Ceph » Messengers

Peers A and B:

- A reached epoch 10, in which A and B should exchange heartbeats
 - A advances to epoch 20, and closes pipe to B
 - B reaches epoch 10, opens connection to A
 - A accepts connection
 - B reaches epoch 20, closes it's end of connection to A
 -> A still has connection state with B, but won't mark it down (again)

I think the solution is for incoming connections to start in 'server' mode, where a failure closes out the connection, unless the receiving end explicitly does a mark_lossless(). This should map cleanly onto the heartbeat exchange code. Need to look more closely at how the connection policy is applied..