I wrote:
> The relevant owner field for a composite type is pg_type.typowner.
> We don't bother to maintain pg_class.relowner for the subsidiary
> pg_class entry.
Actually, now that I look at it, there is a non-cosmetic issue here:
we seem to be creating a dependency link for the pg_class owner field.
We have to either not do that, or be willing to fix it during ALTER TYPE
OWNER. For instance
regression=# create user foo;
CREATE ROLE
regression=# create user bar;
CREATE ROLE
regression=# \c - foo
You are now connected as new user "foo".
regression=> create type mytype as (f1 int);
CREATE TYPE
regression=> \c - postgres
You are now connected as new user "postgres".
regression=# drop user foo;
ERROR: role "foo" cannot be dropped because some objects depend on it
DETAIL: owner of composite type mytype
owner of type mytype
regression=# alter type mytype owner to bar;
ALTER TYPE
regression=# drop user foo;
ERROR: role "foo" cannot be dropped because some objects depend on it
DETAIL: owner of composite type mytype
regression=#
regards, tom lane