when cluster members are added or removed, we should update the state file

[mmerrill3]

The state file used to initialize the etcd server should reflect when etcd members come and go. Currently, we update /etc/hosts when this information changes, but we do not update the state file under the base data directory for etcd-main or etcd-events. I also propose updating the state.Nodes information for the running etcd service to reflect the changes.
This manifests as bad node data being returned by the GetInfo peers service, since the state of the cluster comes from that state file, and it doesn't get updated once its read. If you decide to expand your cluster, and then shrink it (for say, moving to a new subnet), references to the old etcd members will still be seen in the logs when the "master" etcd server calls GetInfo on the other peers.

[adammw]

We're noticing this issue on rolling restarts as well, where we see both old and new IPs listed in the hosts update messages.