As of mirror pods serve as a "mirror" of the state of static pods on a node. If a static pod gets evicted, the mirror pod will reflect this state change, but is never deleted.
When kubelet restarts it restarts previously evicted static pods and so the old mirror pod is no longer valid.

We need to either delete mirror pods once their respective static pods have been completely removed from the node, or
have kubelet avoid restarting static pods based on the presence of corresponding mirror pods.

The latter would require kubelet to wait for access to API server which will affect bootstrapping scenarios (bootstrapping a master node using static pods)

The former might be a better alternative for now until Kubelet can avoid having multiple simultaneous sources of pods.

cc @dashpole @kubernetes/sig-node-bugs