Performance Testing Skill

Performance Testing

Progressive profiles only: Smoke → Load → Stress → Soak → Spike. Never jump to stress first.
Encode SLOs as k6 thresholds that fail the run — eyeballing is not a gate.
Every VU iteration includes sleep() think time (prevents accidental DDoS).
Send traceparent (+ load-test id) so slow percentiles map to OpenTelemetry spans.
Prove resilience with chaos (latency/failure injection); pair resilience patterns.
Run load in CI to catch regressions. Diagnose bottlenecks via traces, not guesses.
Before full profile scripts or chaos helpers, read the matching resource below.

Confirm unit/integration pass. Failing smoke with 1 VU = functional bug, not a perf problem.
Write smoke (1 VU, ~1m, low error rate). Pass before higher load.
Add load at expected traffic with p95/p99 and error-rate thresholds.
Escalate: stress (find break point), soak (hours, watch memory/leak), spike (burst + recovery).
Correlate: inject W3C traceparent and x-load-test-id; query traces for pool exhaustion, N+1, rate limits.
Chaos: inject latency/failures (in-process helpers or Toxiproxy). Verify timeouts, retries, circuit breakers fire. Before implementing injectors, read references/chaos.md.
CI: run load on main/nightly; upload JSON/HTML artifacts; fail on threshold breach.

references/load-profiles.md — smoke/load/stress/soak/spike options, thresholds, k6 CLI, tracing headers. Read when authoring scripts.
references/chaos.md — latency/failure helpers, Toxiproxy, chaos scenarios, CI sketch. Read when proving resilience under fault.
references/rationalizations.md — excuse→reality, pitfalls, red flags. Read when tempted to skip a gate.

Not for unit/mutation tests (writing-tests, mutation-testing) or React render micro-opts (react-development).
Pyramid placement of load/chaos: testing-strategy. Patterns under test: resilience. Trace correlation: observability. Bottleneck follow-up: debugging-methodology.