‘Date Tool’, by Marco Russo: URL is… http://www.sqlbi.eu/Projects/DateTool/tabid/87/Default.aspx b. ‘A Different Approach to Implementing Time Calculations in SSAS”, by David Shroyer -- OLAP Solutions: URL is… http://www.obs3.com. Under ‘Papers’, see ‘Time Calculations’. (D) Here’s the screenshot. Again, to see the role names, please refer to the embedded “Cube Dimension Usage” table. __________________________________________________________________________________________________________________________________________________________________________________ Page 24 of 27
- 25. Reference 3: Big Picture: Marketing Pipeline Intelligence Schema (All Tables and Fields) __________________________________________________________________________________________________________________________________________________________________________________ Page 25 of 27
- 26. Conclusion: Recall the Catch-All Item (K) in the introduction’s “Required Ad-hoc Query Types” section, which stated… Except where the granularity of the business data logically prevents it, all quantitative facts (measures) must be able to be broken out by Distribution Channel, Product, E-Commerce Sales Channel, Consumers demographics, skill-level, sport-activity- level, known attitudes, geography and timing. It seems that we have accomplished that and, as a result, can offer an enormous range of sophisticated ad-hoc querying, which was our major goal. In fact, all measures in all fact tables can be browsed against all attributes from all dimensions, which is valuable given that this schema tightly integrates many separate business processing into a single, flexible analytic interface. With the ‘Fact…AccumSnap’ as the central fact table in our Schema Hub, we also have the extensibility to support the addition of other consumer touch points, whether they are dimension-like or transaction- like, by simply adding them into the Schema Hub via the “Fact…AccumSnap” fact table, and then adding their associated “…Count” and “…LagDays” measure fields there, too. Mission accomplished. Questions, comments and expert critiques from readers are very welcome Since this paper is a draft, please provide your feedback to me in my DecisionLab Forum (vs. the more public Windows Live, MSDN forum’s etc.) at http://forum.decisionlab.net/User/Discussion.aspx?id=203097. In doing so, readers will be able to view and/or comment on, feedback from others. Following expert feedback and possible revision, I would like to publish it more widely (eg. MSDN, Windows Live, etc.), and, in fact, suggestions on places to publish it are most welcome. DecisionLab Forum access is, by default, immediately granted once username and password are set up, so it should be easy and quick. Those who experience problems with forum access or entries should email me at dupton@decisionlab.net. Lastly, I intend to blog on this and similar topics. See http://blog.decisionlab.net ________________________________________________ Daniel Upton dupton@decisionlab.net DecisionLab.Net business intelligence is business performance __________________________________________________________________________________________________________________________________________________________________________________ Page 26 of 27
- 27. DecisionLab.Net _________________________________________________________________________________________________________________________________________________________________________ Daniel Upton DecisionLab http://www.decisionlab.net dupton@decisionlab.net Direct 760.525.3268 http://blog.decisionlab.net Carlsbad, California, USA __________________________________________________________________________________________________________________________________________________________________________________ Page 27 of 27