Fix zombie task handling with multiple schedulers (apache#24906)

Each scheduler was looking at all running tasks for zombies, leading to
multiple schedulers handling the zombies. This causes problems with
retries (e.g. being marked as FAILED instead of UP_FOR_RETRY) and
callbacks (e.g. `on_failure_callback` being called multiple times).

When the second scheduler tries to determine if the task is able to be retried,
and it's already in UP_FOR_RETRY (the first scheduler already finished),
it sees the "next" try_number (as it's no longer running),
which then leads it to be FAILED instead.

The easy fix is to simply restrict each scheduler to its own TIs, as
orphaned running TIs will be adopted anyways.

(cherry picked from commit 1c0d0a5)

airflow/jobs/scheduler_job.py

@@ -1339,7 +1339,8 @@ def check_trigger_timeouts(self, session: Session = None):
     def _find_zombies(self, session):
         """
         Find zombie task instances, which are tasks haven't heartbeated for too long
-        and update the current zombie list.
+        or have a no-longer-running LocalTaskJob, and send them off to the DAG processor
+        to be handled.
         """
         self.log.debug("Finding 'running' jobs without a recent heartbeat")
         limit_dttm = timezone.utcnow() - timedelta(seconds=self._zombie_threshold_secs)
@@ -1355,6 +1356,7 @@ def _find_zombies(self, session):
                     LocalTaskJob.latest_heartbeat < limit_dttm,
                 )
             )
+            .filter(TaskInstance.queued_by_job_id == self.id)
             .all()
         )
 

tests/jobs/test_scheduler_job.py

@@ -3856,7 +3856,6 @@ def test_find_zombies(self):
             session.query(LocalTaskJob).delete()
             dag = dagbag.get_dag('example_branch_operator')
             dag.sync_to_db()
-            task = dag.get_task(task_id='run_this_first')
 
             dag_run = dag.create_dagrun(
                 state=DagRunState.RUNNING,
@@ -3865,21 +3864,33 @@ def test_find_zombies(self):
                 session=session,
             )
 
-            ti = TaskInstance(task, run_id=dag_run.run_id, state=State.RUNNING)
-            local_job = LocalTaskJob(ti)
-            local_job.state = State.SHUTDOWN
-
-            session.add(local_job)
-            session.flush()
-
-            ti.job_id = local_job.id
-            session.add(ti)
-            session.flush()
-
             self.scheduler_job = SchedulerJob(subdir=os.devnull)
             self.scheduler_job.executor = MockExecutor()
             self.scheduler_job.processor_agent = mock.MagicMock()
 
+            # We will provision 2 tasks so we can check we only find zombies from this scheduler
+            tasks_to_setup = ['branching', 'run_this_first']
+
+            for task_id in tasks_to_setup:
+                task = dag.get_task(task_id=task_id)
+                ti = TaskInstance(task, run_id=dag_run.run_id, state=State.RUNNING)
+                ti.queued_by_job_id = 999
+
+                local_job = LocalTaskJob(ti)
+                local_job.state = State.SHUTDOWN
+
+                session.add(local_job)
+                session.flush()
+
+                ti.job_id = local_job.id
+                session.add(ti)
+                session.flush()
+
+            assert task.task_id == 'run_this_first'  # Make sure we have the task/ti we expect
+
+            ti.queued_by_job_id = self.scheduler_job.id
+            session.flush()
+
             self.scheduler_job._find_zombies(session=session)
 
             self.scheduler_job.executor.callback_sink.send.assert_called_once()