kimtth Goal: know how the system behaves under realistic and peak load.
Use for:
- validating capacity before a launch or sale
- finding bottlenecks and breaking points
- catching performance regressions
Workflow:
- Define the scenario: traffic shape, mix, and target SLOs.
- Model realistic user behavior and data, not uniform hits.
- Ramp load gradually; record latency percentiles and errors.
- Push to the breaking point to find the real limit.
- Correlate results with resource and dependency metrics.
- Fix the top bottleneck, then re-test.
Measure:
- p50/p95/p99 latency, not just averages
- throughput and error rate under load
- resource saturation (CPU, memory, connections)
- downstream dependency limits
Rules:
- test against a production-like environment and data
- look at tail latency; averages hide pain
- change one variable per run to attribute results
- protect real systems; never load-test production blindly