What will happen to evicted pods in kubernetes?

I just saw some of my pods got evicted by kubernetes. What will happen to them? just hanging around like that or I have to delete them manually?

A quick workaround I use, is to delete all evicted pods manually after an incident. You can use this command:

kubectl get pods --all-namespaces -o json | jq '.items[] | select(.status.reason!=null) | select(.status.reason | contains("Evicted")) | "kubectl delete pods \(.metadata.name) -n \(.metadata.namespace)"' | xargs -n 1 bash -c

To delete pods in Failed state in namespace default

kubectl -n default delete pods --field-selector=status.phase=Failed

Evicted pods should be manually deleted. You can use following command to delete all pods in Error state.

kubectl get pods --all-namespaces --field-selector 'status.phase==Failed' -o json | kubectl delete -f -

Depending on if a soft or hard eviction threshold that has been met, the Containers in the Pod will be terminated with or without grace period, the PodPhase will be marked as Failed and the Pod deleted. If your Application runs as part of e.g. a Deployment, there will be another Pod created and scheduled by Kubernetes - probably on another Node not exceeding its eviction thresholds.

Be aware that eviction does not necessarily have to be caused by thresholds but can also be invoked via kubectl drain to empty a node or manually via the Kubernetes API.

The bellow command delete all failed pods from all namespaces

kubectl get pods -A | grep Evicted | awk '{print $2 " -n " $1}' | xargs -n 3 kubectl delete pod

To answer the original question: the evicted pods will hang around until the number of them reaches the terminated-pod-gc-threshold limit (it's an option of kube-controller-manager and is equal to 12500 by default), it's by design behavior of Kubernetes (also the same approach is used and documented for Jobs - https://kubernetes.io/docs/concepts/workloads/controllers/job/#job-termination-and-cleanup). Keeping the evicted pods pods around allows you to view the logs of those pods to check for errors, warnings, or other diagnostic output.

One more bash command to delete evicted pods

kubectl get pods | grep Evicted | awk '{print $1}' | xargs kubectl delete pod

Just in the case someone wants to automatically delete all evicted pods for all namespaces:

• Powershell

    Foreach( $x in (kubectl get po --all-namespaces --field-selector=status.phase=Failed --no-headers -o custom-columns=:metadata.name)) {kubectl delete po $x --all-namespaces }

• Bash

kubectl get po --all-namespaces --field-selector=status.phase=Failed --no-headers -o custom-columns=:metadata.name | xargs kubectl delete po --all-namespaces

Kube-controller-manager exists by default with a working K8s installation. It appears that the default is a max of 12500 terminated pods before GC kicks in.

Directly from the K8s documentation: https://kubernetes.io/docs/reference/command-line-tools-reference/kube-controller-manager/#kube-controller-manager

> --terminated-pod-gc-threshold int32     Default: 12500
> Number of terminated pods that can exist before the terminated pod garbage collector starts deleting terminated pods. If <= 0, the terminated pod garbage collector is disabled.

In case you have pods with a Completed status that you want to keep around:

kubectl get pods --all-namespaces --field-selector 'status.phase==Failed' -o json | kubectl delete -f -

Another way still with awk.

To prevent any human error that could make me crazy (deleting desirable pods), I check before the result of the get pods command :

kubectl -n my-ns get pods --no-headers --field-selector=status.phase=Failed     

If that looks good, here we go :

kubectl -n my-ns get pods --no-headers --field-selector=status.phase=Failed | \
awk '{system("kubectl -n my-ns delete pods " $1)}'

Same thing with pods of all namespaces.

Check :

kubectl get -A pods --no-headers --field-selector=status.phase=Failed     

Delete :

kubectl get -A pods --no-headers --field-selector status.phase=Failed | \
awk '{system("kubectl -n " $1 " delete pod " $2 )}'

OpenShift equivalent of Kalvin's command to delete all 'Evicted' pods:

eval "$(oc get pods --all-namespaces -o json | jq -r '.items[] | select(.status.phase == "Failed" and .status.reason == "Evicted") | "oc delete pod --namespace " + .metadata.namespace + " " + .metadata.name')"

To delete all the Evicted pods by force, you can try this one-line command:

$ kubectl get pod -A | sed -nr '/Evicted/s/(^\S+)\s+(\S+).*/kubectl -n \1 delete pod \2 --force --grace-period=0/e'

Tips: use the p modifier of s command of sed instead of e will just print the real command to do the deletion job:

$ kubectl get pod -A | sed -nr '/Evicted/s/(^\S+)\s+(\S+).*/kubectl -n \1 delete pod \2 --force --grace-period=0/p'

below command will get all evicted pods from the default namespace and delete them

kubectl get pods | grep Evicted | awk '{print$1}' | xargs -I {} kubectl delete pods/{}

Here is the 'official' guide for how to hard code the threshold(if you do not want to see too many evicted pods): kube-controll-manager

But a known problem is how to have kube-controll-manager installed...