On Thu, Nov 12, 2015 at 10:23 AM, Amit Kapila <amit.kapila16@gmail.com> wrote:
> Thanks for the report. The reason for this problem is that instrumentation
> information from workers is getting aggregated multiple times. In
> ExecShutdownGatherWorkers(), we call ExecParallelFinish where it
> will wait for workers to finish and then accumulate stats from workers.
> Now ExecShutdownGatherWorkers() could be called multiple times
> (once we read all tuples from workers, at end of node) and it should be
> ensured that repeated calls should not try to redo the work done by first
> call.
> The same is ensured for tuplequeues, but not for parallel executor info.
> I think we can safely assume that we need to call ExecParallelFinish() only
> when there are workers started by the Gathers node, so on those lines
> attached patch should fix the problem.
I suggest that we instead fix ExecParallelFinish() to be idempotent.
Add a "bool finished" flag to ParallelExecutorInfo and return at once
if it's already set. Get rid of the exposed
ExecParallelReinitializeTupleQueues() interface and have
ExecParallelReinitialize(pei) instead. Have that call
ReinitializeParallelDSM(), ExecParallelSetupTupleQueues(pei->pcxt,
true), and set pei->finished = false. I think that would give us a
slightly cleaner separation of concerns between nodeGather.c and
execParallel.c.
Your fix seems a little fragile. You're relying on node->reader !=
NULL to tell you whether the readers need to be cleaned up, but in
fact node->reader is set to a non-NULL value AFTER the pei has been
created. Granted, we currently always create a reader unless we don't
get any workers, and if we don't get any workers then failing to call
ExecParallelFinish is currently harmless, but nonetheless I think we
should be more explicit about this so it doesn't accidentally get
broken later.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
