Re: plpgsql versus domains

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Thu, Feb 26, 2015 at 1:53 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> To some extent this is a workaround for the fact that plpgsql does type
>> conversions the way it does (ie, by I/O rather than by using the parser's
>> cast mechanisms). We've talked about changing that, but people seem to
>> be afraid of the compatibility consequences, and I'm not planning to fight
>> two compatibility battles concurrently ;-)

> A sensible plan, but since we're here:

> - I do agree that changing the way PL/pgsql does those type
> conversions is scary. I can't predict what will break, and there's no
> getting around the scariness of that.

> - On the other hand, I do think it would be good to change it, because
> this sucks:

> rhaas=# do $$ declare x int; begin x := (3.0::numeric)/(1.0::numeric); end $$;
> ERROR: invalid input syntax for integer: "3.0000000000000000"
> CONTEXT: PL/pgSQL function inline_code_block line 1 at assignment

If we did want to open that can of worms, my proposal would be to make
plpgsql type conversions work like so:

* If there is a cast pathway known to the core SQL parser, use that
mechanism.

* Otherwise, attempt to convert via I/O as we do today.

This seems to minimize the risk of breaking things, although there would
probably be corner cases that work differently (for instance I bet boolean
to text might turn out differently). In the very long run we could perhaps
deprecate and remove the second phase.

regards, tom lane