Horton

Re: FailedAssertion("pd_idx == pinfo->nparts", File: "execPartition.c", Line: 1689)

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Amit Langote <amitlangote09@gmail.com>
Cc: Justin Pryzby <pryzby@telsasoft.com>, Robert Haas <robertmhaas@gmail.com>, "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Date: 2020-08-05T17:30:30Z
Lists: pgsql-hackers

Attachments

Amit Langote <amitlangote09@gmail.com> writes:
> The crash reported here is in the other case where the concurrently
> added partitions cause the execution-time PartitionDesc to have more
> partitions than the one that PartitionedRelPruneInfo is based on.
> I was able to reproduce such a crash as follows:

Yeah, I can repeat the case per these directions.  I concur that the
issue is that ExecCreatePartitionPruneState is failing to cope with
zeroes in the relid_map.

> The attached patch should fix that.

I don't like this patch at all though; I do not think it is being nearly
careful enough to ensure that it's matched the surviving relation OIDs
correctly.  In particular it blithely assumes that a zero in relid_map
*must* match the immediately next entry in partdesc->oids, which is easy
to break if the new partition is adjacent to the one the planner managed
to prune.  So I think we should do it more like the attached.

I'm strongly tempted to convert the trailing Assert to an actual
test-and-elog, too, but didn't do so here.

			regards, tom lane

Commits

  1. Support testing of cases where table schemas change after planning.

  2. Fix matching of sub-partitions when a partitioned plan is stale.

  3. Speed up planning when partitions can be pruned at plan time.

  4. Allow ATTACH PARTITION with only ShareUpdateExclusiveLock.