Обсуждение: BUG #13796: ALTER TYPE DROP COLUMN -- unexpected behavior ?

The following bug has been logged on the website:

Bug reference:      13796
Logged by:          Peter Plachta
Email address:      pplachta@gmail.com
PostgreSQL version: 9.4.5
Operating system:   mac os x
Description:

Here is a testcase:
create type complex as (a1 int, a2 numeric, a3 text, a4 int, a5 int);
create or replace function foo(arg complex) returns complex as $$
begin
  return ( select arg );
end; $$ language plpgsql;
-- modify type test
alter type complex drop attribute a4;
select foo(row(1, 1.1, 'one', 111));
======================
The last one prints:
     foo
--------------
 (1,1.1,one,)
(1 row)
======================
Why is the last element NULLed out? It's not like I can pass 5 elements to
the function.

pplachta@gmail.com writes:
> create type complex as (a1 int, a2 numeric, a3 text, a4 int, a5 int);
> create or replace function foo(arg complex) returns complex as $$
> begin
>   return ( select arg );
> end; $$ language plpgsql;
> alter type complex drop attribute a4;
> [ foo() stops working ]

Yeah, the problem is that since "arg" has a named composite type, it is
handled using the PLPGSQL_DTYPE_ROW code path, which sets up a plpgsql
Datum for each column at function compile time.  So the rowtype is baked
into the function at that point.  If you start a fresh session everything
is fine.

A real fix might involve switching over to the PLPGSQL_DTYPE_REC code
path, which I've advocated for for some time but it'd be pretty invasive.
Or perhaps we could arrange to force recompilation of a plpgsql function
if any composite type it depends on has changed.  Nobody's really gotten
excited enough about this to do either ...

            regards, tom lane

Thanks for looking Tom !

Yeah, I have looked at the PLPGSQL_DTYPE_REC code path and that looks hard.
Let me look at the recompilation angle, if I have a fix of some sort I'll
let you know.

  peter

On Tue, Dec 8, 2015 at 1:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:

> pplachta@gmail.com writes:
> > create type complex as (a1 int, a2 numeric, a3 text, a4 int, a5 int);
> > create or replace function foo(arg complex) returns complex as $$
> > begin
> >   return ( select arg );
> > end; $$ language plpgsql;
> > alter type complex drop attribute a4;
> > [ foo() stops working ]
>
> Yeah, the problem is that since "arg" has a named composite type, it is
> handled using the PLPGSQL_DTYPE_ROW code path, which sets up a plpgsql
> Datum for each column at function compile time.  So the rowtype is baked
> into the function at that point.  If you start a fresh session everything
> is fine.
>
> A real fix might involve switching over to the PLPGSQL_DTYPE_REC code
> path, which I've advocated for for some time but it'd be pretty invasive.
> Or perhaps we could arrange to force recompilation of a plpgsql function
> if any composite type it depends on has changed.  Nobody's really gotten
> excited enough about this to do either ...
>
>                         regards, tom lane
>