Re: knngist - 0.8

On Fri, Oct 15, 2010 at 7:10 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> I still feel vaguely uneasy about the fact that the proposed patch
>> can't handle ASC/DESC or NULLS FIRST/LAST, and that unease grew a bit
>> more last night when I read Peter's patch to add collation support.
>
> Good point.
>
>> We could possibly cram ASC/DESC and NULLS FIRST/LAST in by defining
>> four new categories of operator strategies rather than one, but
>> there's no way that's going to work for collations.  Is there some
>> other way to approach this problem?  Can we leave pg_amop as it is,
>> and pass the context (sort vs. qual, ASC/DESC, NULLS FIRST/LAST,
>> collation, whatever...) to the index via some sort of side channel?
>
> Well, we cannot avoid changing pg_amop, or at least changing its
> interpretation, because the current scheme simply can't represent
> indexable operators that are used for anything except simple boolean
> WHERE tests.

What exactly do you mean by that?

It has always seemed to me that the operator class mechanism is a
complicated abstraction mechanism that actually adds only a very small
amount of abstraction.  Instead of referring to operators by name or
OID we refer to them by a number that maps onto a name or OID.  That
allows the user to change the name or OID without breaking anything,
but that's about it.  Perhaps we should think of pg_amop not so much
as a way to tell the AM what to do, but just a way to tell it what
operator is logically involved without relying on the name or OID.

> I agree though that we do *not* want pg_amop involved
> in the details of exactly what sort ordering is referenced by a sortable
> operator.  Somehow that needs to be passed in a side channel.

Yeah.

> Maybe we should think in terms of a side channel for Peter's patch
> as well.  I share your feeling that trying to propagate collation
> the same way we now propagate typmod is a recipe for serious pain.

I'm not sure what you're thinking of here.  I think we can have the
idea of a FullySpecifiedType = <typid, typmod, collationoid>, but
that's not so much a side channel as an abstraction layer.  It
absolutely won't work to stuff the collations in a global variable or
something like that, if that's what you're imagining.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company