Organizational Semantic Model
I am delivering a data governance assessment for an enterprise client. As a part of the effort to migrate reporting from MicroStrategy to Power BI, the client wants to improve data analytics. The gap analysis interviews with the business leaders revealed common pitfalls: no single version of truth, data is hard to come by, business users don’t know what data sources exist, business users spend more time in data wrangling than analytics, data quality is bad, IT is overwhelmed with report requests, report proliferation and duplication, and so on…
Sounds familiar? As I mentioned many times in my blog, an enterprise data warehouse (EDW) plays a critical role in overcoming the above challenges, but it’s not enough. A semantic model is needed and I extolled its virtues in my “Why Semantic Layer?” newsletter. In the Microsoft BI world, Analysis Services Tabular is commonly used to implement such models that are typically layered on top of EDW . In general, there are two ways to approach the model implementation:
- (Self-service BI path) Business users create self-service semantic models using Power BI Desktop. Behind the scenes, Power BI creates databases hosted in the Analysis Services Tabular server from the *.pbix files.
- (Organizational BI path) BI developers implement organizational semantic models.
Since both implementation paths lead to the same technology, it boils down to ownership, vision, and purpose.
Because IT is overwhelmed, the temptation is to transfer the semantic model development to business users. The issue with this approach is business users seldom have the skills, time, and vision to do so. And the end, mini “semantic models” (“spreadmarts”) are produced and the same problems are perpetuated.
In most cases, my recommendation is for IT to own the semantic model because they own the data warehouse and the “Discipline at the Core” vision. And yes, unless operational and security requirements dictate otherwise, it should strive for a single centralized semantic model that spans all subject areas. If the technology you use for semantic modeling can’t deliver acceptable performance with large models, then it’s time to change it.
Of course, not all data exists or will exist in EDW. This is where self-service “Flexibility at the Edge” comes in. I have high expectations for the forthcoming “Composite models over Power BI datasets and Azure Analysis Services” mega feature (public preview expected in November 2020). This will enable the following scenario that Power BI cannot deliver today:
- Business user starts by connecting live to corporate data in the organizational semantic model. Every new report requirement should start with evaluating if all or some of the data is in the semantic model, and if so, instructing the user to connect to the semantic model to avoid data modeling and data duplication.
- Business user wants to mash up this data with some data that is not in EDW by retaining the live connection to the organizational semantic model and importing (or connective live with DirectQuery) to other datasets.
This will give us the best of both worlds and help us achieve “Discipline at the Core, Flexibility at the Edge“.
