Ceph

We currently have a paxos instance per state machine, which is silly for a bunch of reasons. The big one is that a monitor might be up to date on, say, the osdmap, but out of date on the pgmap, which leads to weird availability states.

Also, when a monitor rejoins and is far behind, it tries to catch up immediately. We need a 'recovering' state during which the monitor can stream data that it is behind on, and then try to join the quorum.