Re: Minimal logical decoding on standbys

Hi,

On 1/24/23 1:46 AM, Andres Freund wrote:
> Hi,
>
> On 2023-01-19 10:43:27 +0100, Drouvot, Bertrand wrote:
>>>> With a reload in place in my testing, now I notice that the catalog_xmin
>>>> is updated on the primary physical slot after logical slots invalidation
>>>> when reloading hot_standby_feedback from "off" to "on".
>>>>
>>>> This is not the case after a re-start (aka catalog_xmin is NULL).
>>>>
>>>> I think a re-start and reload should produce identical behavior on
>>>> the primary physical slot. If so, I'm tempted to think that the catalog_xmin
>>>> should be updated in case of a re-start too (even if all the logical slots are invalidated)
>>>> because the slots are not dropped yet. What do you think?
>>>
>>> I can't quite follow the steps leading up to the difference. Could you list
>>> them in a bit more detail?
>>>
>>>
>>
>> Sure, so with:
>>
>> 1) hot_standby_feedback set to off on the standby
>> 2) create 2 logical replication slots on the standby and activate one
>> 3) Invalidate the logical slots on the standby with VACUUM FULL on the primary
>> 4) change hot_standby_feedback to on on the standby
>>
>> If:
>>
>> 5) pg_reload_conf() on the standby, then on the primary we get a catalog_xmin
>> for the physical slot that the standby is attached to:
>>
>> postgres=# select slot_type,xmin,catalog_xmin from pg_replication_slots ;
>> slot_type | xmin | catalog_xmin
>> -----------+------+--------------
>> physical | 822 | 748
>> (1 row)
>
> How long did you wait for this to change?

Almost instantaneous after pg_reload_conf() on the standby.

> I don't think there's anything right
> now that'd force a new hot-standby-feedback message to be sent to the primary,
> after slots got invalidated.
>
> I suspect that if you terminated the walsender connection on the primary,
> you'd not see it anymore either?
>

Still there after the standby is shutdown but disappears when the standby is re-started.

> If that isn't it, something is broken in InvalidateObsolete...
>

Will look at what's going on and ensure catalog_xmin is not sent to the primary after pg_reload_conf() (if the slots are invalidated).

>
>> No, but a question still remains to me:
>>
>> Given the fact that the row removal case is already done
>> in the next test (aka Scenario 2), If we want to replace the "vacuum full" test
>> on the database (done in Scenario 1) with a cheaper one at the table level,
>> what could it be to guarantee an invalidation?
>>
>> Same as scenario 2 but with "vacuum full pg_class" would not really add value
>> to the tests, right?
>
> A database wide VACUUM FULL is also just a row removal test, no?

Yeah, so I was wondering if Scenario 1 was simply not just useless.

> I think it
> makes sense to test that both VACUUM and VACUUM FULL both trigger conflicts,
> because they internally use *very* different mechanisms.

Got it, will do and replace Scenario 1 as you suggested initially.

> It'd probably be
> good to test at least conflicts triggered due to row removal via on-access
> pruning as well. And perhaps also for btree killtuples. I think those are the
> common cases for catalog tables.
>

Thanks for the proposal, will look at it.

Regards,

--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com