Load testing before a launch, again, and the same finding

A second launch, a second load test, and the same class of finding as last year for a different reason.

$ k6 run --vus 80 --duration 3m script.js
  http_req_duration  p95=380ms  p99=4.1s
  http_req_failed    0.4%

# last year: php-fpm children exceeded max_connections.
# this year: the connection pool is fine and the
# OPCACHE is not — 80 concurrent requests across 20
# workers pushed the interned strings buffer over its
# limit and triggered a restart mid-test.

Two years, two resource limits, both of them a configured number that was fine at the previous concurrency. The general finding is that a load test finds whichever limit is nearest, and fixing one moves the ceiling to the next — which is an argument for running it before every launch rather than treating last year’s result as knowledge.