To: Greg Stark <gsstark@MIT.EDU>
cc: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] GiST: PickSplit and multi-attr indexes 
In-reply-to: <87r7mt7b1i.fsf@stark.xeocode.com> 
References: <419723A7.5040505@samurai.com> <27866.1100476459@sss.pgh.pa.us>
	<1100492090.23420.28.camel@localhost.localdomain>
	<7879.1100531985@sss.pgh.pa.us>
	<1100578365.23420.71.camel@localhost.localdomain>
	<17937.1100616003@sss.pgh.pa.us> <87wtwl7bqe.fsf@stark.xeocode.com>
	<87r7mt7b1i.fsf@stark.xeocode.com>
Comments: In-reply-to Greg Stark <gsstark@MIT.EDU>
	message dated "16 Nov 2004 16:12:57 -0500"
Date: Tue, 16 Nov 2004 16:37:20 -0500
Message-ID: <2761.1100641040@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>
Lines: 22
Xref: stark.xeocode.com misc:37450

Greg Stark <gsstark@MIT.EDU> writes:
> The approach they take is to have a function which calculates an
> abstract "distance" between any two entries. There's an algorithm that
> they use to pick the split based on this distance function.

> If you abandoned "PickSplit" and instead exposed this distance
> function as the external API then the behaviour for multi-column
> indexes is clear. You calculate the distance along all the axes and
> calculate the diagonal distance.

Hmm ... the problem with that is the assumption that different opclasses
will compute similarly-scaled distances.  If opclass A generates
distances in the range (0,1e6) while B generates in the range (0,1),
combining them with Euclidean distance won't work well at all.  OTOH you
can't blindly normalize, because in some cases maybe the data is such
that a massive difference in distances is truly appropriate.

I'm also a bit leery of the assumption that every GiST application can
reduce its PickSplit logic to Euclidean distances.

			regards, tom lane