Re: Avoid full GIN index scan when possible

Tomas Vondra <tomas(dot)vondra(at)2ndquadrant(dot)com> writes:
> On Fri, Jun 28, 2019 at 03:03:19PM -0400, Tom Lane wrote:
>> I not only don't want that function in indxpath.c, I don't even want
>> it to be known/called from there. If we need the ability for the index
>> AM to editorialize on the list of indexable quals (which I'm not very
>> convinced of yet), let's make an AM interface function to do it.

> Wouldn't it be better to have a function that inspects a single qual and
> says whether it's "optimizable" or not? That could be part of the AM
> implementation, and we'd call it and it'd be us messing with the list.

Uh ... we already determined that the qual is indexable (ie is a member
of the index's opclass), or allowed the index AM to derive an indexable
clause from it, so I'm not sure what you envision would happen
additionally there. If I understand what Julien is concerned about
--- and I may not --- it's that the set of indexable clauses *as a whole*
may have or lack properties of interest. So I'm thinking the answer
involves some callback that can do something to the whole list, not
qual-at-a-time. We've already got facilities for the latter case.

> But that kinda resembles stuff we already have - selectivity/cost. So
> why shouldn't this be considered as part of costing?

Yeah, I'm not entirely convinced that we need anything new here.
The cost estimate function can detect such situations, and so can
the index AM at scan start --- for example, btree checks for
contradictory quals at scan start. There's a certain amount of
duplicative effort involved there perhaps, but you also have to
keep in mind that we don't know the values of run-time-determined
comparison values until scan start. So if you want certainty rather
than just a cost estimate, you may have to do these sorts of checks
at scan start.

regards, tom lane