Does fancing work with control-plane nodes too?

[tumurzakov]

I manage bare metal cluster and suffer from downtime when nodes failing

1. Worker node failed - status become NotReady but traffic still come to the pods that was on this node. Often this leads to downtime if there was critical components. Now I need to delete this node from cluster to reshedule pods. Nothing special for your project.
2. Control plane node failed - that was only one time. I have 3 control plane nodes. After one node failed i immediatly deleted it from cluster. But nothing works because kubedns stoped resolving internal names. As I understand this happend because remaing etcd daemons didn't reach an agreement who become a master. Only systemctl restart docker on one of two working control plane nodes resolves an issue. So, could kube-fencing be used on control-plane nodes as well?

[kvaps]

Do you have coredns pods running on single node?

[tumurzakov]

No,

# kubectl -n kube-system get pods -o wide|grep coredns
coredns-758cc77499-crbmv                1/1     Running   0          22d     10.244.0.4       kube-master-3   <none>           <none>
coredns-758cc77499-z2flm                1/1     Running   0          40d     10.244.7.16      kube-master-2   <none>           <none>

I have 3 master nodes, but coredns deployment contain 2 replicas. Something wrong in configuration?