Working Code skill
Profile application performance
Profile application performance with a representative workload, synchronized measurements, call-path evidence, and before-after verification.
Measure the bottleneck before optimizing.
When to use
- Use after a performance symptom and success threshold are defined.
- Do not profile production with unsafe overhead, sensitive captures, or unapproved load.
Procedure
- Define journey, workload, environment, version, metric, percentile, baseline, and target.
- Reproduce with production-shaped data and realistic concurrency while controlling warmup and cache state.
- Synchronize request traces, CPU profiles, allocation or heap data, I/O, database, network, runtime, and host metrics.
- Identify where elapsed time or resource use accumulates across the critical path.
- Form a ranked hypothesis and choose the smallest discriminating experiment.
- Change one material factor while preserving correctness and recording side effects.
- Repeat enough trials to compare distributions, not a single fast run.
- Run correctness, load, and regression checks and monitor after rollout.
--- name: profile-application-performance category: code description: Profile application performance with a representative workload, synchronized measurements, call-path evidence, and before-after verification. Use when latency, throughput, CPU, memory, I/O, or cost needs improvement. --- # profile-application-performance Measure the bottleneck before optimizing. ## When to use - Use after a performance symptom and success threshold are defined. - Do not profile production with unsafe overhead, sensitive captures, or unapproved load. ## Procedure 1. Define journey, workload, environment, version, metric, percentile, baseline, and target. 2. Reproduce with production-shaped data and realistic concurrency while controlling warmup and cache state. 3. Synchronize request traces, CPU profiles, allocation or heap data, I/O, database, network, runtime, and host metrics. 4. Identify where elapsed time or resource use accumulates across the critical path. 5. Form a ranked hypothesis and choose the smallest discriminating experiment. 6. Change one material factor while preserving correctness and recording side effects. 7. Repeat enough trials to compare distributions, not a single fast run. 8. Run correctness, load, and regression checks and monitor after rollout. ## Done - A performance profile report records representative workload, baseline, synchronized evidence, bottleneck, and hypothesis - Repeated before-after tests verify target improvement, correctness, resource tradeoffs, and production monitoring