Testing

This page is for anyone deciding whether to trust fpm-tune near a host: what the suite covers, and which commands are safe to try.

What stands behind it

Unit tests cover the decision logic: the budget, the measurement, the allocator (a randomised sweep of generated plans holds the never-over-budget invariant) and hysteresis. make check runs fmt-check, tidy-check, vet, lint, test, vulncheck and license-check.

The part that can take a host down is tested against a real php-fpm. CI installs one and runs:

testing/e2e.sh: the end-to-end path with the real binary. A dry run that changes nothing, a reload the master survives with its sockets intact, and a second run refused while the first holds the lock.
testing/chaos.sh: the perturbations a host has. A site added, a site removed and php-fpm reloaded before the tool notices, the master restarted underneath it, and a breakage elsewhere the tool must not react to by deleting its own work. It exists because a soak on a VM found a removed pool still declared in zz-fpm-tune.conf, which kept the master from starting.
testing/mutations.py: every named safety guard removed from the source in turn, with the suite required to fail each time. A test that passes with its guard deleted is testing nothing.

CI runs these as separate jobs: test, lint, vulncheck, sbom (with the license check), mutations, integration, and cross-compiled builds for linux and darwin on amd64 and arm64. On a machine with php-fpm installed, make integration runs e2e and chaos, and make mutations runs the sweep.

Trying it without risk

fpm-tune plan writes no configuration and reloads nothing. It records what it observed to the state file; --no-learn turns that off.
sudo fpm-tune apply --dry-run renders and validates the change against a copy of the pool directory and writes nothing.
fpm-tune serve and sudo fpm-tune install-service run in advisory mode: they measure, publish metrics and write the recommendation file, and change nothing.
apply --no-learn and serve --apply --no-learn are refused, because a run that writes has to record it, or the next one reloads the pool this one just reloaded.