Performance Smoke Testing

Not full load testing — a smoke check run locally, never against stage or prod.

API

for i in $(seq 20); do curl -s -o /dev/null -w '%{time_total}\n' "$BASE/api/tasks?limit=50"; done | sort -n | awk '{a[NR]=$1} END{print "p50="a[int(NR*.5)]" p95="a[int(NR*.95)]" max="a[NR]}'

Then run the same with 10x the seeded rows. p95 growing more than 3x for 10x the data is an N+1 suspicion — report it with the two numbers. Fail bar: p95 >1s for a simple read on localhost, or any number a criterion states explicitly. 0.5-1s is a note, not a fail.

Concurrency smoke

Only when the change implies it (a shared resource, a write path under contention):

npx -y autocannon -c 10 -d 10 <url>

Errors must be 0.

Web

Measure a production build, never a dev server — dev builds are unoptimized and that is where false fails come from:

npm run build && (nohup npx vite preview --port 4173 > "$QA/preview.log" 2>&1 &)
CHROME_PATH="$CHROME_BIN" npx -y lighthouse@13 http://localhost:4173/<page> --only-categories=performance --form-factor=mobile --screenEmulation.width=360 --chrome-flags="--headless=new" --output=json --output-path="$QA/lh-perf.json"

Report LCP/CLS/TBT against "good": LCP ≤2.5s, CLS ≤0.1, TBT ≤200ms. Fail only for a criterion's own number or a CLS shift you can see on a changed screen; other misses are notes, not fails.

Never

Load-test stage or prod. Use k6 only when the repo already ships k6 scripts — not as the default tool.

Recording

Record cases as category: other with a title prefix perf: — the category enum has no performance value of its own.