Paul Turley shares some advice:
Since the introduction of Power Pivot for Excel, SQL Server Analysis Services Tabular, Azure Analysis Services and Power BI; the native mode for storing data in a semantic data model (previously called a “dataset” in Power BI) has been a proprietary file structure consisting of binary and XML files. These file structures were established in the early days of multidimensional SSAS back in 2000 and 2005. When an Import mode model is published to the Power BI service, deployed to an SSAS server or when Power BI Desktop is running, data for the model is loaded into memory where it remains as long as the service is running. When users interact with a report or when DAX queries are run against the model, results are retrieved very quickly from the data residing in memory. There are some exceptions for very large models or when many models in the service don’t all fit into memory at the same time, the service will page some or all of the model in and out of memory to make sure that the most-often used model pages remain in memory for the next user request. But, for argument’s sake, the entire semantic model sits in memory, waiting for the next report or user request.
Rather than the proprietary SSAS file structure, Direct Lake models use the native Delta-parquet files that store structured data tables for a Fabric lakehouse or warehouse in One Lake. And rather than making a copy of the data in memory, the semantic model is a metadata structure that shares the same Delta-parquet file storage. As soon as a report runs against a model, all of the model data is paged into memory which then behaves a lot like an Import mode model. This means than while the model remains in memory, performance should about the same as Import, with a few exceptions.
Read on to see what the capabilities of Direct Lake mode are today, as well as a few design considerations for your Microsoft Fabric architecture.