Merge builds fail for Volto 19alphas where an unknown digest fails to fetch from the registry

[fredvd]

To improve our image building with the available github runners instead of depending on hosted runners, we now split the container image builds between the amd64 and arm architectures. These are build on the matching runner architecture, the 'parts' are uploaded as digest without tag information to the registry, and one final 'merge' command to the registry connects/bundles the 2 digest in a tagged ready to use public image. only the digest hash matters and is written in the filename of to an otherwise empty file during the run and upload as artefact.

Our builds sometimes fail with a digest that cannot be found on the registry. I added 'ls' command in the workflow to check which digest hash is uploaded on the separate image architecture and which hashes are retrieved for the merge. I think something is wrong with the cache keys:

I checked the Volto 19alpha19 GHA you mentioned yesterday. It's weird. there is a digest fiel retrieved as artefact which sha I cannot find in the rest of the run as being generated.

I did notice something odd, which makes me suspect we need to add something to the cache keys, maybe the run job count.

Seeing this workflow run, with 2 retries, failing to build the pnpm-dev-image: https://github.com/plone/plone-frontend/actions/runs/20026446241/job/57425911221

attempt 1:
a282e2e119c8ae95a9d18499b34fb6d7fcbc257a21a591dcc90ac5521fcd2587
c1f728109455bb70b92e0da9f445a1132132aee1f90b1d6754b03739b4d46d11
2587 not found
attempt 2:
145b38a7c4cba3458d15bc1bdcb2a15526c416d0f1c5d20c2577149606769cf5
a282e2e119c8ae95a9d18499b34fb6d7fcbc257a21a591dcc90ac5521fcd2587
2587 not found.
attempt 3:
88ce1e145a0c9b54eeabbb0a20301515afa947371dd70a4882e6a77fdc5103dc
de385b0b6240bed3b99c4ebd111ebf091950a9d48ff7aa16a3527106e7b4dbdb
-> goes well.

So It fails 2 times to merge the pnpm-dev-image, but one of the hashes from attempts 1, shows up again at attempt 2....

[fredvd]

One thing I considered while implementing the worklfow was to include the run attempt number in the artefact cache key. Now that I see cache keys reappear from earlier test runs, lets try that first.