Paul Turley talks about fashion:
One of the core best practice guidance principals for Power BI modeling is to avoid including columns that aren’t absolutely necessary for analytic reporting. Every column uses precious memory and especially long, unique values that don’t compress very well. When consulting clients bring me large models that require expensive capacity licensing and pose report performance issues, my first inclination is to see what column data can be carved out of the model; and perhaps moved to another table for a drill-through report.
The product team came up with a very clever way to reduce the in-memory footprint of a Direct Lake semantic model: hold a popularity contest! The semantic model engine will only keep columns in memory based on their hotness. I mean this literally…
Read on to learn a bit more about the algorithm in play and how it differs from a naive Least Recently Used cache.