[Experimental] Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows by potiuk · Pull Request #66348 · apache/airflow

potiuk · 2026-05-04T10:52:02Z

Summary

The combined ci-amd-arm.yml alternated ARM and AMD on the canary slot via a
runtime RUNNERS_TYPE_CROSS_MAPPING lookup of the previous workflow run's
labels. That worked but tied the two architectures together — one canary slot
had to flip between them, badges showed whichever ran last, notifications
spoke for both, and a runtime API call decided the platform on every push.

This change splits the file into a reusable pipeline + two thin wrappers:

_ci-pipeline.yml — reusable workflow (workflow_call) with the
entire pipeline. Takes platform and runner-type as inputs;
build-info echoes the inputs instead of computing them via
cross-mapping. Single source of truth — no per-platform duplication.
ci-arm.yml — slim wrapper. Schedule (canary cron) +
workflow_dispatch only. Calls _ci-pipeline.yml with
platform: linux/arm64.
ci-amd.yml — slim wrapper. pull_request + push (to v3-N-test etc)
- workflow_dispatch. Calls _ci-pipeline.yml with
  platform: linux/amd64. No schedule for now; ARM keeps the canary
  slot. Add a schedule: block here if/when AMD should also act as a
  scheduled canary — the two workflows run in parallel with separate
  concurrency groups.

The runtime cross-mapping in selective_checks.runner_type is left in place
because is_disabled_integration still consumes it; the API lookup against
the now-removed ci-amd-arm.yml returns nothing and falls back to AMD,
which is harmless. Removing the cross-mapping entirely is a follow-up.

Reference updates (every consumer of the old file name)

README.md and dev/update_github_branch_config.py: badges → ci-amd.yml
(the workflow that runs on every push to main)
.github/workflows/ci-notification.yml: workflow-id → ci-amd.yml
.github/workflows/e2e-flaky-tests-report.yml: WORKFLOW_NAME →
ci-amd.yml
scripts/ci/analyze_e2e_flaky_tests.py: default WORKFLOW_NAME →
ci-amd.yml
dev/README_AIRFLOW3_DEV.md, dev/README_RELEASE_AIRFLOW.md: release-cut
verification URL → ci-amd.yml
generated/PYPI_README.md: regenerated from README.md
.github/workflows/update-constraints-on-push*.yml,
release_dockerhub_image.yml: comments referring back to the old file →
_ci-pipeline.yml

Test plan

prek run yamllint clean on all changed workflow files
prek run zizmor clean (template-injection on the print-platform
job's inputs.platform was hoisted to env; secrets: inherit replaced
with explicit DOCS_AWS_* / SLACK_BOT_TOKEN forwarding)
prek run --files <changed> clean across all hooks
All 162 test_selective_checks.py tests pass (the runner_type
tests still pass since they mock requests.get and don't depend on the
real workflow file existing)
Smoke-test in CI: opening this PR will trigger ci-amd.yml. The ARM
canary will fire on its next scheduled cron tick — if both look healthy,
ready for review.

Why draft

Posting as draft for an initial review of the architecture before letting
CI exercise it for the first time. Once the AMD run on this PR completes
green, I'll mark it ready for review.

Was generative AI tooling used to co-author this PR?

Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines

… wrappers Replaces the combined ci-amd-arm.yml workflow with two thin wrappers carrying the full pipeline inline: - ci-arm.yml: schedule (canary cron) + workflow_dispatch only; runs on linux/arm64. - ci-amd.yml: pull_request + push (to v3-N-test etc) + workflow_dispatch; runs on linux/amd64. No schedule for now; ARM keeps the canary slot. Each wrapper carries the full pipeline jobs identically except for platform / runner-type / triggers, so per-job UI grouping is preserved on the GitHub Actions page. The runtime cross-mapping in selective_checks.runner_type is left in place because is_disabled_integration still consumes it; the API lookup against the now-removed ci-amd-arm.yml falls back to AMD harmlessly. References to the old file name are updated in: README.md, dev/update_github_branch_config.py, ci-notification.yml, e2e-flaky-tests-report.yml, scripts/ci/analyze_e2e_flaky_tests.py, dev/README_AIRFLOW3_DEV.md, dev/README_RELEASE_AIRFLOW.md, generated/PYPI_README.md, update-constraints-on-push*.yml, release_dockerhub_image.yml.

The two CI workflow files are physical copies of each other (GH Actions has no cross-file YAML include) — they should differ only in: - header intro comment - workflow name (ARM vs AMD) - triggers (schedule vs pull_request+push) - concurrency group prefix - build-info platform / runner-type outputs - print-platform job name + echo Drift outside this set is a bug. The new `scripts/ci/prek/check_ci_workflows_in_sync.py` normalizes both files against the documented divergences (regex line rules + arch-only block removals) and asserts the rest matches byte-for-byte. Wired up as a local prek hook gated on either workflow file changing. Local UX: when run interactively, the script also diffs each file against `upstream/main` (falls back to `origin/main`, then `main`) and attributes the drift — "only ci-arm.yml changed since baseline; mirror to ci-amd.yml" with the exact diff to copy, OR "add to LINE_RULES / ARM_ONLY_BLOCK in this script if intentional". Falls back to the non-attribution view in CI / non-TTY runs. Set `FORCE_INTERACTIVE=1` to opt in from a non-TTY shell.

jscheffl

I would be good with this. But also have no strong opinion. That that YAML can not include, would save a lot of redundant code.. but with tooling as you provided is acceptable burden.

github-actions · 2026-05-07T23:09:27Z

Backport failed to create: v3-2-test. View the failure log Run details

Note: As of Merging PRs targeted for Airflow 3.X
the committer who merges the PR is responsible for backporting the PRs that are bug fixes (generally speaking) to the maintenance branches.

In matter of doubt please ask in #release-management Slack channel.

Status	Branch	Result
❌	v3-2-test

You can attempt to backport this manually by running:

cherry_picker 8f93637 v3-2-test

This should apply the commit to the v3-2-test branch and leave the commit in conflict state marking
the files that need manual conflict resolution.

After you have resolved the conflicts, you can continue the backport process by running:

cherry_picker --continue

If you don't have cherry-picker installed, see the installation guide.

…rkflows (#66348) (#66664) * Split ci-amd-arm.yml into ci-arm.yml (canary) and ci-amd.yml (per-PR) wrappers Replaces the combined ci-amd-arm.yml workflow with two thin wrappers carrying the full pipeline inline: - ci-arm.yml: schedule (canary cron) + workflow_dispatch only; runs on linux/arm64. - ci-amd.yml: pull_request + push (to v3-N-test etc) + workflow_dispatch; runs on linux/amd64. No schedule for now; ARM keeps the canary slot. Each wrapper carries the full pipeline jobs identically except for platform / runner-type / triggers, so per-job UI grouping is preserved on the GitHub Actions page. The runtime cross-mapping in selective_checks.runner_type is left in place because is_disabled_integration still consumes it; the API lookup against the now-removed ci-amd-arm.yml falls back to AMD harmlessly. References to the old file name are updated in: README.md, dev/update_github_branch_config.py, ci-notification.yml, e2e-flaky-tests-report.yml, scripts/ci/analyze_e2e_flaky_tests.py, dev/README_AIRFLOW3_DEV.md, dev/README_RELEASE_AIRFLOW.md, generated/PYPI_README.md, update-constraints-on-push*.yml, release_dockerhub_image.yml. * Add prek hook that asserts ci-arm.yml and ci-amd.yml stay in sync The two CI workflow files are physical copies of each other (GH Actions has no cross-file YAML include) — they should differ only in: - header intro comment - workflow name (ARM vs AMD) - triggers (schedule vs pull_request+push) - concurrency group prefix - build-info platform / runner-type outputs - print-platform job name + echo Drift outside this set is a bug. The new `scripts/ci/prek/check_ci_workflows_in_sync.py` normalizes both files against the documented divergences (regex line rules + arch-only block removals) and asserts the rest matches byte-for-byte. Wired up as a local prek hook gated on either workflow file changing. Local UX: when run interactively, the script also diffs each file against `upstream/main` (falls back to `origin/main`, then `main`) and attributes the drift — "only ci-arm.yml changed since baseline; mirror to ci-amd.yml" with the exact diff to copy, OR "add to LINE_RULES / ARM_ONLY_BLOCK in this script if intentional". Falls back to the non-attribution view in CI / non-TTY runs. Set `FORCE_INTERACTIVE=1` to opt in from a non-TTY shell. (cherry picked from commit 8f93637)

…llision (#66662) PR #66348 split the combined ci-amd-arm.yml into per-platform wrappers and intentionally deferred re-adding a schedule trigger to ci-amd.yml ('No schedule for now; ARM keeps the canary slot. Add a schedule: block here if/when AMD should also act as a scheduled canary'). As a result: - The AMD 'Tests (AMD)' badge on the README pointed at a workflow that hadn't run on main since 2025-10-25 (it now only runs on PRs and release-branch pushes), so the badge was stale. - The notify-slack job at the bottom of ci-amd.yml is gated on github.event_name == 'schedule' and was therefore dead code: there was no schedule trigger left to fire it. - main on AMD had no scheduled CI monitor at all (ci-notification.yml only watches v3-2-test x ci-amd.yml, and the comment there assumes ARM is reported by its own notify-slack). This change: 1. Restores the previous AMD canary cron in ci-amd.yml at minute :58 of the same hours ARM uses, so the two scheduled runs don't compete for runners at exactly the same minute. 2. Splits the README build-status row for main into two badges (AMD and ARM), drops the obsolete 2.x row (still using the legacy ci.yml), and keeps 3.x AMD-only (ARM has no v3-2-test schedule). Click-throughs go to the workflow page rather than the catch-all /actions URL. 3. Disambiguates the slack-state artifact names that ci-amd.yml, ci-arm.yml, and ci-image-checks.yml share. Before this PR the latent collision did not bite (AMD's notify-slack never fired); restoring the AMD schedule activates it. Each platform now writes its own slack-state-{tests,inventory}-<branch>-{amd,arm} artifact so the de-dup tracker in scripts/ci/slack_notification_state.py keeps independent state per platform. 4. Updates scripts/ci/prek/check_ci_workflows_in_sync.py to allow the new AMD-only schedule block and the per-platform artifact-name suffix as documented divergences. generated/PYPI_README.md is regenerated automatically by the generate-pypi-readme prek hook to mirror README.md.

…llision (#66662) (#66665) PR #66348 split the combined ci-amd-arm.yml into per-platform wrappers and intentionally deferred re-adding a schedule trigger to ci-amd.yml ('No schedule for now; ARM keeps the canary slot. Add a schedule: block here if/when AMD should also act as a scheduled canary'). As a result: - The AMD 'Tests (AMD)' badge on the README pointed at a workflow that hadn't run on main since 2025-10-25 (it now only runs on PRs and release-branch pushes), so the badge was stale. - The notify-slack job at the bottom of ci-amd.yml is gated on github.event_name == 'schedule' and was therefore dead code: there was no schedule trigger left to fire it. - main on AMD had no scheduled CI monitor at all (ci-notification.yml only watches v3-2-test x ci-amd.yml, and the comment there assumes ARM is reported by its own notify-slack). This change: 1. Restores the previous AMD canary cron in ci-amd.yml at minute :58 of the same hours ARM uses, so the two scheduled runs don't compete for runners at exactly the same minute. 2. Splits the README build-status row for main into two badges (AMD and ARM), drops the obsolete 2.x row (still using the legacy ci.yml), and keeps 3.x AMD-only (ARM has no v3-2-test schedule). Click-throughs go to the workflow page rather than the catch-all /actions URL. 3. Disambiguates the slack-state artifact names that ci-amd.yml, ci-arm.yml, and ci-image-checks.yml share. Before this PR the latent collision did not bite (AMD's notify-slack never fired); restoring the AMD schedule activates it. Each platform now writes its own slack-state-{tests,inventory}-<branch>-{amd,arm} artifact so the de-dup tracker in scripts/ci/slack_notification_state.py keeps independent state per platform. 4. Updates scripts/ci/prek/check_ci_workflows_in_sync.py to allow the new AMD-only schedule block and the per-platform artifact-name suffix as documented divergences. generated/PYPI_README.md is regenerated automatically by the generate-pypi-readme prek hook to mirror README.md. (cherry picked from commit 022e35d)

…rkflows (apache#66348) * Split ci-amd-arm.yml into ci-arm.yml (canary) and ci-amd.yml (per-PR) wrappers Replaces the combined ci-amd-arm.yml workflow with two thin wrappers carrying the full pipeline inline: - ci-arm.yml: schedule (canary cron) + workflow_dispatch only; runs on linux/arm64. - ci-amd.yml: pull_request + push (to v3-N-test etc) + workflow_dispatch; runs on linux/amd64. No schedule for now; ARM keeps the canary slot. Each wrapper carries the full pipeline jobs identically except for platform / runner-type / triggers, so per-job UI grouping is preserved on the GitHub Actions page. The runtime cross-mapping in selective_checks.runner_type is left in place because is_disabled_integration still consumes it; the API lookup against the now-removed ci-amd-arm.yml falls back to AMD harmlessly. References to the old file name are updated in: README.md, dev/update_github_branch_config.py, ci-notification.yml, e2e-flaky-tests-report.yml, scripts/ci/analyze_e2e_flaky_tests.py, dev/README_AIRFLOW3_DEV.md, dev/README_RELEASE_AIRFLOW.md, generated/PYPI_README.md, update-constraints-on-push*.yml, release_dockerhub_image.yml. * Add prek hook that asserts ci-arm.yml and ci-amd.yml stay in sync The two CI workflow files are physical copies of each other (GH Actions has no cross-file YAML include) — they should differ only in: - header intro comment - workflow name (ARM vs AMD) - triggers (schedule vs pull_request+push) - concurrency group prefix - build-info platform / runner-type outputs - print-platform job name + echo Drift outside this set is a bug. The new `scripts/ci/prek/check_ci_workflows_in_sync.py` normalizes both files against the documented divergences (regex line rules + arch-only block removals) and asserts the rest matches byte-for-byte. Wired up as a local prek hook gated on either workflow file changing. Local UX: when run interactively, the script also diffs each file against `upstream/main` (falls back to `origin/main`, then `main`) and attributes the drift — "only ci-arm.yml changed since baseline; mirror to ci-amd.yml" with the exact diff to copy, OR "add to LINE_RULES / ARM_ONLY_BLOCK in this script if intentional". Falls back to the non-attribution view in CI / non-TTY runs. Set `FORCE_INTERACTIVE=1` to opt in from a non-TTY shell.

…rkflows (#66348) (#66664) * Split ci-amd-arm.yml into ci-arm.yml (canary) and ci-amd.yml (per-PR) wrappers Replaces the combined ci-amd-arm.yml workflow with two thin wrappers carrying the full pipeline inline: - ci-arm.yml: schedule (canary cron) + workflow_dispatch only; runs on linux/arm64. - ci-amd.yml: pull_request + push (to v3-N-test etc) + workflow_dispatch; runs on linux/amd64. No schedule for now; ARM keeps the canary slot. Each wrapper carries the full pipeline jobs identically except for platform / runner-type / triggers, so per-job UI grouping is preserved on the GitHub Actions page. The runtime cross-mapping in selective_checks.runner_type is left in place because is_disabled_integration still consumes it; the API lookup against the now-removed ci-amd-arm.yml falls back to AMD harmlessly. References to the old file name are updated in: README.md, dev/update_github_branch_config.py, ci-notification.yml, e2e-flaky-tests-report.yml, scripts/ci/analyze_e2e_flaky_tests.py, dev/README_AIRFLOW3_DEV.md, dev/README_RELEASE_AIRFLOW.md, generated/PYPI_README.md, update-constraints-on-push*.yml, release_dockerhub_image.yml. * Add prek hook that asserts ci-arm.yml and ci-amd.yml stay in sync The two CI workflow files are physical copies of each other (GH Actions has no cross-file YAML include) — they should differ only in: - header intro comment - workflow name (ARM vs AMD) - triggers (schedule vs pull_request+push) - concurrency group prefix - build-info platform / runner-type outputs - print-platform job name + echo Drift outside this set is a bug. The new `scripts/ci/prek/check_ci_workflows_in_sync.py` normalizes both files against the documented divergences (regex line rules + arch-only block removals) and asserts the rest matches byte-for-byte. Wired up as a local prek hook gated on either workflow file changing. Local UX: when run interactively, the script also diffs each file against `upstream/main` (falls back to `origin/main`, then `main`) and attributes the drift — "only ci-arm.yml changed since baseline; mirror to ci-amd.yml" with the exact diff to copy, OR "add to LINE_RULES / ARM_ONLY_BLOCK in this script if intentional". Falls back to the non-attribution view in CI / non-TTY runs. Set `FORCE_INTERACTIVE=1` to opt in from a non-TTY shell. (cherry picked from commit 8f93637)

…llision (#66662) (#66665) PR #66348 split the combined ci-amd-arm.yml into per-platform wrappers and intentionally deferred re-adding a schedule trigger to ci-amd.yml ('No schedule for now; ARM keeps the canary slot. Add a schedule: block here if/when AMD should also act as a scheduled canary'). As a result: - The AMD 'Tests (AMD)' badge on the README pointed at a workflow that hadn't run on main since 2025-10-25 (it now only runs on PRs and release-branch pushes), so the badge was stale. - The notify-slack job at the bottom of ci-amd.yml is gated on github.event_name == 'schedule' and was therefore dead code: there was no schedule trigger left to fire it. - main on AMD had no scheduled CI monitor at all (ci-notification.yml only watches v3-2-test x ci-amd.yml, and the comment there assumes ARM is reported by its own notify-slack). This change: 1. Restores the previous AMD canary cron in ci-amd.yml at minute :58 of the same hours ARM uses, so the two scheduled runs don't compete for runners at exactly the same minute. 2. Splits the README build-status row for main into two badges (AMD and ARM), drops the obsolete 2.x row (still using the legacy ci.yml), and keeps 3.x AMD-only (ARM has no v3-2-test schedule). Click-throughs go to the workflow page rather than the catch-all /actions URL. 3. Disambiguates the slack-state artifact names that ci-amd.yml, ci-arm.yml, and ci-image-checks.yml share. Before this PR the latent collision did not bite (AMD's notify-slack never fired); restoring the AMD schedule activates it. Each platform now writes its own slack-state-{tests,inventory}-<branch>-{amd,arm} artifact so the de-dup tracker in scripts/ci/slack_notification_state.py keeps independent state per platform. 4. Updates scripts/ci/prek/check_ci_workflows_in_sync.py to allow the new AMD-only schedule block and the per-platform artifact-name suffix as documented divergences. generated/PYPI_README.md is regenerated automatically by the generate-pypi-readme prek hook to mirror README.md. (cherry picked from commit 022e35d)

boring-cyborg Bot added area:dev-tools backport-to-v3-2-test labels May 4, 2026

potiuk marked this pull request as ready for review May 4, 2026 16:03

potiuk changed the title ~~Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows~~ [Experimental] Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows May 4, 2026

potiuk requested review from amoghrajesh, ashb, bugraoz93, choo121600, ephraimbuddy, gopidesupavan, jason810496, jedcunningham, jscheffl, kaxil, shahar1, vatsrahul1001 and vincbeck as code owners May 4, 2026 16:10

potiuk force-pushed the split-ci-arm-amd-workflows branch from c3a7012 to 32c769e Compare May 4, 2026 16:12

potiuk added 2 commits May 7, 2026 04:32

potiuk force-pushed the split-ci-arm-amd-workflows branch from 32c769e to cb7492c Compare May 7, 2026 02:33

jscheffl approved these changes May 7, 2026

View reviewed changes

potiuk merged commit 8f93637 into apache:main May 7, 2026
143 checks passed

potiuk deleted the split-ci-arm-amd-workflows branch May 7, 2026 23:07

This was referenced May 10, 2026

Restore AMD scheduled canary, split README badges, fix Slack-state collision #66662

Merged

[v3-2-test] Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows (#66348) #66664

Merged

potiuk mentioned this pull request May 10, 2026

[v3-2-test] Restore AMD scheduled canary, split README badges, fix Slack-state collision (#66662) #66665

Merged

2 tasks

FrankYang0529 mentioned this pull request Jul 17, 2026

Fix CI integration tests gated by the wrong platform #70015

Merged

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Experimental] Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows#66348

[Experimental] Split ci-amd-arm.yml into ARM canary and AMD per-PR workflows#66348
potiuk merged 2 commits into
apache:mainfrom
potiuk:split-ci-arm-amd-workflows

potiuk commented May 4, 2026

Uh oh!

jscheffl left a comment

Uh oh!

Uh oh!

github-actions Bot commented May 7, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

potiuk commented May 4, 2026

Summary

Reference updates (every consumer of the old file name)

Test plan

Why draft

Was generative AI tooling used to co-author this PR?

Uh oh!

jscheffl left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot commented May 7, 2026

Backport failed to create: v3-2-test. View the failure log Run details

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants