On 13/12/16 01:49, Andres Freund wrote:
> On 2016-12-13 01:42:48 +0100, Petr Jelinek wrote:
>> On 13/12/16 01:40, Tom Lane wrote:
>>> Petr Jelinek <petr.jelinek@2ndquadrant.com> writes:
>>>> On 13/12/16 00:39, Tom Lane wrote:
>>>>> Hm, buildfarm says this didn't fix it. Where exactly does the dropping
>>>>> of the slot happen ... is it not complete by the time the backend exits?
>>>
>>>> Yes, I've seen. The cleanup of slots is done in ProcKill(), I wonder,
>>>> since it's on_shmem_exit hook, and the pgstat_beshutdown_hook is as
>>>> well, maybe the pgstat_beshutdown_hook is called before ProcKill and the
>>>> query is lucky to hit right in between.
>>>
>>> Hm. That seems like a pretty bogus place to do it. An awful lot of the
>>> backend infrastructure is already gone by then, if I recall the ordering
>>> correctly. Maybe ShutdownPostgres would be a saner place; but it really
>>> depends on what you think the module layering is for this facility.
>>> I would definitely not think it is proc.c's responsibility, though.
>>>
>>
>> Well, the problem is that that's the place where the currently active
>> slot is released (if there was an active one). So we'd need to move that
>> part somewhere else as well. I am not sure what's the reasoning for
>> releasing it at that specific spot so CCing Andres.
>
> Why don't we just instead make the loop over pg_replication_slots?
>
I mentioned that as possible solution upthread, I am only worried that
the failure scenario is basically infinite loop.
--
Petr Jelinek http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services