Making the uploads non-fatal moved the failure to the downloads (#74)

[sotashimozono]

Follow-up to #76, which was incomplete. Closes the rest of #74's resilience half.

What #76 got wrong

#76 stopped a full artifact storage quota from failing the label job, and it worked — the shards actually ran for the first time. All eight then failed one step later:

Unable to download artifact(s): Artifact not found for name: testshards-timings

An upload that is allowed to fail has a consumer that is not. Both halves have to degrade together.

…and the obvious fix would be wrong

continue-on-error on the download is not available here, and the reason is already written next to it:

The ONE history labels read, not one this shard fetched for itself. A per-shard fetch let a single transient failure put one shard on a different assignment function from the rest — which double-runs some units and skips others, with every shard still green.

A per-shard tolerance reintroduces exactly that divergence, and it is the silent kind.

The shape that works

The decision is made once, in the job that knows, and shared:

  labels:
    outputs:
      timings: ${{ steps.timings.outcome }}
  test:
    - uses: actions/download-artifact@v8
      if: ${{ needs.labels.outputs.timings == 'success' }}

needs.labels.outputs.timings is one value for the whole matrix, so either every shard has the history or none does — the invariant the comment protects is preserved, not weakened.

"None" is a state the package already defines. load_timings returns an empty table for a missing path (ownership.jl:200), which is what the first run of any suite gets. So the fallback is the ordinary unweighted assignment, not a new code path. A ::warning:: says the balance may be worse and that the result is not.

prebuild gets the same shape

It is an optimisation outright: a shard without the shared compiled cache precompiles for itself and computes the same answer, slower. Its upload becomes non-fatal, its outcome an output, and both the download and the adopt step gate on it — otherwise a full quota still takes the whole matrix down through prebuild, which is what would have happened next.

Unchanged

Still fatal, still ungated: Every unit ran, exactly once, and the refusal inside restore_counters. Coverage and timings are reports; completeness is a claim.

0.3.36 → 0.3.37.

[github-actions]

📚 Docs preview: https://codes.sota-shimozono.com/TestShards.jl/previews/PR77/

(updates on each push to this PR)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!