In the case that the BlockingJoinNode runs the build asynchrously in one
fragment instance but not another, a deadlock between the instances is
possible (see IMPALA-1863 for the details). To avoid this deadlock
potential, close the bulid side child on error, which will deregister
any datastream receivers from that side, breaking the cycle that leads
to the deadlock.
Change-Id: I2de06615897b4bcaa5855449a98984f11c948dc4
Reviewed-on: http://gerrit.cloudera.org:8080/242
Reviewed-by: Dan Hecht <dhecht@cloudera.com>
Tested-by: Internal Jenkins
Conflicts:
be/src/exec/blocking-join-node.cc