Replacing a failed etcd replica

This document describes how to replace a failed etcd replica in a high availability (HA) user cluster.

The instructions given here apply to an HA user cluster that uses kubeception; that is, a user cluster that does not have Controlplane V2 enabled. If you need to replace an etcd replica in a user cluster that has Controlplane V2 enabled, contact Cloud Customer Care.

Before you begin

Make sure the admin cluster is working correctly.
Make sure the other two etcd members in the user cluster are working correctly. If more than one etcd member has failed, see Recovery from etcd data corruption or loss.

Replacing a failed etcd replica

Back up a copy of the etcd PodDisruptionBudget (PDB) so you can restore it later.
```
kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG -n USER_CLUSTER_NAME get pdb kube-etcd-pdb -o yaml > PATH_TO_PDB_FILE
```
Where:
- ADMIN_CLUSTER_KUBECONFIG is the path to the kubeconfig file for the admin cluster.
- USER_CLUSTER_NAME is the name of the user cluster that contains the failed etcd replica.
- PATH_TO_PDB_FILE is the path where you want to save the etcd PDB file, for instance /tmp/etcpdb.yaml.

Delete the etcd PodDisruptionBudget (PDB).

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG -n USER_CLUSTER_NAME delete pdb kube-etcd-pdb

Run the following command to open the kube-etcd StatefulSet in your text editor:

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG -n USER_CLUSTER_NAME edit statefulset kube-etcd

Change the value of the --initial-cluster-state flag to existing.

containers:
    - name: kube-etcd
      ...
      args:
        - --initial-cluster-state=existing
      ...

Drain the failed etcd replica node.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG drain NODE_NAME --ignore-daemonsets --delete-local-data

Where NODE_NAME is the name of the failed etcd replica node.

Create a new shell in the container of one of the working kube-etcd pods.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG exec -it \
   KUBE_ETCD_POD --container kube-etcd --namespace USER_CLUSTER_NAME \
   -- bin/sh

Where KUBE_ETCD_POD is the name of the working kube-etcd pod. For example, kube-etcd-0.

From this new shell, run the following commands:

Remove the failed etcd replica node from the etcd cluster.

First, list all the members of the etcd cluster:

etcdctl member list -w table

The output shows all the member IDs. Determine the member ID of the failed replica.

Next, remove the failed replica:

export ETCDCTL_CACERT=/etcd.local.config/certificates/etcdCA.crt
export ETCDCTL_CERT=/etcd.local.config/certificates/etcd.crt
export ETCDCTL_CERT=/etcd.local.config/certificates/etcd.crt
export ETCDCTL_KEY=/etcd.local.config/certificates/etcd.key
export ETCDCTL_ENDPOINTS=https://127.0.0.1:2379
etcdctl member remove MEMBER_ID

Where MEMBER_ID is the hex member ID of the failed etcd replica pod.

Follow steps 1-3 of Deploying the utility Pods to create a utility Pod in the admin cluster. This Pod is used to access the PersistentVolume (PV) of the failed etcd member in the user cluster.

Clean up the etcd data directory from within the utility Pod.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG exec -it -n USER_CLUSTER_NAME etcd-utility-MEMBER_NUMBER -- /bin/bash -c 'rm -rf /var/lib/etcd/*'

Delete the utility Pod.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG delete pod -n USER_CLUSTER_NAME etcd-utility-MEMBER_NUMBER

Uncordon the failed node.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG uncordon NODE_NAME

Open the kube-etcd StatefulSet in your text editor.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG -n USER_CLUSTER_NAME edit statefulset kube-etcd

Change the value of the --initial-cluster-state flag to new.

containers:
    - name: kube-etcd
      ...
      args:
        - --initial-cluster-state=new
      ...

Restore the etcd PDB which was deleted in step 1.

kubectl --kubeconfig ADMIN_CLUSTER_KUBECONFIG apply -f /path/to/etcdpdb.yaml