Omar Hassan

Got it working! The issue was that my cloud connection was using my personal OAuth token. For production, I'll switch to a service principal-based connection as recommended in the docs.

One thing I'd add to the DR docs: the timing of bacpac exports matters. If you export at 2am and the writeback database is being actively written during a planning session, you may have partial transactions. Recommend documenting the freeze-period…

We do exactly this — Python notebooks in Fabric generate Prophet forecasts for our demand planning, write to SQL, and Fabric Plan's Planning sheet shows the statistical baseline alongside the planner's override. Best of both worlds.

Also check the Capacity Metrics app while loading. If you see CU spikes on that first query, you may be triggering Direct Lake fallback to DirectQuery mode, which is significantly slower.

One practical difference: Reference columns support 'Match Multiple' — returning multiple values from all matching records in the source table. This is useful for one-to-many relationships where LOOKUPVALUE() would only return the first match.