From: Tom Lane <tgl@sss.pgh.pa.us>
To: Sergei Kornilov <sk@zsrv.org>
cc: pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #18172: High memory usage in tSRF function context
In-reply-to: <223611698479775@omv3tj5zwvd3plqc.iva.yp-c.yandex.net>
References: <18172-9b8c5fc1d676ded3@postgresql.org>
 <1736370.1698463142@sss.pgh.pa.us>
 <223611698479775@omv3tj5zwvd3plqc.iva.yp-c.yandex.net>
Comments: In-reply-to Sergei Kornilov <sk@zsrv.org>
	message dated "Sat, 28 Oct 2023 10:56:15 +0300"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <1977914.1698516908.1@sss.pgh.pa.us>
Date: Sat, 28 Oct 2023 14:15:08 -0400
Message-ID: <1977915.1698516908@sss.pgh.pa.us>
Archived-At: 
 <https://www.postgresql.org/message-id/1977915.1698516908%40sss.pgh.pa.us>
Precedence: bulk

Sergei Kornilov <sk@zsrv.org> writes:
> Thank you! The patch works for me too.

Thanks for testing!  I looked at ExecProjectSet a second time
and realized that I probably still hadn't thought hard enough,
because it's also responsible for doing ResetExprContext(econtext)
and that wasn't happening either when looping around after the
SRF returns no rows.  So if the SRF itself leaks some memory
in its CurrentMemoryContext and then returns nothing, we can
see bloat in that context.  I failed to produce such a behavior
with jsonb_array_elements, but after trying some other things
I found a test case that still leaked even with the initial
patch:

create table zed as
  select repeat('xyzzy', 10000) as f1, repeat('g', 10000) as flags
  from generate_series(1,100000);

select regexp_matches(f1, 'q', flags) from zed;

The leak here stems from regexp_matches detoasting its flags
argument in the CurrentMemoryContext.  A big flags argument
is surely not real-world usage, but perhaps there are other
cases that are more likely to happen.

The fix of course is just to make sure we also do ResetExprContext
when looping around.  Applied to all branches.

			regards, tom lane