Re: TODO: Fix CREATE CAST on DOMAINs

Gevik Babakhani <pgdev(at)xs4all(dot)nl> writes:
> First I would like to know how PG's code looked like without the
> domains.

IIRC, as far as the datatype coercion and operator/function resolution
code were concerned, the domain patch basically consisted of dropping
getBaseType() calls in at a bunch of choke points, so that these
routines all treat a domain the same as its base type.

I was never real happy about this, mainly because it adds extra syscache
lookups that buy you nothing when you're not using domains.

What would be nice is to revert all that, and instead have domain
creation insert explicit cast entries between a domain and its base type
into pg_cast, so that domains don't require any special-case code in
this part of the system. The problem with that glib answer is that
maintaining anything like the current behavior would seem to require
that we abandon the current principle that we consider only one-step
cast pathways when trying to match arguments to functions. (For
instance, a varchar value can be passed to a text-taking function
because varchar->text is an implicit cast according to pg_cast. What
of a domain over varchar?) And allowing multi-step casts to be chosen
automatically is unbelievably scary --- I think it would basically
destroy the system's ability to resolve overloaded functions at all,
because you can get from any type to any other if you allow enough cast
steps. (The unreasonably large number of implicit casts to text aren't
helping any here :-()

So the hard part of this doesn't really require any understanding of
code at all. What we need is a proposal for an algorithm that loosens
the casting rules "just enough" to make explicit pg_cast entries for
domains work the way we would like them to, without wholesale breakage
of situations that have nothing to do with domains. See
http://developer.postgresql.org/pgdocs/postgres/typeconv.html

Rereading what I just wrote, it might be as simple as allowing a
two-step cast in certain cases, only if the first step is a domain to
base type coercion (which we assume would be specially marked in
pg_cast). But the devil is in the details ... and anyway there might
be a cleaner approach than that.

regards, tom lane