Fabric Throttling: Diagnose the Operation and Time Window
Use operation evidence and capacity metrics to investigate delays without treating a burst as unlimited compute.
英语参考文章
Recognize the symptom before choosing a fix
A delayed report, rejected request, or late background job can have several causes. Preserve the operation time, workspace, request details, and visible error before changing capacity. Distinguish capacity pressure from a source-system timeout, a permissions failure, or an application error. Those failures require different responses.
Bursting is not a permanent entitlement
Fabric can account for consumption over smoothing windows, and some work can temporarily consume beyond a capacity's nominal rate. This does not create unlimited free compute or a universal maximum burst multiplier for every workload. Consult the current Microsoft throttling guidance and the Capacity Metrics app. The absence of an immediate rejection does not prove that a schedule is sustainable.
A practical investigation
In an illustrative scenario, a report becomes slow after a large batch starts. Compare operation metrics before the batch, during it, and after it finishes. Check whether the same symptoms occur on another day with a different schedule. Avoid repeatedly restarting the job without understanding retries; that can add work and complicate the evidence.
Recovery and validation
- Identify the resource-intensive operation and its owner.
- Check the current consumption and rejection details.
- Evaluate query or job changes and scheduling alternatives.
- Test any proposed scale change on a representative workload.
- Define who can pause, resume, or resize the capacity and how failures are communicated.
Record whether the intervention restored the expected reporting and refresh behavior. A larger SKU may provide headroom, but it does not repair an incorrect transformation, unbounded retry loop, or unsupported data source. Keep the diagnosis and the response linked so the next incident can be assessed using evidence rather than a generic scaling rule.