Performance Investigation
Measure before optimizing. A performance suspicion is not evidence of a bottleneck.
Workflow
- Define the affected workload, user impact, target metric, percentile, environment, and acceptable threshold.
- Establish a repeatable baseline with representative data and controlled conditions.
- Profile the relevant resource: wall time, CPU, allocations, memory growth, queries, network, disk, rendering, or bundle composition.
- Rank bottlenecks by measured contribution and form falsifiable hypotheses.
- Change one factor at a time and compare against the same baseline.
- Check correctness, cold/warm behavior, tail latency, variance, resource trade-offs, and scale characteristics.
- Add a stable regression benchmark or budget when the project can maintain it.
Report environment, commands, dataset, repetitions, before/after results, uncertainty, trade-offs, and remaining bottlenecks. Do not generalize microbenchmark results beyond their measured workload.
1---2name: performance-investigation3description: Investigate latency, throughput, memory, CPU, I/O, bundle size, or scalability using measurement, profiling, experiments, and benchmarks.4---56# Performance Investigation78Measure before optimizing. A performance suspicion is not evidence of a bottleneck.910## Workflow11121. Define the affected workload, user impact, target metric, percentile, environment, and acceptable threshold.132. Establish a repeatable baseline with representative data and controlled conditions.143. Profile the relevant resource: wall time, CPU, allocations, memory growth, queries, network, disk, rendering, or bundle composition.154. Rank bottlenecks by measured contribution and form falsifiable hypotheses.165. Change one factor at a time and compare against the same baseline.176. Check correctness, cold/warm behavior, tail latency, variance, resource trade-offs, and scale characteristics.187. Add a stable regression benchmark or budget when the project can maintain it.1920Report environment, commands, dataset, repetitions, before/after results, uncertainty, trade-offs, and remaining bottlenecks. Do not generalize microbenchmark results beyond their measured workload.
egeetas