Raghu Angadi wrote:
As Doug mentioned does will HADOOP-2346 do for now? In fact write
timeout can be made configurable.
I fail to see how HADOOP-2346, as is, does anything for us.
Studying a loaded-up instance of regionserver where we have lots of open
MapFiles, sockets seem overwhelmingly to be in the CLOSE_WAIT state.
IIRC, the datanode has shut its side of the connection and client is
oblivious of the fact -- at least until it goes about its next read.
So, HADOOP-2346 adds nothing to help w/ the clean up of client side
sockets in CLOSE_WAIT. The remainder are usually in the ESTABLISHED
state. If code in HADOOP-2346 runs, IIUC, all that will happen is that
the ESTABLISHED connections will be converted to CLOSE_WAITs.
Thanks,
St.Ack
