Never force-delete in Kubernetes

The rule

Do not use --force, --grace-period=0, or strip finalizers to delete a Kubernetes object. Use a normal graceful kubectl delete and let termination complete. This applies to every resource — pods, jobs, deployments, StatefulSets, PVCs, namespaces.

Specifically, never do (on your own initiative):

• kubectl delete pod X --force --grace-period=0
• kubectl delete ... --grace-period=0
• kubectl patch ns/pvc ... -p '{"metadata":{"finalizers":[]}}' (or --finalize / editing finalizers to null) to clear a "stuck Terminating" object
• kubectl delete namespace X when it hangs, then force-finalizing it

Why it's dangerous

--force --grace-period=0 deletes the API object without confirming the workload actually stopped. Consequences:

• StatefulSet / single-writer pods: the replacement pod starts while the old container may still be running → two writers at once (split-brain), data corruption, duplicate external side effects. Force-deleting a StatefulSet pod is explicitly unsafe.
• PVCs / volumes: force/finalizer-stripping orphans the real backing volume or leaves it attached, and can lose data.
• Namespaces: clearing finalizers abandons real resources (cloud LBs, disks, external records) that the finalizer existed to clean up — silent leaks and cost.

A "stuck Terminating" object is a symptom, not the problem. Forcing it hides the cause.

Do this instead

1. kubectl describe the object and look at events, deletionTimestamp, and finalizers.
2. Find what's blocking termination: a controller/operator, a volume that won't detach, an admission/finalizer webhook that's down, a node that's NotReady, a pod with a stuck preStop hook.
3. Fix the cause (bring the webhook/controller back, cordon+investigate the node, resolve the volume). Graceful deletion then completes on its own.
4. If a pod is stuck because its node is genuinely gone, delete/​drain the Node object (or let the node controller evict) rather than force-deleting the pod.

The one exception

Force-delete only with explicit human authorization, on a specifically diagnosed resource where the graceful path is confirmed impossible (e.g. a pod on a permanently-dead node with no StatefulSet single-writer risk). Never as routine, never as a shortcut to "make it go away," never on a StatefulSet pod without confirming the old container is truly dead. State why in the moment.