Clearing a DAG run/task instance with a DagRunQueuedAtDeadline fails with TypeError: unsupported type for timedelta seconds component: dict #70368

Apache Airflow version

3.3.0 and later

What happened

Clearing a failed DAG run or task instance returns HTTP 500 ("Internal Service Error") in the UI when the DAG defines a deadline that references the DAG run's queued_at time (DeadlineReference.DAGRUN_QUEUED_AT). The API server logs:

TypeError: unsupported type for timedelta seconds component: dict

What you think should happen instead

The clear should succeed, and the queued-at deadline should be recalculated against the new queued_at.

Root cause

In 3.3.0, migration 0117_3_3_0_change_deadline_interval_to_json.py (from #64751, which added VariableInterval support) changed the deadline_alert.interval column from FLOAT to JSON. Existing values were rewritten into the serde envelope, e.g. {"__classname__": "datetime.timedelta", "__version__": 2, "__data__": 3600.0}.

_recalculate_dagrun_queued_at_deadlines() in airflow-core/src/airflow/models/taskinstance.py (called from clear_task_instances) still reads the raw column and does timedelta(seconds=deadline_alert.interval), passing the JSON dict straight into timedelta. Every other consumer (DagRun creation in serialization/definitions/dag.py, the model __repr__) decodes the JSON first. Only the clear path was missed.

How to reproduce

1. Define a DAG with a deadline referencing DeadlineReference.DAGRUN_QUEUED_AT.
2. Let a run queue so a Deadline row is created.
3. Clear the DAG run or one of its task instances.

The API server returns 500 with the TypeError above.

Workarounds

• Delete the task instance in the UI instead of clearing it.
• Temporarily remove the deadline parameter from the DAG.

Are you willing to submit a PR?

Yes.