Joe McDonnell has uploaded this change for review. ( http://gerrit.cloudera.org:8080/23264
Change subject: PROTOTYPE: IMPALA-14271: Reapply the core piece of IMPALA-6984 ...................................................................... PROTOTYPE: IMPALA-14271: Reapply the core piece of IMPALA-6984 IMPALA-6984 changed the behavior to cancel backends when the query reaches the RETURNED_RESULTS state. This ran into a regression on large clusters where a query would end up waiting 10 seconds. IMPALA-10047 reverted the core piece of the change. For tuple caching, we found that a scan node can get stuck waiting for a global runtime filter. It turns out that the coordinator will not send out global runtime filters if the query is in a terminal state. Tuple caching was causing queries to reach the RETURNED_RESULTS phase before the runtime filter could be sent out. Reenabling the core part of IMPALA-6984 sends out a cancel as soon as the query transitions to RETURNED_RESULTS and wakes up any fragment instances waiting on runtime filters. The underlying cause of IMPALA-10047 is a tangle of locks that causes us to exhaust the RPC threads. The coordinator is holding a lock on the backend state while it sends the cancel synchronously. Other backends that complete during that time run Coordinator::BackendState::LogFirstInProgress(), which iterates through backend states to find the first that is not done. The check to see if a backend state is done takes a lock on the backend state. The problem case is that the coordinator may be sending a cancel to a backend on itself. In that case, it needs an RPC thread on the coordinator to be available to process the cancel. If all of the RPC threads are processing updates, they can all call LogFirstInProgress() and get stuck on the backend state lock for the coordinator's fragment. In that case, it becomes a temporary deadlock as the cancel can't be processed and the coordinator won't release the lock. It only gets resolved by the RPC timing out. To resolve this, this introduces an atomic is_done_ variable on Coordinator::BackendState that is set to true when reaching a terminal state. This introduces a IsDoneLockless() method to fetch the state from that atomic variable without getting the lock and switches LogFirstInProgress() to use it. This prevents the deadlock, as the RPC threads don't need to get a lock on another backend to make progress. Testing: - Hand tested with 30 impalads and control_service_num_svc_threads=1 Without the fix, it reproduces easily after reverting IMPALA-10047. With the fix, it doesn't reproduce. Change-Id: Ia058b03c72cc4bb83b0bd0a19ff6c8c43a647974 --- M be/src/runtime/coordinator-backend-state.cc M be/src/runtime/coordinator-backend-state.h M be/src/runtime/coordinator.cc 3 files changed, 42 insertions(+), 3 deletions(-) git pull ssh://gerrit.cloudera.org:29418/Impala-ASF refs/changes/64/23264/1 -- To view, visit http://gerrit.cloudera.org:8080/23264 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-MessageType: newchange Gerrit-Change-Id: Ia058b03c72cc4bb83b0bd0a19ff6c8c43a647974 Gerrit-Change-Number: 23264 Gerrit-PatchSet: 1 Gerrit-Owner: Joe McDonnell <joemcdonn...@cloudera.com>
