Re: Wrong results from Parallel Hash Full Join

On Wed, Apr 12, 2023 at 2:59 PM Melanie Plageman
<melanieplageman(at)gmail(dot)com> wrote:
>
> On Wed, Apr 12, 2023 at 2:14 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
> >
> > Hi,
> >
> > On 2023-04-12 10:57:17 -0400, Melanie Plageman wrote:
> > > HeapTupleHeaderHasMatch() checks if HEAP_TUPLE_HAS_MATCH is set.
> > >
> > > In htup_details.h, you will see that HEAP_TUPLE_HAS_MATCH is defined as
> > > HEAP_ONLY_TUPLE
> > > /*
> > > * HEAP_TUPLE_HAS_MATCH is a temporary flag used during hash joins. It is
> > > * only used in tuples that are in the hash table, and those don't need
> > > * any visibility information, so we can overlay it on a visibility flag
> > > * instead of using up a dedicated bit.
> > > */
> > > #define HEAP_TUPLE_HAS_MATCH HEAP_ONLY_TUPLE /* tuple has a join match */
> > >
> > > If you redefine HEAP_TUPLE_HAS_MATCH as something that isn't already
> > > used, say 0x1800, the query returns correct results.
> > > [...]
> > > The question is, why does this only happen for a parallel full hash join?
> >
> > I'd guess that PHJ code is missing a HeapTupleHeaderClearMatch() somewhere,
> > but the non-parallel case isn't.
>
> Indeed. Thanks! This diff fixes the case Richard provided.
>
> diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
> index a45bd3a315..54c06c5eb3 100644
> --- a/src/backend/executor/nodeHash.c
> +++ b/src/backend/executor/nodeHash.c
> @@ -1724,6 +1724,7 @@ retry:
> /* Store the hash value in the HashJoinTuple header. */
> hashTuple->hashvalue = hashvalue;
> memcpy(HJTUPLE_MINTUPLE(hashTuple), tuple, tuple->t_len);
> + HeapTupleHeaderClearMatch(HJTUPLE_MINTUPLE(hashTuple));
>
> /* Push it onto the front of the bucket's list */
> ExecParallelHashPushTuple(&hashtable->buckets.shared[bucketno],
>
> I will propose a patch that includes this change and a test.
>
> I just want to convince myself that ExecParallelHashTableInsertCurrentBatch()
> covers the non-batch 0 cases and we don't need to add something to
> sts_puttuple().

So, indeed, tuples in batches after batch 0 already had their match bit
cleared by ExecParallelHashTableInsertCurrentBatch().

Attached patch includes the fix for ExecParallelHashTableInsert() as
well as a test. I toyed with adapting one of the existing parallel full
hash join tests to cover this case, however, I think Richard's repro is
much more clear. Maybe it is worth throwing in a few updates to the
tables in the existing queries to provide coverage for the other
HeapTupleHeaderClearMatch() calls in the code, though.

- Melanie