Thread: Selectivity estimation for intarray

Alexander Korotkov <aekorotkov@gmail.com> writes:
> My proposal is to let ALTER OPERATOR change restrict and join selectivity
> functions of the operator. Also it would be useful to be able to change
> commutator and negator of operator: extension could add commutators and
> negators in further versions. Any thoughts?

I'm pretty dubious about this, because we lack any mechanism for undoing
parser/planner decisions based on operator properties.  And there's quite
a lot of stuff that is based on the assumption that operator properties
will never change.

An example of the pitfalls here is that we can never allow ALTER OPERATOR
RENAME, because for example if you rename '<' to '~<~' that will change
its precedence, and we have no way to fix the parse trees embedded in
stored views to reflect that.

For the specific cases you mention, perhaps it would be all right if we
taught plancache.c to blow away *all* cached plans upon seeing any change
in pg_operator; but that seems like a brute-force solution.
        regards, tom lane

Oleg Bartunov <obartunov@gmail.com> writes:
> Any chance to have this patch in 9.5 ? Many intarray users will be happy.

Considering how desperately behind we are on reviewing/committing patches
that were submitted by the deadline, I don't think it would be appropriate
or fair to add on major new patches that came in months late.  Please add
this to the first 9.6 commitfest, instead.
        regards, tom lane

* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Alexander Korotkov <aekorotkov@gmail.com> writes:
> > My proposal is to let ALTER OPERATOR change restrict and join selectivity
> > functions of the operator. Also it would be useful to be able to change
> > commutator and negator of operator: extension could add commutators and
> > negators in further versions. Any thoughts?
>
> I'm pretty dubious about this, because we lack any mechanism for undoing
> parser/planner decisions based on operator properties.  And there's quite
> a lot of stuff that is based on the assumption that operator properties
> will never change.
>
> An example of the pitfalls here is that we can never allow ALTER OPERATOR
> RENAME, because for example if you rename '<' to '~<~' that will change
> its precedence, and we have no way to fix the parse trees embedded in
> stored views to reflect that.
>
> For the specific cases you mention, perhaps it would be all right if we
> taught plancache.c to blow away *all* cached plans upon seeing any change
> in pg_operator; but that seems like a brute-force solution.

Agreed that it is- but is that really a problem...?  I've not run into
many (any?) systems where pg_operator is getting changed often...  The
worst part would be adding new operators/extensions, but perhaps we
could exclude that specific case from triggering the cache invalidation?
Thanks!
    Stephen

Stephen Frost <sfrost@snowman.net> writes:
> * Tom Lane (tgl@sss.pgh.pa.us) wrote:
>> For the specific cases you mention, perhaps it would be all right if we
>> taught plancache.c to blow away *all* cached plans upon seeing any change
>> in pg_operator; but that seems like a brute-force solution.

> Agreed that it is- but is that really a problem...?

Perhaps it isn't; we certainly have assumptions that pg_amop, for
instance, changes seldom enough that it's not worth tracking individual
changes.  The same might be true of pg_operator.  I'm not sure though.

The core point I'm trying to make is that making pg_operator entries
mutable is something that's going to require very careful review.
        regards, tom lane