[mistraloz]

Summary

Note

The builder uses the release candidate moby/buildkit:v0.35.0-rc1 (see below): bump it in workflow.yml when
BuildKit 0.35.0 is released.

This PR divides the runner minutes of the CI by 2 to 3.5, and makes every run faster (see the measures below):

• Bug
• Feature
• Breaking changes

• The 3 variants of a PHP version are built in one job, from common layers: the base and the extensions are built once
  instead of once per variant. The extensions are installed once, then the slim part of the variant is applied on them for
  the fat image (see Contributing). The fat images no longer contain the layers of the slim
  image of their variant, their content is unchanged.
• The tests of a job run 2 test files at a time (tests-suite/run_tests.sh), the NodeJS images are now tested.
• The images are pushed 4 at a time, with retries (ghcr.io sometimes answers "unknown blob").
• The layers are cached in ghcr.io/thecodingmachine/php-cache:
  • the runs of v5 read and write it, so a push or a manual run only rebuilds what changed;
  • a separate workflow (purge-cache.yml) deletes the cache each Sunday, 30 minutes before the weekly run, so that
    the images take the updates of the packages each week. It can also be run manually;
  • the pull requests only read it: no pull request can write in the cache of the published images.

BuildKit 0.34 does not find in the cache the layers of a target built from another target, with a new builder as on
GitHub. BuildKit 0.35.0-rc1 fixes it.

It also fixes mod_cgid (APACHE_EXTENSION_CGID=1), which did not start in the slim-apache image and with the
built-in Apache of the fpm variant: its socket directory was missing.

Duration of the whole workflow, from its start to its end, and runner minutes:

Run	Before	After
Weekly run of v5 (from an empty cache)	18 to 19 min, 211	15 min, 96 (fork)
Push or manual run during the week	18 min, 211	10.8 min, 69 (fork)
Pull request, without cache	13.2 min, 168	10.3 min, 75
Pull request, with the cache of v5		7.9 min, 58 (fork)

"Before" is measured on the last runs of this repository (weekly runs of October and November, CI of #429): their
24 build jobs waited up to 7 min in the queue. The new workflow runs 8 jobs at a time, they start within 5 s. The
cache does not exist in this repository yet, so the runs marked "fork" were measured on a fork (with the same queue
time). The purge of the cache (about 1 min) runs 30 min before the weekly run, it is not counted.

Build of PHP 8.4 measured locally with a new builder: 753 s of CPU without cache, 297 s with the cache.

Test plan (required)

• The images of PHP 8.4 built before and after were compared (configuration and files): only the socket
  directory of mod_cgid differs.
• New test test_cgid (variant-apache.sh), it fails on the current images.
• The whole test suite passes on the 4 PHP versions, locally and in the runs of the fork.
• The cache was checked on the fork: the purge workflow empties it, the next run fills it, a later run and a pull request
  reuse it. Locally, a new builder finds the 195 build steps of PHP 8.4 in the cache.
• test_dockerStop waited 3 seconds before stopping the container: it now waits until its processes are started (it
  failed once on a loaded runner).

Checklist

• I followed the guidelines in CONTRIBUTING guide
• I have squashed any insignificant commits
• This change has comments for package types, values, functions, and non-obvious lines of code