Isaac Morland <isaac.morland@gmail.com> writes:
>> so 7. 3. 2020 v 22:20 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:
>>> Actually ... have you given any thought to just deciding that ranges and
>>> multiranges are the same type? That is, any range can now potentially
>>> contain multiple segments?
> Definitely agreed that range and multirange (or whatever it's called)
> should be different. In the work I do I have a number of uses for ranges,
> but not (yet) for multiranges. I want to be able to declare a column as
> range and be sure that it is just a single range, and then call lower() and
> upper() on it and be sure to get just one value in each case; and if I
> accidentally try to take the union of ranges where the union isn’t another
> range, I want to get an error rather than calculate some weird (in my
> context) multirange.
I do not find that argument convincing at all. Surely you could put
that constraint on your column using "CHECK (numranges(VALUE) <= 1)"
or some such notation.
Also, you're attacking a straw man with respect to lower() and upper();
I did not suggest changing them to return arrays, but rather interpreting
them as returning the lowest or highest endpoint, which I think would be
transparent in most cases. (There would obviously need to be some other
functions that could dissect a multirange more completely.)
The real problem with the proposal as it stands, I think, is exactly
that range union has failure conditions and you have to use some other
operator if you want to get a successful result always. That's an
enormously ugly kluge, and if we'd done it right the first time nobody
would have objected.
Bottom line is that I don't think that we should add a pile of new moving
parts to the type system just because people are afraid of change;
arguably, that's *more* change (and more risk of bugs), not less.
Unifying the types would, for example, get rid of the pesky question
of what promoting a range to multirange should look like exactly,
because it'd be a no-op.
regards, tom lane