The GitHub Actions job "Tests" on airflow.git/v3-1-test has failed.
Run started by GitHub user jscheffl (triggered by jscheffl).
Head commit for run:
da4a5eaf8d6c06d433da6792ad5fea96ab60f367 / github-actions[bot]
<41898282+github-actions[bot]@users.noreply.github.com>
[v3-1-test] Fix: TriggerDagRunOperator stuck in deferred state with
reset_dag_run (#57756) (#57968) (#58333)
When TriggerDagRunOperator is used with deferrable=True,
wait_for_completion=True,
reset_dag_run=True, and a fixed trigger_run_id, the operator becomes
permanently
stuck in deferred state after clearing and re-running.
Root cause:
When reset_dag_run=True is used with a fixed run_id, the database preserves
the
original logical_date from the first run. However, on subsequent runs after
clearing,
the operator calculates a NEW logical_date based on the current time. The
DagStateTrigger
was being created with this newly calculated logical_date, causing a mismatch
when
querying the database - the trigger looked for a DAG run with the new
logical_date
but the database contained the original logical_date, causing the query to
return
zero results indefinitely.
Solution:
- Modified _handle_trigger_dag_run() in task_runner.py to pass
execution_dates=None
to DagStateTrigger when run_ids is provided, since run_id alone is
sufficient and
globally unique
- Added test test_handle_trigger_dag_run_deferred_with_reset_uses_run_id_only
to
verify the fix and prevent regression
The fix ensures that both deferrable and non-deferrable modes use identical
logic
for determining DAG run completion - querying by run_id and state only,
without
filtering by logical_date which can become stale when resets are involved.
(cherry picked from commit 4f3d0c55beab8d999c5526e36065f25809afd13f)
Co-authored-by: Mykola Shyshov <[email protected]>
Report URL: https://github.com/apache/airflow/actions/runs/19378592657
With regards,
GitHub Actions via GitBox
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
