d6d7ff8c9e made iter_role_images yield one ref per version_variants suffix so
the mirror copies the extra tags. Two of its four consumers assumed one ref per
service. The resolver writes services[service] = {image, version}, so the last
ref won and the applications override carried v1.9.6-cuda; the engine role
appends -cuda itself and asked the mirror for v1.9.6-cuda-cuda, a tag nothing
ever pushed. The version updater builds one entry per ref and writes back into
meta/services.yml, so it would have replaced the declared version with a cuda
tag once this branch reaches the bot on main.
ImageRef now carries `derived`. The resolver and the updater skip those refs;
the mirror sync keeps iterating all of them, which is what the variants exist
for, and the mirror wait already keyed on the full ref including the version.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Autoformat output from a quality run, carried along since. The syntax tree is
unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four libretranslate jobs in run 37173067619 failed with
failed to solve: ghcr.io/…/mirror/docker.io/libretranslate/libretranslate:v1.9.6-cuda: not found
svc-ai-libretranslate-engine and web-svc-libretranslate append '-cuda' to the version
in vars/main.yml when the gpu flag is set. iter_role_images is the only enumeration
the mirror pipeline sees and it reads image and version alone, so the derived tag was
never a candidate for mirroring and resolved against a registry that has no such tag.
Upstream does publish it: Docker Hub answers 200 for v1.9.6-cuda, the mirror 404.
The suffixes are now declared next to the version they extend, and discovery yields a
ref per suffix. A suffix rather than a full tag keeps the version number in one place
when it is bumped.
The gpu flag is static: it is set in meta/services.yml, no inventory overrides it, and
resource_filter reads configuration rather than hardware. Every host therefore takes
the cuda path today, which is why all four jobs failed rather than some.
Rejected: mirror_policy: skip, which has no users in this repository and would route
these two images around the mirror to Docker Hub permanently, reintroducing the rate
limit the mirror exists to avoid. Also rejected: baking '-cuda' into version, which
fixes the mirror but removes the GPU switch.
The test pins both halves; without the discovery change its two variant cases fail
and its two base cases still pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
kevinveenbirkenbachandClaude Opus 5