Re: "tuple concurrently updated" in pg_restore --jobs

Justin Pryzby <pryzby(at)telsasoft(dot)com> writes:
> I hit this issue intermittently (roughly half the time) while working with a
> patch David submitted, and finally found a recipe to reproduce it on an
> unpatched v12 instance.

> I was surprised to see pg_restore -j2 is restoring ACLs in pre-data in
> parallel.

It's not pre-data. But it's true that pg_restore figures it can restore
ACLs in parallel during the ACL-restoring pass, on the theory that pg_dump
will not emit two different ACL entries for the same object, so that we
can do all the catalog updates in parallel without conflicts.

This works about 99% of the time, in fact. It falls down in the --clean
case if we have to revoke existing table permissions, because in that case
the REVOKE at table level is required to clear the table's per-column ACLs
as well, so that that ACL entry involves touching the same catalog rows
that the per-column ACLs want to touch.

I think the right fix is to give the per-column ACL entries dependencies
on the per-table ACL, if there is one. This will not fix the problem
for the case of restoring from an existing pg_dump archive that lacks
such dependency links --- but given the lack of field complaints, I'm
okay with that.

This looks straightforward, if somewhat tedious because we'll have to
change the API of pg_dump's dumpACL() function, which is called by
a lot of places. Barring objections, I'll go do that.

regards, tom lane