The rolling web build lost a run to ghcr.io answering the login with a TLS
handshake timeout. Nothing was wrong with the code: the app image had already
been built and pushed, and the worker image job failed before it started, on
the one step in the whole workflow that had no way to try again. Both jobs run
on the same self-hosted runner and reach for the registry in the same second,
which is presumably why one of the pair got the timeout and the other did not.
So the login is a plain `docker login` with five attempts and a backoff rather
than an action with no retry in it. The release workflow gets the same
treatment and needs it more: it is dispatched by hand and has already tagged
the release by the time it reaches this step.
Logging in by hand means logging out by hand. The action did that in a post
step, and this runner is self-hosted, so without it the credentials would sit
in the docker config after the job is over.
Felitendo
90d169130e