Files
cloud-apache-container/docs/arm64-publish-policy.md
shadowdaoandClaude Sonnet 5.5 344c5e8d0d ci: address Opus review of the arm64 policy PR
- job timeouts 130 (cac, cac-fpm) / 80 (shared-httpd) so the retry fits; test pins the budget
- verify-image-arch.py: bounded docker calls, pull retried 3x, exit 4 'could not pull' (not an
  arch verdict), pulled image removed afterwards
- ARM64_DIGEST only taken from the first attempt when it succeeded
- comments/docs: max-parallel is per-job and unverified; guard also fires on skipped/cancelled

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0185ukXi4YX5FxJGqhi6qab6
2026-11-19 21:12:56 -08:00

1000 B

arm64 in build-push.yaml (cac, cac-fpm, shared-httpd)

The arm64 leg is best effort (QEMU on the x86 runner) but never silently lost:

  • Matrix families run at most 2 legs at once (max-parallel: 2); 2026-11-19 showed that 8 concurrent QEMU legs fail on network steps.
  • A failed arm64 leg is retried ONCE after 90 s (Retry arm64). Assemble uses the first success.
  • Before publishing, scripts/assemble-multiarch-index.sh --verify-bin checks ELF machine types (amd64 = x86-64, arm64 = AArch64). A wrong arm64 counts as a failed leg; a wrong amd64 aborts.
  • --guard-existing-arm64: if the build ends without arm64 (failed, skipped or cancelled leg) and the current tag has an arm64 entry, the assemble step FAILS and publishes nothing (the amd64 digest is in the error). Options: re-dispatch only=<leg>, or dispatch with allow_amd64_only=true (security rebuilds) and then restore arm64 with the arm64-attach workflow. Push-triggered runs have no input, so the guard is always on there.