This bug contributed to the crash discussed in
HADOOP-572.
ipc.Client is trying to establish connection with its server with an infinite timeout.
For an unknown to me reason infinity equals 3 minutes in this case.
I guess it is configured somewhere in the native socket implementation.
With this timeout data-nodes had only 3 chances to send heartbeats during the 10
minute expiration interval. And with a very busy name-node this makes their
chances to be accepted close to 0.
I included an explicit call of Socket.connect() with a timeout set to 1 min, which is
our default for all connections.
Modified a log message to include information that turned out to be useful for debugging.
Removed unnecessary imports.
Bug
Closed
Major
Explicit timeout for ipc.Client
