To: Robert Haas <robertmhaas@gmail.com>
cc: Kevin Grittner <Kevin.Grittner@wicourts.gov>,
	Mladen Gogala <mladen.gogala@vmsinfo.com>,
	"pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: anti-join chosen even when slower than old plan 
In-reply-to: <AANLkTi=dzhRTTNm1LkyNCAC5+PWbp1jWtKFh4T1Bu4nj@mail.gmail.com> 
References: <4CD9663502000025000374B1@gw.wicourts.gov>
	<20961.1289347680@sss.pgh.pa.us>
	<4CDA62C60200002500037512@gw.wicourts.gov>
	<AANLkTi=eNXKEBo5dXJDd4+aHhfmkibH++fqjj12+DA3o@mail.gmail.com>
	<4CDACBBF020000250003755B@gw.wicourts.gov>
	<4CDB7BAA.7020506@vmsinfo.com>
	<4CDBB0A40200002500037571@gw.wicourts.gov>
	<AANLkTi=dzhRTTNm1LkyNCAC5+PWbp1jWtKFh4T1Bu4nj@mail.gmail.com>
Comments: In-reply-to Robert Haas <robertmhaas@gmail.com>
	message dated "Thu, 11 Nov 2010 12:57:54 -0500"
Date: Thu, 11 Nov 2010 13:23:03 -0500
Message-ID: <18450.1289499783@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>

Robert Haas <robertmhaas@gmail.com> writes:
> Let's back up a moment and talk about what the overall goal is, here.
> Ideally, we would like PostgreSQL to have excellent performance at all
> times, under all circumstances, with minimal tuning.  Therefore, we do
> NOT want to add variables that will, by design, need constant manual
> adjustment.  That is why I suggested that Tom's idea of an
> assume_cached GUC is probably not what we really want to do.   On the
> other hand, what I understand Mladen to be suggesting is something
> completely different.  He's basically saying that, of course, he wants
> it to work out of the box most of the time, but since there are
> guaranteed to be cases where it doesn't, how about providing some
> knobs that aren't intended to be routinely twaddled but which are
> available in case of emergency?  Bravo, I say!

Um ... those are exactly the same thing.  You're just making different
assumptions about how often you will need to twiddle the setting.
Neither assumption is based on any visible evidence, unfortunately.
I was thinking of assume_cached as something that could be
set-and-forget most of the time, and you're entirely right to criticize
it on the grounds that maybe it wouldn't.  But to support a proposal
that doesn't even exist yet on the grounds that it *would* be
set-and-forget seems a tad inconsistent.  We can't make that judgment
without a whole lot more details than have been provided yet for any
idea in this thread.

I do think that something based around a settable-per-table caching
percentage might be a reasonable way to proceed.  But the devil is in
the details, and we don't have those yet.

			regards, tom lane