Date: Mon, 7 Mar 2022 07:09:22 -0600
From: Justin Pryzby <pryzby@telsasoft.com>
To: Magnus Hagander <magnus@hagander.net>
Cc: pgsql-hackers@lists.postgresql.org
Subject: Re: Add parameter jit_warn_above_fraction
Message-ID: <20220307130922.GQ27651@telsasoft.com>
References: <CABUevExR_9ZmkYj-aBvDreDKUinWLBBpORcmTbuPdNb5vGOLtA@mail.gmail.com>
 <20220225162304.GD9008@telsasoft.com>
 <CABUevExe1fV0RzjA7vQRp_WkjtTVF1O9E5STKodWa2CmpFoqHA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CABUevExe1fV0RzjA7vQRp_WkjtTVF1O9E5STKodWa2CmpFoqHA@mail.gmail.com>
User-Agent: Mutt/1.9.4 (2018-02-28)
Archived-At: <https://www.postgresql.org/message-id/20220307130922.GQ27651%40telsasoft.com>
Precedence: bulk

On Mon, Mar 07, 2022 at 01:10:32PM +0100, Magnus Hagander wrote:
> On Fri, Feb 25, 2022 at 5:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> >
> > I think it should be a NOTICE (or less?)
> 
> Hmm. I'm not so sure.
> 
> Other similar parameters often use LOG, but the downside of that is
> that it won't be sent to the client.
> 
> The problem with using NOTICE is that it won't go to the log by
> default. It needs to be at least warning to do that.

Anyone who uses this parameter is aleady going to be changing GUCs, so it
doesn't need to work by default.  The people who are likely to enable this
already monitor their logs and have probably customized their logging
configuration.

> > Is it more useful if this is applied combined with log_min_duration_statement ?
> >
> > It's easy to imagine a query for which the planner computes a high cost, but
> > actually runs quickly.  You might get a bunch of WARNINGs that the query took
> > 10 MS and JIT was 75% of that, even if you don't care about queries that take
> > less than 10 SEC.
> 
> Yeah, that's definitely a problem. But which way would you want to tie
> it to log_min_duration_statement?  That a statement would both have to
> take longer than log_min_duration_statement *and* have JIT above a
> certain percentage? In my experience that is instead likely to miss
> most of the interesting times.

I don't understand - why doesn't it combine trivially with
log_min_duration_statement?  Are you saying that the default / pre-existing min
duration may not log all of the intended queries ?  I think that just means the
configuration needs to be changed.  The GUC needs to allow/help finding these
JIT issues, but going to require an admin's interaction in any case.  Avoiding
false positives may be important for it to be useful at all.

Hmm .. should it also combine with min_sample_rate ?  Maybe that addresses your
concern.

-- 
Justin