Load testing one endpoint before a launch

A full load test of a system is a project; a load test of the one endpoint a campaign will hit is an afternoon.

$ k6 run --vus 50 --duration 2m script.js

  http_req_duration  avg=142ms  p95=410ms  p99=2.1s
  http_reqs          21,408     178/s
  http_req_failed    0.02%

# the p99 is the finding: 2.1 seconds, from connection
# pool exhaustion at 50 concurrent — max_connections was
# 60 and php-fpm was configured for 80 children.

The mismatch between the worker count and the connection limit is the kind of thing that only appears under concurrency, and it appeared in twenty minutes of testing rather than during the campaign. Testing one endpoint at a realistic concurrency answers the question that was actually asked, and the temptation to broaden it into a full performance exercise is what stops these from happening at all.