Watch depends on long running connections, which intervening proxies
may break without the control of the remote server. Specific errors
handled are io.EOF, io.EOF wrapped by *url.Error, http connection
reset errors (caused by race conditions in golang http code), and
connection reset by peer (simply tolerated).
Implemented via HTTP and websocket. A test is present but this isn't
yet wired into anything.
Eventual purpose of this is to allow a scheduler to watch for new pods.
Or allow replication controller to watch for new items it controlls.
Generally, it'll be good to turn everything possible into a push instead
of a poll.
