Since the buildbot workers run inside k8s pods, they get killed when a maintenance window starts. They don't necessarily restart correctly and need to be manually restarted again to reconnect to the buildbot master.
This should be fixed so that it is automated and no intervention is needed from the oncall.