On 11/8/23 16:52, Tom Lane wrote:
> Tomas Vondra <tomas.vondra@enterprisedb.com> writes:
>> On 11/8/23 16:10, Justin Pryzby wrote:
>>> I found this in our logs, and reproduced it under v11-v16.
>>>
>>> CREATE TABLE t(a int, b int);
>>> INSERT INTO t SELECT generate_series(1,999);
>>> CREATE STATISTICS t_stats ON a,b FROM t;
>>>
>>> while :; do psql postgres -qtxc "ANALYZE t"; done &
>>> while :; do psql postgres -qtxc "begin; DROP STATISTICS t_stats"; done &
>>>
>>> It's known that concurrent DDL can hit elog(). But in this case,
>>> there's only one DDL operation.
>
>> AFAICS this happens because store_statext (after ANALYZE builds the new
>> statistics) does this:
>
> Shouldn't DROP STATISTICS be taking a lock on the associated table
> that is strong enough to lock out ANALYZE?
>
Yes, I think that's the correct thing to do. I recall having a
discussion about this with someone while working on the patch, leading
to the current code. But I haven't managed to find that particular bit
in the archives :-(
Anyway, the attached patch should fix this by getting the lock, I think.
- RemoveStatisticsById is what gets called drop DROP STATISTICS (or for
dependencies), so that's where we get the AE lock
- RemoveStatisticsDataById gets called from ANALYZE, so that already
should have a lock (so no need to acquire another one)
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company