In the two previous posts of the Incremental Refresh in Power BI series, we have learned what incremental refresh is, how to implement it, and best practices on how to safely publish the semantic model changes to Microsoft Fabric (aka Power BI Service). This post focuses on a couple of more best practices in implementing incremental refresh on large semantic models in Power BI.

Note

Since May 2023 that Microsoft announced Microsoft Fabric for the first time, Power BI is a part of Microsoft Fabric. Hence, we use the term Microsoft Fabric throughout this post to refer to Power BI or Power BI Service.

The Problem

Implementing incremental refresh on Power BI is usually straightforward if we carefully follow the implementation steps. However in some real-world scenarios, following the implementation steps is not enough. In different parts of my latest book, Expert Data Modeling with Power BI, 2’nd Edition, I emphasis the fact that understanding business requirements is the key to every single development project and data modelling is no different. Let me explain it more in the context of incremental data refresh implementation.

Let’s say we followed all the required implementation steps and we also followed the deployment best practices and everything runs pretty good in our development environment; the first data refresh takes longer, we we expected, all the partitions are also created and everything looks fine. So, we deploy the solution to production environment and refresh the semantic model. Our production data source has substantially larger data than the development data source. So the data refresh takes way too long. We wait a couple of hours and leave it to run overnight. The next day we find out that the first refresh failed. Some of the possibilities that lead the first data refresh to fail are Timeout, Out of resources, or Out of memory errors. This can happen regardless of your licensing plan, even on Power BI Premium capacities.

Another issue you may face usually happens during development. Many development teams try to keep their development data source’s size as close as possible to their production data source. And… NO, I am NOT suggesting using the production data source for development. Anyway, you may be tempted to do so. You set one month’s worth of data using the RangeStart and RangeEnd parameters just to find out that the data source actually has hundreds of millions of rows in a month. Now, your PBIX file on your local machine is way too large so you cannot even save it on your local machine.

This post provides some best practices. Some of the practices this post focuses on require implementation. To keep this post at an optimal length, I save the implementations for future posts. With that in mind, let’s begin.

Best Practices

So far, we have scratched the surface of some common challenges that we may face if we do not pay attention to the requirements and the size of the data being loaded into the data model. The good news is that this post explores a couple of good practices to guarantee smoother and more controlled implementation avoiding the data refresh issues as much as possible. Indeed, there might still be cases where we follow all best practices and we still face challenges.

Note

While implementing incremental refresh is available in Power BI Pro semantic models, but the restrictions on parallelism and lack of XMLA endpoint might be a deal breaker in many scenarios. So many of the techniques and best practices discussed in this post require a premium semantic model backed by either Premium Per User (PPU), Power BI Capacity (P/A/EM) or Fabric Capacity.

The next few sections explain some best practices to mitigate the risks of facing difficult challenges down the road.

Practice 1: Investigate the data source in terms of its complexity and size

This one is easy; not really. It is necessary to know what kind of beast we are dealing with. If you have access to the pre-production data source or to the production, it is good to know how much data will be loaded into the semantic model. Let’s say the source table contains 400 million rows of data for the past 2 years. A quick math suggests that on average we will have more than 16 million rows per month. While these are just hypothetical numbers, you may have even larger data sources. So having some data source size and growth estimation is always helpful for taking the next steps more thoroughly.

Practice 2: Keep the date range between the RangeStart and RangeEnd small

Continuing from the previous practice, if we deal with fairly large data sources, then waiting for millions of rows to be loaded into the data model at development time doesn’t make too much sense. So depending on the numbers you get from the previous point, select a date range that is small enough to let you easily continue with your development without needing to wait a long time to load the data into the model with every single change in the Power Query layer. Remember, the date range selected between the RangeStart and RangeEnd does NOT affect the creation of the partition on Microsoft Fabric after publishing. So there wouldn’t be any issues if you chose the values of the RangeStart and RangeEnd to be on the same day or even at the exact same time. One important point to remember is that we cannot change the values of the RangeStart and RangeEnd parameters after publishing the model to Microsoft Fabric.