So far in this series I have compared and contrasted the Data Vault approach to that of using a replicated ODS or a dimensional Kimball-style structure for an enterprise data warehouse repository. Let me thank everyone who has replied or commented for taking these posts as they were intended, keeping the conversation professional and not starting any flame wars. 🙂
So to close out the thread I must of course discuss using classic Inmon-style CIF (Corporate Information Factory) approach. Many of you may also think of this as a 3NF EDW approach.
To start it should be noted that I started my data warehouse career building 3NF data warehouses and applying Bill Inmon’s principles. Very shortly after I began learning about data warehousing I had the incredible privilege of getting to not only meet Mr. Inmon and learn from him but I got to co-author a book (my first) with him. That book was The Data Model Resource Book. I got to work with Bill on the data warehouse chapters and because of that experience became well versed in his theories and principles. For many years I even gave talks at user group events about how to convert an enterprise data model to an enterprise data warehouse model.
So how does this approach work? Basically you do some moderate denormalization of the source system model (where it is not already denormalized) and add a snapshot date to all the primary keys (to track changes over time). This of course is an oversimplification – there are a number of denormaliztion techniques that could be used in build a data warehouse using Bill’s original thesis.
Additionally this approach (CIF or even DW 2.0) calls for then building dimensional data marts on top of that EDW (along with other reporting structures as needed). Risks here are similar to those mentioned in the previous posts.
The primary one being that the resulting EDW data structure is usually pretty tightly coupled to the OLTP model so the risk of reworking and reloading data is very high as the OLTP structure changes over time. This of course would have impacts downstream to the dimensional models, reports, and dependent extracts.
The addition of snapshot dates to all the PKs in this style data warehouse model also adds quite a bit of complexity to the load and query logic as the dates cascade down through parent-child-child-type relationships. Getting data out ends up needing lots of nested Max(Date) sorts of sub-queries. Miss a sub-query or get it wrong and you get the wrong data. Overall a fairly fragile architecture in the long run.
Also like the dimensional approach, I have encountered few teams that have been successful trying to implement this style of data warehouse in an incremental or agile fashion. My bad luck? Maybe…
The loosely coupled Data Vault data model mitigates these risks and also allows for agile deployment.
As discussed in the previous posts, the data model for a Data Vault based data warehouse is based on business keys and processes rather than the model of any one source system. The approach was specifically developed to mitigate the risks and struggles that were evident in the traditional approaches to data warehousing, including what we all considered the Inmon approach.
As I mentioned earlier I got to interact with Bill Inmon while we worked on a book. The interaction did not stop there. I have had many discussions over the years with Bill on many topics related to data warehousing, which of course includes talking about Data Vault. Both Dan and I talked with Bill about the ideas in the Data Vault approach. I spent a number of lunches telling him about my real-world experience with the approach and how it compared to his original approach (since I had done both). There were both overlaps and differences. Initially, Bill simply agreed it sounded like a reasonable approach (which was a relief to me!).
Over a period of time, many conversations with many people, study, and research, we actually won Bill Inmon over and got his endorsement for Data Vault. In June of 2007, Bill Inmon stated for the record:
“The Data Vault is the optimal choice for modeling the EDW in the DW 2.0 framework.”
So if Bill Inmon agrees that Data Vault is a better approach for modeling an enterprise data warehouse, why would anyone keep using his old methods and not at least consider learning more about Data Vault?
Something to think about, eh?
I hope you enjoyed this little series about Data Vault and will keep it in mind as you get into your new data warehouse projects for 2013.