Files
impala/be
Sailesh Mukil ff629c2deb IMPALA-3875: Thrift threaded server hang in some cases
We use socket timeouts for our backend connections. We set these
timeouts only after we've open()'d the connection, which ideally
should be fine. However, our TSaslTransport stack does read()'s and
write()'s over the network on an open(), which means that on a secure
cluster we send and recieve non-TCP-handshake packets on open(). This
is because the current code tries to establish a SASL handshake during
open().

If for any reason the peer server does not respond to the read()'s
during the open() call (after connect() is successful), the client
will wait on read() indefinitely. This patch sets the socket timeout
before we call open(), so that the read()'s and write()'s during the
open() are subject to the timeout as well.

We should also consider making a larger change where this SASL
handshake does not take place during an open(), but instead after the
open() call is completed, so as to have the open() semantics be the
same for both secure and insecure clusters.

Change-Id: I6c8f91a88f723e0e58e81bb385c5a8f190021868
Reviewed-on: http://gerrit.cloudera.org:8080/5263
Reviewed-by: Sailesh Mukil <sailesh@cloudera.com>
Tested-by: Internal Jenkins
2016-12-03 23:21:10 +00:00
..