Re: Range Types, constructors, and the type system

On Mon, 2011-06-27 at 14:50 -0400, Robert Haas wrote:
> Couldn't we also do neither of these things? I mean, presumably
> '[1,10]'::int8range had better work.

I think that if we combine this idea with Florian's "PAIR" suggestion
here:
http://archives.postgresql.org/message-id/AD4FC75D-DB99-48ED-9082-52EE3A4D74A6@phlo.org

then I think we have a solution.

If we add a type RANGEINPUT that is not a pseudotype, we can use that as
an intermediate type that is returned by range constructors. Then, we
add casts from RANGEINPUT to each range type. That would allow
range(1,2)::int8range
to work without changing the type system around, because range() would
have the signature:
range(ANYELEMENT, ANYELEMENT) -> RANGEINPUT
and then the cast would change it into an int8range. But we only need
the one cast per range type, and we can also support all of the other
kinds of constructors like:
range_cc(ANYELEMENT, ANYELEMENT) -> RANGEINPUT
range_linf_c(ANYELEMENT) -> RANGEINPUT
without additional hassle.

The RANGEINPUT type itself would hold similar information to actual
range types: the subtype OID (instead of the range type, because it's
not a range yet), optionally the two bounds (depending on the flags),
and the flags byte. The cast to a real range type would read the
subtype, and try to coerce the bounds to the subtype of the range you're
casting to, set the range type oid, leave the flags byte the same, and
it's done.

So, in effect, RANGEINPUT is a special type used only for range
constructors. If someone tried to output it, it would throw an
exception, and we'd even have enough information at that point to print
a nice error message with a hint.

Actually, this is pretty much exactly Florian's idea (thanks again,
Florian), but at the time I didn't like it because "pair" didn't capture
everything that I wanted to capture, like infinite bounds, etc. But
there's no reason that it can't, and your point made me realize that --
you are effectively just using TEXT as the intermediate type (which
works, but has some undesirable characteristics).

Do we think that this is a good way forward? The only thing I can think
of that's undesirable is that it's not normal to be required to cast the
result of a function, and might be slightly difficult to explain in the
documentation in a straightforward way.

Regards,
Jeff Davis