Seven Fabric capacity sizing mistakes
(and the honest fix).

1 · Sizing to the peak instead of the smoothed profile

Fabric smooths consumption over time windows, so short bursts above nominal capacity are absorbed. Teams that size to their scariest five minutes pay peak prices twenty-four hours a day. Fix: chart a realistic day — pipeline windows, analyst hours, overlaps — and size to that profile, accepting that rare collisions queue briefly.

2 · Modeling compute without the licensing interplay

In BI-heavy organizations, how report viewers are licensed can swing the total more than one capacity step. Capacity size and Power BI licensing interact (the F64 boundary in current terms). Fix: put viewer licensing and compute into the same model and verify the current threshold terms the week you decide — never compare capacities on compute alone.

3 · One capacity for everything

A single pool means a runaway dev notebook and your CEO's dashboard share the same capacity units. Fix: isolate noisy or experimental workloads onto a small pay-as-you-go capacity that pauses at night. The isolation usually costs less than the incidents it prevents.

4 · Forgetting that storage compounds

OneLake bills per GB-month on its own meter and never pauses. CDC feeds, snapshots and abandoned experiment tables accrete. Fix: model storage with a monthly growth rate and a retention policy, and review what's actually growing each quarter.

5 · Reservation as a default instead of a conclusion

Reservations trade commitment for a lower rate — and bill every hour, used or not. Bought before usage is understood, they lock in the wrong size at a discount. Fix: run pay-as-you-go with pause discipline until the pattern is boring, then reserve the baseline you have proven and keep bursty remainder on PAYG.

6 · No pause discipline on non-production capacities

A development capacity left running around the clock bills three to four times the hours anyone works. Fix: schedule pauses (nights, weekends) for every capacity that doesn't serve production consumers, and make un-pausing a one-click norm rather than keeping things hot "just in case".

7 · Modeling once and never re-checking

Rates change, workloads drift, licensing terms evolve. A model from last budget season quietly becomes fiction. Fix: keep the model as a living scenario — re-run it quarterly with current rates and actual usage, and record what changed.