GitHub user arkadiuszbach added a comment to the discussion: Unexpected SIGTERM on Tasks (Airflow 2.10.5 on GKE with KubernetesExecutor and No Resource Constraints)
>From the [GKE >docs](https://docs.cloud.google.com/kubernetes-engine/docs/best-practices/upgrading-clusters#graceful-termination), > with the default surge-upgrade strategy, a pod has 1 hour to terminate >gracefully during node upgrades — after that it gets evicted, so maybe that is >the reason? The cluster autoscaler could also evict the pod, but if safe-to-evict is false (that is the default in the helm chart) then the autoscaler won't remove the node at all. However, if safe-to-evict is not false, the autoscaler has its own graceful termination timeout that overrides the pod's terminationGracePeriodSeconds — and on GKE [this is not configurable](https://docs.cloud.google.com/kubernetes-engine/docs/concepts/cluster-autoscaler): it's 10 minutes for GKE versions before 1.32.7, and 1 hour for GKE 1.32.7+. So even with a large terminationGracePeriodSeconds, the autoscaler will force-terminate the pod after that timeout. GitHub link: https://github.com/apache/airflow/discussions/62978#discussioncomment-17155595 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
