Overview
This is a typical engagement for a data platform team that shares one Fabric capacity between report users, scheduled pipelines and notebooks. Users see capacity errors at some hours and not others, and the team cannot say which item causes them. AceMQ engineers, with a partner subject-matter expert where needed, read the capacity's history and write the findings.
Challenge
Microsoft documents that Fabric smooths background operations over 24 hours and interactive operations over 5 to 64 minutes, gives 10 minutes of overage protection, then delays interactive operations by 20 seconds, rejects them once the next 60 minutes are spent and rejects all requests once the next 24 hours are spent. Because the throttling state belongs to the capacity, one heavy job can affect every workspace assigned to it. Raising the SKU is one fix, but pausing a capacity bills its accumulated future usage and overage billing costs three times the normal rate.
Environment
One Fabric capacity assigned to several workspaces, scheduled Data Factory pipelines, Spark notebooks, a Warehouse, Power BI semantic models and reports, and the Capacity Metrics app installed by a capacity admin.
Approach
AceMQ engineers read the Utilization, Throttling and Overages views and the System events table of the Capacity Metrics app, drill to the timepoints where limits were crossed, and list the interactive and background operations behind them. They compare three options on the measured usage: a larger SKU, a second capacity for the heaviest workspace, and a changed job schedule.
Solution
- Timeline of throttling events with the stage reached and the minutes to burndown for each
- Top operations by capacity unit use at each crossing, split into interactive and background
- Workspace to capacity layout and a proposal for a second capacity where one workload dominates
- SKU, pause and resume, and overage billing compared on cost and risk, with the assumptions written down
- Schedule changes for jobs that start together, and alerting before users see errors
- Rehearsal plan for the first change, with a rollback point and an owner for each step
Outcome
The engagement delivers a written review with the throttling timeline, the operations that drove it, a recommended capacity layout or schedule, the cost comparison with its assumptions and a rehearsal plan for the first change. The customer's team decides from the facts and runs the change, or AceMQ can run it alongside them.