Pods of a Deployment (or StatefulSet) can be terminated during normal operation in Kubernetes. When a pod is terminated, Kubernetes executes a well-defined sequence of steps. Most notably, Kubernetes calls a preStop-hook in each container of the Pod and subsequently sends SIGTERM to the process with PID 1 in each container. Applications should ensure that they can deal with being terminated at any time by implementing custom preStop-hooks and reacting to SIGTERM with application specificshutdown logic. A sudden termination should never result in an inconsistent application state. If the shutdown procedure of an application takes exceptionally long, the application should specify its own terminationGracePeriodSeconds to ensure that its Pods receive enough time to terminate cleanly.
Here are some examples:
- When your application needs to guarantee that it always writes certain log lines when one of its replicas is terminated.
- When your application needs to perform some database operation (commit, rollback) on termination of one of its replicas.
- When your application's replicas need to write some persistent data when terminating.
- When your application's replicas need to make an announcement to their peers (e.g. in a database cluster implemented as a StatefulSet) when being terminated.
A Deployment which can reacts to SIGTERM and implememts a preStop-hook. Create it using the following command:
kubectl apply -f examples/Tail the logs of the associated pod using the following command
kubectl logs $(kubectl get pods --selector=app=graceful-shutdown -o jsonpath='{.items[0].metadata.name}') -fOpen a new shell and delete the deployment using the following command, observing the log lines concerning SIGTERM and the preStop-hook:
kubectl delete -f examples/