Content-Type: text/plain;
	charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 16.0 \(3826.700.81\))
Subject: Re: Optimize LISTEN/NOTIFY
From: Chao Li <li.evan.chao@gmail.com>
In-Reply-To: <7556f0d4-03fd-451a-bd34-5f62b424319a@app.fastmail.com>
Date: Thu, 30 Oct 2025 11:22:40 +0800
Cc: pgsql-hackers <pgsql-hackers@postgresql.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <290910DE-9A03-4AE6-B348-073D5DA96ACC@gmail.com>
References: <6899c044-4a82-49be-8117-e6f669765f7e@app.fastmail.com>
 <165530.1752362320@sss.pgh.pa.us>
 <02a7cd37-e2fc-4212-8b19-f8c239c95fb8@app.fastmail.com>
 <e396ecf3-4227-4918-b9ff-e9568dcebcf0@app.fastmail.com>
 <b6a427fe-2ed6-4568-85d9-207a68172617@app.fastmail.com>
 <CAK80=jhmE40KVqQ3ho37MArS7cAED1p9m7uikDxcnDmqdW7t8A@mail.gmail.com>
 <96f00bf1-cc9d-4520-9d02-9e14e7767c88@app.fastmail.com>
 <30c2aa7d-dd6c-4b68-a2e4-f217a1a34acf@app.fastmail.com>
 <CA+hUKGLrMGkWDB0cwTa0RqD+AF7O-Ywgck8aVYKwOQnZgYRRug@mail.gmail.com>
 <0b4d402a-9ac2-4aa8-acf8-8231dbe579ea@app.fastmail.com>
 <c8e84bb2-d2e5-4a77-b547-3a93c5778e7e@app.fastmail.com>
 <3095599.1758644879@sss.pgh.pa.us>
 <0dc6a2cc-5216-4dc1-9dd2-430cafc6095b@app.fastmail.com>
 <E7D8A5CD-C445-4713-B6B7-6CC1D3A71161@gmail.com>
 <baf71275-9083-41be-ae13-f80821541a1a@app.fastmail.com>
 <52CC167F-763B-4ECA-B0B4-DAB381816828@gmail.com>
 <d35a114d-204a-4b40-9284-06749532b724@app.fastmail.com>
 <9186C6D0-F7A9-482A-9183-89E530B57E36@gmail.com>
 <da0996!!!!!!d8-ecf9-490d-bbd0-2a5bdb5b115a@app.fastmail.com>
 <D1CF7CA5-213A-4FC1-86DB-052D58904A0A@gmail.com>
 <cad78ff4-aae9-4c1c-8e16-ff0ccd4957f1@app.fastmail.com>
 <1073593.1759423179@sss.pgh.pa.us>
 <4bd5e6c4-6fa7-44bb-869d-59a32a331fa8@app.fastmail.com>
 <CAFY6G8dap-bCnAnMG-2Gzew8yv2Vbi9gsx9+yszKMmd57ygfvA@mail.gmail.com>
 <85828f29-e72e-4400-94f3-9a69bc8dc239@app.fastmail.com>
 <2495353.1759860890@sss.pgh.pa.us>
 <8aeae418-92a6-4bbd-9c06-9574c79e59f7@app.fastmail.com>
 <2531672.1759868124@sss.pgh.pa.us>
 <474efa78-337c-41cd-a73a-f845a0115109@app.fastmail.com>
 <2749343.1759949176@sss.pgh.pa.us>
 <8bfca2be-1ec0-4e15-aafb-0b7b661fe936@app.fastmail.com>
 <9eba307f-f2fb-48f0-9507-2e197f39ef9e@app.fastmail.com>
 <cde2a91d-d47a-43b5-ab27-f98aa739d38e@app.fastmail.com>
 <8c71183a-0d28-4bcf-a806-78446ff95404@app.fastmail.com>
 <1009807.1760476747@sss.pgh.pa.us>
 <1F7227F5-C33D-4E2C-8511-33F1468590D0@gmail.com>
 <0a5a20d3-4621-46b3-b2ab-903f63a20dea@app.fastmail.com>
 <6F913129-ABEF-4004-AAF3-F22FC34!29AE8@gmail.com>
 <cbabbb45-ec27-47f1-a8f3-eb29073e9aef@app.fastmail.com>
 <ceab2c54-3463-4029-a7d8-895a617bb457@app.fastmail.com>
 <1547585.1760645808@sss.pgh.pa.us>
 <b8df8040-1b15-4a1a-b5f6-74acebbd2d0c@app.fastmail.com>
 <CAE7r3MLivh1sHWF06hrVXkiQbw-KChPcQsh+9CheXprm5vRVMQ@mail.gmail.com>
 <14865EB6-0BF4-462B-9072-10BDAC10C052@gmail.com>
 <CAE7r3MK-3AOdh1mpZ8hw9h6F_i0D5RMoAy7CttnfCJRpB8GJDA@mail.gmail.com>
 <0BCA1C2D-B92C-459E-B1A6-6D06BA4C62CF@gmail.com>
 <55d24cbb-e9ef-491f-a99b-b3dbd7cecdf9@app.fastmail.com>
 <38574cad-e90d-47b7-a015-753bb6bbc360@app.fastmail.com>
 <66631FB7-5BEA-4ED5-A694-9AD8B9CCFEE8@gmail.com>
 <4b7b49a5-5e1a-44a8-93e0-60457d15cb1d@app.fastmail.com>
 <82DEA2B6-6FC5-4A79-BDE3-1FD72F104A6E@gmail.com>
 <38de1036-d8cf-420c-b845-edb5a946b191@app.fastmail.com>
 <87E40BF8-8877-4DBD-9040-99AF8A4E6358@gmail.com>
 <d29ae0a8-b28c-4450-bd9f-c56657e57f45@app.fastmail.com>
 <DF4BDA30-CC41-4BAF-9852-E399C3F273EC@gmail.com>
 <7556f0d4-03fd-451a-bd34-5f62b424319a@app.fastmail.com>
To: Joel Jacobson <joel@compiler.org>
Archived-At: 
 <https://www.postgresql.org/message-id/290910DE-9A03-4AE6-B348-073D5DA96ACC%40gmail.com>
Precedence: bulk


> On Oct 29, 2025, at 18:33, Joel Jacobson <joel@compiler.org> wrote:
>=20
> On Wed, Oct 29, 2025, at 08:05, Chao Li wrote:
>>=20
>> I think the current implementation still has a race problem.
>>=20
>> Let=E2=80=99s say notifier N1 notifies listener=E2=80=99s L1 to read =
message.
>> L1 starts to read: it acquires the look, gets reading range, then=20
>> releases the lock, start performs reading without holding the lock.
>> Notifier N2 comes, N2 doesn=E2=80=99t have anything L1 is interested =
in. N2 now=20
>> holds the look, when it checks "if (QUEUE_POS_EQUAL(pos,=20
>> queueHeadBeforeWrite))=E2=80=9D, here comes the race. Because the =
lock is in=20
>> N2=E2=80=99s hand, L1 cannot get the lock to update its pos, so "if=20=

>> (QUEUE_POS_EQUAL(pos, queueHeadBeforeWrite))=E2=80=9D will not be =
satisfied, so=20
>> direct advancement won=E2=80=99t happen.
>=20
> I'm not sure I agree that qualifies as a race "problem" per se, since =
I
> think that just sounds like a case where we would do an unnecessary
> wakeup, right?
>=20

Why unnecessary? Say there are 100 listeners L1 - L100. When N2 is =
checking state of L1, L100 has finished reading, ideally L100 should =
update its pos, then when N2 reaches L100, it should do direct =
advancement, right?

But now the problem is, we use a single notification lock to handle all =
notifiers and listeners. Assume if every backend process has a =
notification lock, then the race is no longer there. When N2 is checking =
state of L1, it just holds L1=E2=80=99s lock, so L100 can go ahead =
update its pos, then when N2 reaches L100, N2 can do direct advancement.

I ever thought to propose to use a lock for every backend process, but I =
didn=E2=80=99t, because a lock is underlying an expensive semaphore, if =
there are hundreds of backends, adding the same number of semaphores =
doesn=E2=80=99t seem a good thing, which would be a too many overheads =
to the system.

> Without more sophisticated data structures (e.g. skip ranges) and
> increased code complexity, there will always be cases where we will by
> do unnecessary wakeups, which IMO need not be a design goal to
> completely avoid, until we have benchmark data that indicates =
otherwise.
>=20

The other problem I see is that, we don=E2=80=99t have a way to evaluate =
if the =E2=80=9Cdirect advancement=E2=80=9D is really effective, such as =
1) if a case that can perform =E2=80=9Cdirect advancement=E2=80=9D is =
really applied the advancement; 2) in a test model, how many =E2=80=9Cdire=
ct advancement=E2=80=9D are applied.

> I think we should iterate by first trying to reason about correctness =
of
> the code, trying to prove/disprove if a notifications could ever end =
up
> not being delivered. The bug I fixed in v22 is an example of such a
> case, that would cause a listening backend to never be awaken, since
> notifiers would not signal it due to the pending wake that was not
> cleared.
>=20
> I wonder if there could be more such serious bugs in the current code. =
I
> will focus my efforts now trying to answer that question. Would be
> really nice if we could find a way to reason formally about this. I've
> been looking into the P programming language, which seems suitable for
> modeling and verifying these kind of asynchronous concurrency =
protocols,
> I will give it a try.
>=20

I don=E2=80=99t think we need to rush. =46rom my observation, none of =
the =E2=80=9Cbig=E2=80=9D patches can get merged quickly anyway. Rather =
than hurrying to make it =E2=80=9Cready,=E2=80=9D I think it=E2=80=99s =
better to take the time to make it =E2=80=9Cperfect=E2=80=9D. I have =
also spent a lot of time on this patch, and I don=E2=80=99t mind to =
spend more. If you need a hand, I will be happy to offer.

TBH, with all the problems I described earlier still in my brain, I just =
cannot convince myself to let this patch go yet. Sorry about that.

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/