During etcd restore, we had problems because the keepalived VIP was assigned to a non-recovery control plane node. In order to be safe, we should ensure that this never happens.

The way to fix would be:
- On step 4 (steps to be run in non-recovery control plane nodes), add the following sub-steps after substep 4f
  - If "/etc/kubernetes/manifests/keepalived.yaml" file exists, move it way with this command: "mv /etc/kubernetes/manifests/keepalived.yaml /root"
  - Wait until there is no running keepalived container as per this command output "crictl ps --name keepalived"
  - Check that the control plane has no VIP assigned as per "ip -o | grep -E '<apiVIP>|<ingressVIP>". For each VIP reported by the command above, run "ip address del <reportedWrongVIP> dev <deviceOfTheWrongReportedVIP>"

- After step 5, add a step to double-check that the recovery control plane node owns the VIP by running "ip -o | grep -E '<apiVIP>'"

Note that we don't have to start the keepalived static pods again because the non-recovery control plane nodes will be replaced and the replacement nodes will have the static pods in place.

All currently supported versions

Sometimes

1. Read docs
2.
3.

Wrong docs

Right docs

The exact problem we found was reported at https://issues.redhat.com/browse/OCPBUGS-18940 . It is still worth that they fix the keepalived to take machine-config-server health into consideration, but we also should not allow keepalived proxy to stay running in nodes that are going to be no longer part of the cluster.