Re: POC, WIP: OR-clause support for indexes

On Thu, Oct 3, 2024 at 4:15 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> Now that you're explicitly creating RestrictInfos for a particular
> index, I suppose that it might be easier to do this kind of thing --
> you have more context. Perhaps the patch can be made to recognize
> a mix of constants like this as all being associated with the same
> B-Tree operator family (the opfamily that the input opclass belongs
> to)? Perhaps the constants could all be normalized to the same type via
> casts/coercions into the underlying B-Tree input opclass -- that
> extra step should be correct ("64.1.2. Behavior of B-Tree Operator Classes"
> describes certain existing guarantees that this step would need to rely
> on).

I don't think you can convert everything to the same type because we
have to assume that type conversions can fail. An exception is if the
types are binary-compatible but that's not the case here. If there's a
way to fix this problem, it's probably by doing the first thing you
suggest above: noticing that all the constants belong to the same
opfamily. I'm not sure if that approach can work either, but I think
it has better chances.

Personally, I don't think this particular limitation is a problem. I
don't think it will be terribly frequent in practice, and it doesn't
seem any weirder than any of the other things that happen as a result
of small and large integer constants being differently typed.

--
Robert Haas
EDB: http://www.enterprisedb.com