Re: pg_get_wal_replay_pause_state() should not return 'paused' while a promotion is ongoing. - Mailing list pgsql-hackers

From Fujii Masao
Subject Re: pg_get_wal_replay_pause_state() should not return 'paused' while a promotion is ongoing.
Date
Msg-id 25396c69-9f70-ca3e-5de3-f3a108ae2a26@oss.nttdata.com
Whole thread Raw
In response to Re: pg_get_wal_replay_pause_state() should not return 'paused' while a promotion is ongoing.  (Kyotaro Horiguchi <horikyota.ntt@gmail.com>)
Responses Re: pg_get_wal_replay_pause_state() should not return 'paused' while a promotion is ongoing.
List pgsql-hackers

On 2021/05/19 15:25, Kyotaro Horiguchi wrote:
> At Wed, 19 May 2021 11:19:13 +0530, Dilip Kumar <dilipbalaut@gmail.com> wrote in
>> On Wed, May 19, 2021 at 10:16 AM Fujii Masao
>> <masao.fujii@oss.nttdata.com> wrote:
>>>
>>> On 2021/05/18 15:46, Michael Paquier wrote:
>>>> On Tue, May 18, 2021 at 12:48:38PM +0900, Fujii Masao wrote:
>>>>> Currently a promotion causes all available WAL to be replayed before
>>>>> a standby becomes a primary whether it was in paused state or not.
>>>>> OTOH, something like immediate promotion (i.e., standby becomes
>>>>> a primary without replaying outstanding WAL) might be useful for
>>>>> some cases. I don't object to that.
>>>>
>>>> Sounds like a "promotion immediate" mode.  It does not sound difficult
>>>> nor expensive to add a small test for that in one of the existing
>>>> recovery tests triggerring a promotion.  Could you add one based on
>>>> pg_get_wal_replay_pause_state()?
>>>
>>> You're thinking to add the test like the following?
>>> #1. Pause the recovery
>>> #2. Confirm that pg_get_wal_replay_pause_state() returns 'paused'
>>> #3. Trigger standby promotion
>>> #4. Confirm that pg_get_wal_replay_pause_state() returns 'not paused'
>>>
>>> It seems not easy to do the test #4 stably because
>>> pg_get_wal_replay_pause_state() needs to be executed
>>> before the promotion finishes.
>>
>> Even for #2, we can not ensure that whether it will be 'paused' or
>> 'pause requested'.
> 
> We often use poll_query_until() to make sure some desired state is
> reached.

Yes.

>  And, as Michael suggested, the function
> pg_get_wal_replay_pause_state() still works at the time of
> recovery_end_command.  So a bit more detailed steps are:

IMO this idea is tricky and fragile, so I'm inclined to avoid that if possible.
Attached is the POC patch to add the following tests.

#1. Check that pg_get_wal_replay_pause_state() reports "not paused" at first.
#2. Request to pause archive recovery and wait until it's actually paused.
#3. Request to resume archive recovery and wait until it's actually resumed.
#4. Request to pause archive recovery and wait until it's actually paused.
        Then, check that the paused state ends and promotion continues
        if a promotion is triggered while recovery is paused.

In #4, pg_get_wal_replay_pause_state() is not executed while promotion
is ongoing. #4 checks that pg_is_in_recovery() returns false and
the promotion finishes expectedly in that case. Isn't this test enough for now?

Regards,

-- 
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION

Attachment

pgsql-hackers by date:

Previous
From: Fabien COELHO
Date:
Subject: Re: pgbench test failing on 14beta1 on Debian/i386
Next
From: Michael Paquier
Date:
Subject: Re: pgbench test failing on 14beta1 on Debian/i386