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

On Sun, Aug 2, 2020 at 2:11 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> Based on commit logs, I suspect this may be an "older bug", specifically maybe
> with:
>
> |commit 898e5e3290a72d288923260143930fb32036c00c
> |Author: Robert Haas <rhaas@postgresql.org>
> |Date:   Thu Mar 7 11:13:12 2019 -0500
> |
> |    Allow ATTACH PARTITION with only ShareUpdateExclusiveLock.
>
> I don't think it matters, but the process surrounding the table being INSERTed
> INTO is more than a little special, involving renames, detaches, creation,
> re-attaching within a transaction.  I think that doesn't matter though, and the
> issue is surrounding the table being SELECTed *from*, which is actually behind
> a view.

That's an entirely reasonable guess, but it doesn't seem easy to
understand exactly what happened here based on the provided
information. The assertion failure probably indicates that
pinfo->relid_map[] and partdesc->oids[] differ in some way other than
additional elements having been inserted into the latter. For example,
some elements might have disappeared, or the order might have changed.
This isn't supposed to happen, because DETACH PARTITION requires
heavier locking, and the order changing without anything getting
detached should be impossible. But evidently it did. If we could dump
out the two arrays in question it might shed more light on exactly how
things went wrong.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company