Re: pg_dump(1) failures when concurrently refreshing mat views

Nico Williams <nico(at)cryptonector(dot)com> writes:
> On Wed, Apr 19, 2017 at 05:20:31PM -0400, Tom Lane wrote:
>> ... uh, really? It's not supposed to take anything higher than ACCESS
>> SHARE, which should coexist fine with anything short of ACCESS EXCLUSIVE.

> Whoops, sorry, that's a mistake in drafting the report. You're right,
> it takes an ACCESS SHARE MODE lock, NOWAIT. The remainder of the report
> remains correct.

But REFRESH CONCURRENTLY takes EXCLUSIVE lock, not ACCESS EXCLUSIVE lock,
so it won't conflict with pg_dump. There's something missing from your
explanation.

Also, the only use of NOWAIT I see is in a parallel pg_dump worker.
It is known that parallel pg_dump can lead to deadlocks if third
parties are taking ACCESS EXCLUSIVE locks --- but you haven't identified
where the ACCESS EXCLUSIVE request is coming from.

regards, tom lane