Re: GiST: PickSplit and multi-attr indexes

Neil Conway <neilc@samurai.com> writes:
> On Mon, 2004-11-15 at 10:19 -0500, Tom Lane wrote:
>> I'm not familiar with the details of the GiST code, but would it work to
>> generalize PickSplit to return a three-way classification?  That is,
>> instead of actually splitting the node, have it identify each item as
>> "definitely left", "definitely right", or "don't care" (think "less",
>> "greater", or "equal to" the desired split value).

> I had thought about this solution, but I was worried that it will
> actually make GiST a less general framework, because it requires that
> the PickSplit() method effectively order the leaves of the tree. For
> btree that is a reasonable assumption to make, but I'm not sure if it
> can or should be made in general.

If there are no don't-care cases, then you're effectively saying that
the first column's PickSplit has sole control over the tree shape,
which is where we're at now.  ISTM the entire point of a multi-column
index is that the first column has duplicates, or at least values that
are similar enough to qualify as don't-cares.
        regards, tom lane