Fix airflow db clean hang on MySQL when delete fails

[Vamsi-klu]

Fixes a MySQL hang in airflow db clean --skip-archive when a delete fails on a foreign-key constraint.

On the failure path, _do_delete() now rolls back the session before dropping the temporary archive table, and the drop uses the session's active connection instead of checking out a new pooled connection. This prevents metadata-lock contention from leaving the cleanup command stuck indefinitely.

Also adds:

• unit coverage for rollback-before-drop ordering
• a MySQL backend regression test for the deadlock case
• a Postgres backend regression test for the failure path

Validation:

• prek run --from-ref main --stage pre-commit
• prek run --from-ref main --stage manual
• targeted MySQL and Postgres regression tests in Breeze
• full airflow-core/tests/unit/utils/test_db_cleanup.py in Breeze (57 passed, 1 skipped)
• breeze ci selective-check --commit-ref HEAD indicates the shared test utility change expands the CI matrix beyond what was practical to run locally

closes: #66177

---

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

• Yes — Codex; Claude Code (Fable 5)

Generated-by: Codex; Claude Code (Fable 5) following the guidelines

---

Important

🛠️ Maintainer triage note for @Vamsi-klu · by @potiuk · 2026-07-02 17:46 UTC

Helpful heads-up from the maintainers — please address before this PR can be reviewed (see the Pull Request quality criteria):

• This PR has merge conflicts with main. Please rebase onto the latest main and resolve them.

The ball is in your court — you've been assigned to this PR. Fix the above, then mark it Ready for review.

_{Automated triage — may be imperfect; a maintainer takes the next look.}

[csrinivas2130]

@Vamsi-klu can you close your PR, I have made the fix 4 days ago and waiting for the #PR-66249 approval for merge and close this issue.

[potiuk]

@Vamsi-klu A few things need addressing before review — see our Pull Request quality criteria.

• ❌ CI fails: Low dep tests: providers / All-prov:LowestDeps:14:3.10:airbyte...apprise (and possibly other checks — see the Checks tab for the full list).

No rush.

---

Note: This comment was drafted by an AI-assisted triage tool and may contain mistakes. Once you have addressed the points above, an Apache Airflow maintainer — a real person — will take the next look at your PR. We use this two-stage triage process so that our maintainers' limited time is spent where it matters most: the conversation with you.

---

Drafted-by: Claude Code (Opus 4.7); reviewed by @potiuk before posting

[Vamsi-klu]

Gentle bump on this PR. CI is green now and the PR is labeled ready for maintainer review. Could a maintainer please take another look when available and let me know if anything else is needed?

Thanks.

---

Drafted-by: Codex (GPT-5); reviewed by @Vamsi-klu before posting

[ephraimbuddy]

I think this will solve the hang but the cleanup would still raise on every dag with run history since dagversion is fk-pinned to task_instance.

The newsfragments is not necessary.

[Vamsi-klu]

Addressed the review and rebased on main:

• session.rollback() is now wrapped in suppress(Exception), with a unit test pinning that the original delete error propagates when rollback itself raises.
• Added a code comment explaining why the archive-table drop runs on the session's own connection.
• Newsfragment removed.

On dag_version being FK-pinned by task_instance: this PR only converts the hang into a fast, visible failure — #68339 is the right place for skipping pinned rows. To keep the two PRs composable, the two backend regression tests here now disable row-exclusion config on a copied _TableConfig dict, so they keep forcing the FK-failure path that pins the hang regression even after #68339 lands.

---

Drafted with Claude Code; reviewed by @Vamsi-klu before posting

[Vamsi-klu]

#66249 was closed on 2026-05-07 without being merged, and #66177 is still open, so this PR remains the active fix for the hang. It has since been through maintainer review (see @ephraimbuddy's comments above) and CI is green. Happy to coordinate if you plan to reopen yours, but as it stands there is no overlapping open PR.

---

Drafted with Claude Code; reviewed by @Vamsi-klu before posting

[Vamsi-klu]

Hi @ephraimbuddy @potiuk , could you take a maintainer pass on this PR when you have a moment?

The PR is cleanly rebased, and I finished the merge-conflict cleanup. It now preserves the behavior fix and includes a regression test for the session.rollback()/error-propagation path in db_cleanup.

[aaron-y-chen]

Hi, thanks for the fix! While testing this, I found a small edge case where the cleanup path can still mask the original delete error. Would you mind fixing it in this PR as well? :)

[Vamsi-klu]

Good catch, thanks! Fixed: the finally drop/commit is now wrapped so that when we're unwinding from a delete failure, a cleanup error is logged as a warning and the original delete error keeps propagating (Python otherwise makes the finally-raised error the top-level one). On the success path a drop/commit failure still surfaces normally.

Added test_do_delete_original_error_survives_archive_drop_failure (failure path — original IntegrityError propagates even when the drop raises OperationalError) and test_do_delete_success_propagates_archive_drop_error (success path — drop error is not swallowed). No backend test since a real DROP failure can't be forced deterministically; the masking logic is fully covered by the mock tests.

---

Drafted-by: Claude Code (Opus 4.8); reviewed by @Vamsi-klu before posting

[ephraimbuddy]

We shouldn't have PR numbers on tests doc strings

[Vamsi-klu]

Hi @ephraimbuddy , Addressed the review comment regarding the removal of PR numbers from test docstrings. The documentation has been updated to be cleaner and maintain standard practices for the codebase.