Andres Freund <andres@anarazel.de> writes:
> On 2015-08-01 19:13:05 -0400, Noah Misch wrote:
>> That's a bug. The test_ddl_deparse suite leaves a shell type, which
>> pg_upgrade fails to reproduce. Whether to have pg_upgrade support that or
>> just error out cleanly is another question.
> There seems little justification to not support shell types. We should
> also add a shell type to the standard regression testing database,
> they're "weird" enough that some increased exposure seems like a good
> idea.
Agreed. I was a bit surprised to find that pg_dump skips shell types,
actually. Probably that's a hangover from when "create function foo()
returns bogus" would autocreate a shell type named "bogus". In all
modern releases, it's fairly hard to accidentally create a shell type,
so we should probably assume that the user meant it to be there.
regards, tom lane