Message-ID: <6597cf88-a499-e5be-e027-62753b197500@oss.nttdata.com>
Date: Wed, 22 Dec 2021 13:56:23 +0900
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:91.0)
 Gecko/20100101 Thunderbird/91.4.0
Subject: Re: sequences vs. synchronous replication
Content-Language: en-US
To: Tomas Vondra <tomas.vondra@enterprisedb.com>,
 Amit Kapila <amit.kapila16@gmail.com>
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
References: <712cad46-a9c8-1389-aef8-faf0203c9be9@enterprisedb.com>
 <CAA4eK1KN-Pqe1y5w0uXU2Ttn8DK4WfqgwAbCRbcDt4yzQ88wRA@mail.gmail.com>
 <4450cd74-1a41-ec03-83ee-fcabc183715d@enterprisedb.com>
From: Fujii Masao <masao.fujii@oss.nttdata.com>
In-Reply-To: <4450cd74-1a41-ec03-83ee-fcabc183715d@enterprisedb.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Archived-At: 
 <https://www.postgresql.org/message-id/6597cf88-a499-e5be-e027-62753b197500%40oss.nttdata.com>
Precedence: bulk


On 2021/12/22 10:57, Tomas Vondra wrote:
> 
> 
> On 12/19/21 04:03, Amit Kapila wrote:
>> On Sat, Dec 18, 2021 at 7:24 AM Tomas Vondra
>> <tomas.vondra@enterprisedb.com> wrote:
>>>
>>> while working on logical decoding of sequences, I ran into an issue with
>>> nextval() in a transaction that rolls back, described in [1]. But after
>>> thinking about it a bit more (and chatting with Petr Jelinek), I think
>>> this issue affects physical sync replication too.
>>>
>>> Imagine you have a primary <-> sync_replica cluster, and you do this:
>>>
>>>     CREATE SEQUENCE s;
>>>
>>>     -- shutdown the sync replica
>>>
>>>     BEGIN;
>>>     SELECT nextval('s') FROM generate_series(1,50);
>>>     ROLLBACK;
>>>
>>>     BEGIN;
>>>     SELECT nextval('s');
>>>     COMMIT;
>>>
>>> The natural expectation would be the COMMIT gets stuck, waiting for the
>>> sync replica (which is not running), right? But it does not.
>>>
>>
>> How about if we always WAL log the first sequence change in a transaction?
>>
> 
> I've been thinking about doing something like this, but I think it would not have any significant advantages compared to using "SEQ_LOG_VALS 0". It would still have the same performance hit for plain nextval() calls, and there's no measurable impact on simple workloads that already write WAL in transactions even with SEQ_LOG_VALS 0.

Just idea; if wal_level > minimal, how about making nextval_internal() (1) check whether WAL is replicated to sync standbys, up to the page lsn of the sequence, and (2) forcibly emit a WAL record if not replicated yet? The similar check is performed at the beginning of SyncRepWaitForLSN(), so probably we can reuse that code.

Regards,

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