Re: range_agg

On Wed, 2019-08-21 at 21:54 -0700, Paul A Jungwirth wrote:
> Sometimes I think about having a maybe<T> type instead of null/not
> null. SQL nulls are already very "monadic" I think but nullability
> doesn't travel.

Yeah, that would be a great direction, but there is some additional
complexity that we'd need to deal with that a "normal" compiler does
not:

  * handling both existing global types (like int) as well as on-the-
fly types like Maybe<Either<Int,Bool>>
  * types need to do more things, like have serialized representations,
interface with indexing strategies, and present the optimizer with
choices that may influence which indexes can be used or not
  * at some point needs to work with normal SQL types and NULL
  * there are a lot of times we care not just whether a type is
sortable, but we actually care about the way it's sorted (e.g.
localization). typeclasses+newtype would probably be unacceptable for
trying to match SQL behavior here.

I'm all in favor of pursuing this, but it's not going to bear fruit
very soon.

> Btw I have working multirange_{send,recv,in,out} now, and I
> automatically create a multirange type and its array type when
> someone
> creates a new range type. I have a decent start on passing tests and
> no compiler warnings. I also have a start on anymultirange and
> anyrangearray. (I think I need the latter to support a range-like
> constructor function, so you can say `int4multirange(int4range(1,4),
> int4range(8,10))`.) I want to get the any* types done and improve the
> test coverage, and then I'll probably be ready to share a patch.

Awesome!

> Here are a couple other questions:
> 
> - Does anyone have advice for the typanalyze function? I feel pretty
> out of my depth there (although I haven't looked into typanalyze
> stuff
> very deeply yet). I can probably get some inspiration from
> range_typanalyze and array_typanalyze, but those are both quite
> detailed (their statistics functions that is).

I think Alexander Korotkov did a lot of the heavy lifting here, perhaps
he has a comment? I'd keep it simple for now if you can, and we can try
to improve it later.

> - What should a multirange do if you give it an empty range? I'm
> thinking it should just ignore it, but then `'{}'::int4multirange =
> '{empty}'::int4multirange`. Does that seem okay? (It does to me
> actually, if we think of `empty` as the additive identity. Similarly
> mr + empty = mr.

I agree. Multiranges are more than just an array of ranges, so they
coalesce into some canonical form.

> - What should a multirange do if you give it a null, like
> `int4multirange(int4range(1,4), null)`. I'm thinking it should be
> null, just like mr + null = null. Right?

Yes. NULL is for the overall multirange datum (that is, a multirange
column can be NULL), but I don't think individual parts of a datatype
make much sense as NULL. So, I agree that mr + null = null. (Note that
arrays and records can have NULL parts, but I don't see a reason we
should follow those examples for multiranges.)

Regards,
    Jeff Davis