Jeff Janes <jeff.janes@gmail.com> writes:
> I've found an interesting performance problem in the intarray extension
> module. It doesn't seem to be version dependent, I've verified it in 9.4.4
> and 9.6devel.
Seems like this isn't specifically an issue for intarray, but rather one
with the core GIN code not being smart about the combination of selective
and unselective index conditions. In particular, it seems like the smart
thing for GIN to do with this example is just ignore the @@ condition
altogether so far as the index search goes, and mark all the results as
needing recheck so that the @@ operator gets applied as a filter.
You could also complain about the core planner not considering leaving
some potentially-indexable quals out of the actual index condition,
but TBH I don't see that that would be a useful use of planner cycles.
In almost every case it would mean that if there are K quals potentially
usable with a given index, we'd cost out 2^K-1 index paths and immediately
reject all but the use-em-all alternative. That's a lot of cycles to
spend to handle a corner case. It's always been assumed to be the index
AM's problem to optimize use of the index quals it's handed, and I don't
see why that shouldn't be true here.
> The proposed selectivity estimate improvement patch for @@ does not fix
> this (and evaluating that patch was how I found this issue.)
Nah, it wouldn't: the cost estimates are correct, in the sense that
gincostestimate realizes that GIN will be horribly stupid about this.
regards, tom lane