amoghrajesh commented on code in PR #67715:
URL: https://github.com/apache/airflow/pull/67715#discussion_r3338768835
##########
providers/apache/spark/src/airflow/providers/apache/spark/hooks/spark_submit.py:
##########
@@ -802,6 +861,95 @@ def _start_driver_status_tracking(self) -> None:
f"returncode = {returncode}"
)
+ def _poll_k8s_driver_via_api(self) -> None:
+ """Poll the K8s driver pod phase until it reaches a terminal state."""
+ pod_name = self._kubernetes_driver_pod
+ namespace = self._connection["namespace"]
+ app_id = self._kubernetes_application_id or pod_name
+
+ if not pod_name:
+ raise ValueError("K8s driver pod name not set; cannot poll
status.")
+
+ client = kube_client.get_kube_client()
+ poll_interval = max(self._status_poll_interval, 20)
+ if poll_interval != self._status_poll_interval:
+ self.log.info(
+ "status_poll_interval=%ds is below the 20s minimum for K8s API
polling; using 20s.",
+ self._status_poll_interval,
+ )
+ # Mirror `missed_job_status_reports` / `max_missed_job_status_reports`
from
+ # `_start_driver_status_tracking`: tolerate transient failures before
giving up.
+ consecutive_unknown = 0
+ max_consecutive_unknown = 3
+ consecutive_api_errors = 0
+ max_consecutive_api_errors = 3
+ consecutive_pending = 0
+ pending_warn_threshold = 10
+
+ try:
+ while True:
+ try:
+ pod = client.read_namespaced_pod(pod_name, namespace)
+ consecutive_api_errors = 0
+ except kube_client.ApiException as e:
+ consecutive_api_errors += 1
+ self.log.warning(
+ "ApiException polling pod %s (%d/%d): %s",
+ pod_name,
+ consecutive_api_errors,
+ max_consecutive_api_errors,
+ e,
+ )
+ if consecutive_api_errors >= max_consecutive_api_errors:
+ raise RuntimeError(
+ f"K8s API unreachable after
{consecutive_api_errors} consecutive errors "
+ f"while polling {app_id}; giving up."
+ ) from e
+ time.sleep(poll_interval)
+ continue
+
+ phase = pod.status.phase or "Initializing"
+ self.log.info("Application status for %s (phase: %s)", app_id,
phase)
+ if phase == "Succeeded":
+ break
+ if phase == "Failed":
+ container_state = ""
+ if pod.status.container_statuses:
+ cs = pod.status.container_statuses[0]
+ if cs.state and cs.state.terminated:
+ container_state = f"
exit_code={cs.state.terminated.exit_code} reason={cs.state.terminated.reason}"
+ raise RuntimeError(f"Spark application {app_id} failed
(phase=Failed{container_state})")
+ if phase == "Pending":
+ consecutive_pending += 1
+ if consecutive_pending == pending_warn_threshold:
+ self.log.warning(
+ "Driver pod %s has been Pending for %d polls
(~%ds); "
+ "it may be unschedulable. Continuing to wait — set
execution_timeout to bound wait time.",
+ pod_name,
+ consecutive_pending,
+ consecutive_pending * poll_interval,
+ )
+ else:
+ consecutive_pending = 0
+
+ if phase == "Unknown":
+ consecutive_unknown += 1
+ if consecutive_unknown >= max_consecutive_unknown:
+ raise RuntimeError(
+ f"Spark application {app_id} reported Unknown
phase "
+ f"{consecutive_unknown} times consecutively;
giving up."
+ )
+ else:
+ consecutive_unknown = 0
+ time.sleep(poll_interval)
+ try:
+ client.delete_namespaced_pod(pod_name, namespace)
+ self.log.info("Deleted driver pod %s", pod_name)
+ except kube_client.ApiException:
+ self.log.warning("Could not delete driver pod %s after
completion", pod_name)
+ finally:
+ self._run_post_submit_commands()
Review Comment:
Fixed by adding and `not self._should_track_driver_via_k8s_api()` to the
submit() finally guard, so post-submit commands run only once now. Also added a
test which uses assert_called_once() on a patched _run_post_submit_commands to
lock the fix in.
--
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]
