Re: Suggestion to add --continue-client-on-abort option to pgbench

On Thu, 25 Sep 2025 13:49:05 +0900
Fujii Masao <masao.fujii@gmail.com> wrote:

> On Thu, Sep 25, 2025 at 11:17 AM Yugo Nagata <nagata@sraoss.co.jp> wrote:
> >
> > On Thu, 25 Sep 2025 11:09:40 +0900
> > Yugo Nagata <nagata@sraoss.co.jp> wrote:
> >
> > > On Thu, 25 Sep 2025 02:19:27 +0900
> > > Fujii Masao <masao.fujii@gmail.com> wrote:
> > >
> > > > On Tue, Sep 23, 2025 at 11:58 AM Rintaro Ikeda
> > > > <ikedarintarof@oss.nttdata.com> wrote:
> > > > > I think the issue is a new bug because we have transitioned to CSTATE_ABORT
> > > > > immediately after queries failed, without executing discardUntilSync().
> > >
> > > Agreed.
> > >
> > > > If so, the fix in discardUntilSync() should be part of the continue-on-error
> > > > patch. The assertion failure fix should be a separate patch, since only
> > > > that needs to be backpatched (the failure can also occur in back branches).
> > >
> > > +1
> > >
> > > >
> > > > > I've attached a patch that fixes the assertion error. The content of v1 patch by
> > > > > Mr. Nagata is also included. I would appreciate it if you review my patch.
> > >
> > > > + switch (PQresultStatus(res))
> > > > + {
> > > > + case PGRES_PIPELINE_SYNC:
> > > > + received_sync = true;
> > > >
> > > > In the PGRES_PIPELINE_SYNC case, PQclear(res) isn't called but should be.
> > > >
> > > > + case PGRES_FATAL_ERROR:
> > > > + PQclear(res);
> > > > + goto done;
> > > >
> > > > I don't think we need goto here. How about this instead?
> > > >
> > > > -----------------------
> > > > @@ -3525,11 +3525,18 @@ discardUntilSync(CState *st)
> > > >                          * results have been discarded.
> > > >                          */
> > > >                         st->num_syncs = 0;
> > > > -                       PQclear(res);
> > > >                         break;
> > > >                 }
> > > > +               /*
> > > > +                * Stop receiving further results if PGRES_FATAL_ERROR
> > > > is returned
> > > > +                * (e.g., due to a connection failure) before
> > > > PGRES_PIPELINE_SYNC,
> > > > +                * since PGRES_PIPELINE_SYNC will never arrive.
> > > > +                */
> > > > +               else if (PQresultStatus(res) == PGRES_FATAL_ERROR)
> > > > +                       break;
> > > >                 PQclear(res);
> > > >         }
> > > > +       PQclear(res);
> > > >
> > > >         /* exit pipeline */
> > > >         if (PQexitPipelineMode(st->con) != 1)
> > > > -----------------------
> > >
> > > I think Fujii-san's version is better because Ikeda-san's version doesn't
> > > consider the case where PGRES_PIPELINE_SYNC is followed by another one.
> > > In that situation, the loop would terminate without getting NULL, which would
> > > causes an error in PQexitPipelineMode().
> > >
> > > However, I would like to suggest an alternative solution: checking the connection
> > > status when readCommandResponse() returns false. This seems more straightforwad,
> > > since the cause of the error can be investigated immediately after it is detected.
> > >
> > > @@ -4024,7 +4043,10 @@ advanceConnectionState(TState *thread, CState *st, StatsData *agg)
> > >                                         if (PQpipelineStatus(st->con) != PQ_PIPELINE_ON)
> > >                                                 st->state = CSTATE_END_COMMAND;
> > >                                 }
> > > -                               else if (canRetryError(st->estatus))
> > > +                               else if (PQstatus(st->con) == CONNECTION_BAD)
> > > +                                       st->state = CSTATE_ABORTED;
> > > +                               else if ((st->estatus == ESTATUS_OTHER_SQL_ERROR && continue_on_error) ||
> > > +                                                canRetryError(st->estatus))
> > >                                         st->state = CSTATE_ERROR;
> > >                                 else
> > >                                         st->state = CSTATE_ABORTED;
> > >
> > > What do you think?
> > >
> > >
> > > Additionally, I noticed that in pipeline mode, the error message reported in
> > > readCommandResponse() is lost, because it is reset when PQgetResult() returned
> > > NULL to indicate the end of query processing. For example:
> > >
> > >  pgbench: client 0 got an error in command 3 (SQL) of script 0;
> > >  pgbench: client 1 got an error in command 3 (SQL) of script 0;
> > >
> > > This can be fixed this by saving the previous error message and using it
> > > for the report. After the fix:
> > >
> > >  pgbench: client 0 got an error in command 3 (SQL) of script 0; FATAL:  terminating connection due to administrator command
> > >
> > > I've attached update patches.
> > >
> > > 0001 fixes the assersion failure in commandError() and error message lost
> > > in readCommandResponse().
> > >
> > > 0002 was the previous 0001 that adds --continiue-on-error, including the
> > > fix to handle connection termination errors.
> > >
> > > 0003 is for improving error messages for errors that cause client abortion.
> >
> > I think the patch 0001 should be back patched since the issues can occurs
> > even for retries of serialization failure or deadlock detection in pipeline mode.
> 
> Agreed.
> 
> Regarding 0001:
> 
> + /*
> + Errors should only be detected during an SQL command or the \endpipeline
> + meta command. Any other case triggers an assertion failure.
> + */
> 
> * should be added before "Errors" and "meta".

Oops. Fixed.
 
> + errmsg = pg_strdup(PQerrorMessage(st->con));
> 
> It would be good to add a comment explaining why we do this.
> 
> + pg_free(errmsg);
> 
> Shouldn't pg_free() be called also in the error case, i.e., after
> jumping to the error label?

Yes, it should be.
Fixed.

I've attached updated patches.

Regards,
Yugo Nagata

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