Direct Lake is the most important Power BI feature in years: import-mode performance reading directly from Delta tables in OneLake, with no scheduled refresh copies. After 40 production deployments, here's what we tell every team.
1. Design gold tables for the model, not the other way around
Direct Lake rewards star schemas with narrow, well-typed columns. If your gold layer is a wide, denormalized export, fix it upstream.
2. V-Order is not optional
Write gold tables with V-Order enabled. It's the single biggest driver of cold-query performance.
3. Watch your guardrails
Every SKU has row-count and model-size guardrails. Exceed them and queries fall back to DirectQuery, silently. Monitor fallback in production.
4. Framing is your new refresh
Framing points the model at the latest table version. Coordinate framing with your pipeline completion, or users will see half-loaded data.
5. Keep calculated columns out
Push calculated columns into the lakehouse. Direct Lake models are cleaner and faster without them.
Direct Lake doesn't forgive a messy gold layer. It exposes it.
6 through 9
Use composite models sparingly. Test with realistic concurrency, not a single user. Document measures for Copilot from the start. And treat the semantic model as a product with an owner, a roadmap, and a service level.