Re: undetected deadlock in ALTER SUBSCRIPTION ... REFRESH PUBLICATION

On 11/23/23 10:24, Amit Kapila wrote:
> On Wed, Nov 22, 2023 at 4:51 PM Tomas Vondra
> <tomas(dot)vondra(at)enterprisedb(dot)com> wrote:
>>
>> On 11/22/23 11:38, Amit Kapila wrote:
>>>
>>> Okay. IIUC, what's going on here is that the apply worker acquires
>>> AccessShareLock on pg_subscription to update rel state for one of the
>>> tables say tbl-1, and then for another table say tbl-2, it started
>>> waiting for a state change via wait_for_relation_state_change(). I
>>> think here the fix is to commit the transaction before we go for a
>>> wait. I guess we need something along the lines of what is proposed in
>>> [1] though we have solved the problem in that thread in some other
>>> way..
>>>
>>
>> Possibly. I haven't checked if the commit might have some unexpected
>> consequences, but I can confirm I can no longer reproduce the deadlock
>> with the patch applied.
>>
>
> Thanks for the verification. Offhand, I don't see any problem with
> doing a commit at that place but will try to think some more about it.
> I think we may want to call pgstat_report_stat(false) after commit to
> avoid a long delay in stats.
>

Makes sense.

> I haven't verified but I think this will be a problem in
> back-branches as well.
>

Yes. I haven't tried but I don't see why backbranches wouldn't have the
same issue.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company