InvestorsCareersClient support
TulanFabric + Agentic AI

Direct Lake in production: nine lessons from 40 deployments

Direct Lake delivers import-mode speed without import-mode copies, if you design for it. What we'd tell every team before they start.

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.

Let's put agents to work on your Fabric estate.

Book a 45-minute working session with a Tulan architect. Bring your hardest data problem; leave with a plan.