Re: Suggestion to add --continue-client-on-abort option to pgbench - Mailing list pgsql-hackers

From Yugo Nagata
Subject Re: Suggestion to add --continue-client-on-abort option to pgbench
Date
Msg-id 20251111105037.f3fc554616bc19891f926c5b@sraoss.co.jp
Whole thread Raw
In response to Re: Suggestion to add --continue-client-on-abort option to pgbench  (Fujii Masao <masao.fujii@gmail.com>)
Responses Re: Suggestion to add --continue-client-on-abort option to pgbench
Re: Suggestion to add --continue-client-on-abort option to pgbench
List pgsql-hackers
On Fri, 7 Nov 2025 18:33:17 +0900
Fujii Masao <masao.fujii@gmail.com> wrote:

> I plan to commit the patch soon, but let's keep discussing and
> investigating the case you mentioned afterward!

I'm sorry for the late reply and for not joining the discussion earlier.

I've spent some time investigating the code in pgbench and libpq, and
it seems to me that your commit looks fine.

However, I found another issue related to the --continue-on-error option,
where an assertion failure occurs in the following test case:

 $ cat pgbench_error.sql 
 \startpipeline
 select 1/0;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \endpipeline

 $ pgbench -f pgbench_error.sql -M extended --continue-on-error -T 1
 pgbench (19devel)
 starting vacuum...end.
 pgbench: pgbench.c:3594: discardUntilSync: Assertion `res == ((void *)0)' failed.

Even after removing the Assert(), we get the following error:

 pgbench: error: client 0 aborted: failed to exit pipeline mode for rolling back the failed transaction

This happens because discardUntilSync() does not expect that a PGRES_TUPLES_OK may be
received after \syncpipeline, and also fails to discard all PGRES_PIPELINE_SYNC results
when multiple \syncpipeline commands are used.

I've attached a patch to fix this.
If a PGRES_PIPELINE_SYNC is followed by something other than PGRES_PIPELINE_SYNC or NULL,
it means that another PGRES_PIPELINE_SYNC will eventually follow after some other results.
In this case, we should reset the receive_sync flag and continue discarding results.

I think this fix should be back-patched, since this is not a bug introduced by
--continue-on-error. The same assertion failure occurs in the following test case,
where transactions are retried after a deadlock error:

 $ cat deadlock.sql 
 \startpipeline
 select * from a order by i for update;
 select 1;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \endpipeline

 $ cat deadlock2.sql 
 \startpipeline
 select * from a order by i desc for update;
 select 1;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \syncpipeline
 select 1;
 \endpipeline

 $ pgbench -f deadlock.sql -f deadlock2.sql -j 2 -c 2 -M extended  

Regards,
Yugo Nagata

-- 
Yugo Nagata <nagata@sraoss.co.jp>

Attachment

pgsql-hackers by date:

Previous
From: Michael Paquier
Date:
Subject: Re: Sequence Access Methods, round two
Next
From: "Jonathan S. Katz"
Date:
Subject: Re: 2025-11-13 release announcement draft