Re: [HACKERS] pgbench regression test failure

From: Fabien COELHO <coelho(at)cri(dot)ensmp(dot)fr>
To: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Cc: Steve Singer <steve(at)ssinger(dot)info>, PostgreSQL Developers <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: [HACKERS] pgbench regression test failure
Date: 2017-11-20 22:15:14
Message-ID: alpine.DEB.2.20.1711202255090.15686@lancre
Views: Raw Message | Whole Thread | Download mbox
Thread:
Lists: pgsql-hackers


Hello Tom,

>>> 2. ISTM that we should report that 100% of the transactions were
>>> above the latency limit, not 33%; that is, the appropriate base
>>> for the "number of transactions above the latency limit" percentage
>>> is the number of actual transactions not the number of scheduled
>>> transactions.
>
>> Hmmm. Allow me to disagree.
>
> I dunno, it just looks odd to me that when we've set up a test case in
> which every one of the transactions is guaranteed to exceed the latency
> limit, that it doesn't say that they all did. I don't particularly buy
> your assumption that the percentages should sum.

This is a side effect. The reason for me is that the user asked for some
transactions, and the results should be given relative to what was asked.

> Anybody else have an opinion there?

Good question.

>>> I also noticed that if I specify "-f sleep-100.sql" more than once,
>>> the per-script TPS reports are out of line. This is evidently because
>>> that calculation isn't excluding skipped xacts; but if we're going to
>>> define tps as excluding skipped xacts, surely we should do so there too.
>
>> I do not think that we should exclude skipped xacts.
>
> Uh ... why not?

Because I totally misinterpreted your sentence.

Indeed, the skipped transactions needs to be substracted from the count.
This is yet another bug.

>>> but I find that unduly optimistic. It should really read more like
>>> "if no transactions were executed, at best we'll get some platform-
>>> dependent spelling of NaN. At worst we'll get a SIGFPE."
>
>> Hmmm. Alas you must be right about spelling. There has been no report of
>> SIGFPE issue, so I would not bother with that.
>
> The core issue here really is that you're assuming IEEE float arithmetic.
> We have not gone as far as deciding that Postgres will only run on IEEE
> hardware, and I don't want to start in pgbench, especially not in
> seldom-exercised corner cases.

Hmmm. It has already started for some years without complaint. Do as you
feel about NaN. I can only say that I do not like much having zero to
stand for undefined.

--
Fabien.

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message Masahiko Sawada 2017-11-20 22:19:30 Re: [HACKERS] Moving relation extension locks out of heavyweight lock manager
Previous Message Alexander Korotkov 2017-11-20 22:07:37 Re: [HACKERS] CUBE seems a bit confused about ORDER BY