JoelEmmanuelCloud commented on code in PR #76410:
URL: https://github.com/apache/airflow/pull/76410#discussion_r4644216861
##########
airflow-core/src/airflow/jobs/scheduler_job_runner.py:
##########
@@ -435,6 +454,44 @@ def __init__(
self.executors: list[BaseExecutor] = executors if executors else
ExecutorLoader.init_executors()
self.executor: BaseExecutor = self.executors[0]
+ # Each executor's get_health() runs on its own daemon thread, never on
the scheduler loop's own
+ # thread: a check that talks to a broker or a cluster control plane
can be slow or hang, and
+ # must not be able to delay task scheduling. A pool isn't needed since
each executor is only
+ # ever checked by one in-flight call at a time (see
_check_executor_health); a daemon thread
+ # also means a hung check can't block process exit the way a
ThreadPoolExecutor worker would.
+ self._executor_health_check_futures: dict[str, tuple[Future, float]] =
{}
+ # A probe that ignores the timeout passed to it would otherwise occupy
its slot forever: once
+ # a pending future has been sitting this many intervals, it's reported
as stalled instead of
+ # silently skipped on every later tick (see _check_executor_health).
+ self._executor_health_check_stall_after_intervals = 3
+ self._stalled_health_checks: set[str] = set()
+
+ self._executor_health_check_interval = conf.getfloat(
+ "scheduler", "executor_health_check_interval", fallback=30.0
+ )
+ # Must be comfortably below the check interval so a probe that honours
it always finishes
+ # before the next tick fires; half the interval is the starting point.
+ self._executor_health_check_timeout =
self._executor_health_check_interval / 2
+
+ # Only executors that actually override get_health() get a Job row at
all: one that was never
+ # going to report anything isn't a gap in coverage (see
_configured_executor_keys in
+ # airflow_health.py, which applies this same rule from the
/monitor/health side). An executor
+ # with no name is also skipped: _executor_health_key has nothing
stable to key its row under.
+ # issubclass(type(executor), ...), not isinstance(executor, ...): a
test double standing in
+ # for an executor (a MagicMock, even one speced to BaseExecutor)
passes isinstance, since mock
+ # fakes that check, but calling a method on it never runs the real
classmethod body, it just
+ # hands back another mock, which this dict would then read as
"overrides it". type() bypasses
+ # that trick and correctly reports such a double as not a real
BaseExecutor subclass.
+ from airflow.executors.base_executor import BaseExecutor as
_BaseExecutor
+
+ self._executor_health_keys: dict[str, str | None] = {
+ key: executor.team_name
+ for executor in self.executors
+ if issubclass(type(executor), _BaseExecutor)
Review Comment:
you're right, done. the dict now just calls executor.overrides_get_health()
directly, and the mocks in tests set overrides_get_health.return_value = False.
##########
airflow-core/src/airflow/jobs/scheduler_job_runner.py:
##########
@@ -3912,6 +4190,23 @@ def adopt_or_reset_orphaned_tasks(self, *, session:
Session = NEW_SESSION) -> in
self.log.info("Marked %d SchedulerJob instances as
failed", num_failed)
stats.incr(self.__class__.__name__.lower() + "_end",
num_failed)
+ # A separate statement, not folded into the SchedulerJob
one above, so the
+ # "Marked %d SchedulerJob instances as failed" log line
and the _end metric keep
+ # counting only schedulers. These ExecutorJob rows are
left running with no
+ # end_date when their owning scheduler is SIGKILLed or
OOM-killed, since _execute's
+ # own cleanup (_end_executor_jobs) never gets to run for
that case; this also
+ # catches a row a health check still in flight at that
moment recreates after
+ # _end_executor_jobs has already closed the others.
+ session.execute(
Review Comment:
added one, test_adopt_or_reset_orphaned_tasks_fails_stale_executor_jobs.
##########
airflow-core/src/airflow/jobs/scheduler_job_runner.py:
##########
@@ -435,6 +454,44 @@ def __init__(
self.executors: list[BaseExecutor] = executors if executors else
ExecutorLoader.init_executors()
self.executor: BaseExecutor = self.executors[0]
+ # Each executor's get_health() runs on its own daemon thread, never on
the scheduler loop's own
Review Comment:
trimmed them down.
##########
airflow-core/src/airflow/api/common/airflow_health.py:
##########
@@ -54,13 +58,15 @@ class DetailedHealthStatus(str, Enum):
@provide_session
-def get_jobs_health(job_runner_class, *, session: Session = NEW_SESSION) ->
list[Job]:
+def get_jobs_health(job_runner_class: type | str, *, session: Session =
NEW_SESSION) -> list[Job]:
"""
- Return live unfinished jobs and the latest unfinished job, ordered by
heartbeat.
+ Return live unfinished jobs and the latest unfinished job of the given job
type, ordered by heartbeat.
- Keep the latest job even when dead so health responses retain its
heartbeat.
+ Keep the latest job even when dead so health responses retain its
heartbeat. ``job_runner_class`` is
+ a job runner class that reports its own ``job_type`` (e.g.
``SchedulerJobRunner``), or a bare
+ job_type string for a job with no runner class of its own, like the
executor's ``EXECUTOR_JOB_TYPE``.
"""
- job_type = job_runner_class.job_type
+ job_type = getattr(job_runner_class, "job_type", job_runner_class)
Review Comment:
done, simpler this way.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]