Hello Hayato-san,
> I played pgbench with wrong parameters,
That's good:-)
> and I found bug-candidate.
>
> 1. Do initdb and start.
> 2. Initialize schema and data with "scale factor" = 1.
> 3. execute following command many times:
>
> $ pgbench -c 101 -j 10 postgres
>
> Then, sometimes the negative " initial connection time" was returned.
> Lateyncy average is also strange.
>
> ```
> $ pgbench -c 101 -j 10 postgres
> starting vacuum...end.
> pgbench: error: connection to server on socket "/tmp/.s.PGSQL.5432" failed: FATAL: sorry, too many clients already
Hmmm.
AFAICR there was a decision to generate a report even if something went
very wrong, in this case some client could not connect, so some values
are not initialized, hence the absurd figures, as you show below.
Maybe we should revisit this decision.
> initial connection time = -372896921.586 ms
> I sought pgbench.c and found a reason.
> When a thread failed to get some connections, they do not fill any values to thread->bench_start in threadRun().
> And if the failure is caused in the final thread (this means threads[nthreads - 1]->bench_start is zero),
> the following if-statement sets bench_start to zero.
> I cannot distinguish whether we have to fix it, but I attache the patch.
> This simply ignores a result when therad->bench_start is zero.
> How do you think?
Hmmm. Possibly. Another option could be not to report anything after some
errors. I'm not sure, because it would depend on the use case. I guess the
command returned an error status as well.
I'm going to give it some thoughts.
--
Fabien.