performance-testinglisted
Install: claude install-skill YevheniiaQA/qa-agent-toolkit
# Performance Testing
Runs **after** functional testing, not alongside it. A performance figure taken
from a build whose behaviour is not yet verified measures the wrong thing, and
running both at once means each becomes the other's load.
Keep the investment proportionate. Most projects need a smoke test, one load
test at expected traffic, and a baseline to compare against. Stress, spike and
soak are worth building when there is a specific question to answer.
## Before running anything
**Approval is mandatory.** Load testing generates real traffic against real
infrastructure. Against a shared staging environment it degrades everyone else's
work; against production it is an outage you caused. Confirm in chat: which
environment, what peak load, what window, and who has been told.
**Never test production** unless the engineer explicitly directs it, the window
is agreed, and there is a documented abort procedure.
**Record the conditions before the numbers.** Environment size, dataset size,
whether other traffic is present, whether caches are warm. A latency figure
without these is uninterpretable a month later, and comparing two runs with
different conditions is worse than having no comparison at all.
## Test types
| Type | Shape | Answers |
|---|---|---|
| **Smoke** | 1–5 VUs, 30s | Does the script work and the system respond at all? Always run first |
| **Load** | Ramp to expected peak, hold, ramp down | Does it meet its targets under normal traffic? |
| **Stress** | Ra