[ale18V]

What & why

Currently in verify.yml:

1. download-datasets restores datasets/ from actions/cache@v5, and then re-zips and re-uploads the entire folder to actions/upload-artifact@v7 on every single CI run so test jobs can pull it with actions/download-artifact.
2. build waits on download-datasets (needs: [download-datasets, static-analysis]) even though _build.yml never downloads or uses datasets/.
3. build-server-docker (~2m11s) is bundled inside _build.yml alongside build-topological (~1m06s), forcing all 14 test jobs in _test.yml to sit idle for an extra ~1 minute waiting for the Docker smoke test even though all-builds is already ready and no test job uses the Docker image.

Changes

• Warm-only dataset cache (.github/actions/download-datasets/action.yml & pull-dataset-artifacts/action.yml):
  • Use lookup-only: true on actions/cache@v5 in download-datasets so cache hits complete in seconds without downloading or re-uploading a duplicate workflow artifact.
  • Restore datasets/ directly from actions/cache/restore@v5 (with fail-on-cache-miss: true) in pull-dataset-artifacts.
• Remove unnecessary sequential dependencies (.github/workflows/verify.yml & _build.yml):
  • Move download-datasets from build.needs to test.needs.
  • Move build-server-docker out of _build.yml into verify.yml (needs: [static-analysis]) so it runs in parallel with build and test instead of blocking test startup.

[JulienVig]

Good idea to reduce unecessary dependencies! I think we can push this further for the Docker smoke test. However it looks like some dependencies are missing for it in ci-paths.yml .

Looking at the past CI runs it looks like the speedup is pretty unsignificant compared the timing of test/server. It doesn't cost anything to parallelize more but maybe we should focus our efforts on splitting test/server in several parallel jobs?