Re: postgres_fdw fails to see that array type belongs to extension

I wrote:
> There's one big remaining to-do item, I think: experimentation with
> pg_upgrade proves that a binary upgrade fails to fix the extension
> membership of arrays/rowtypes. That's because pg_dump needs to
> manually reconstruct the membership dependencies, and it thinks that
> it doesn't need to do anything for implicit arrays. Normally the
> point of that is that we want to exactly reconstruct the extension's
> contents, but I think in this case we'd really like to add the
> additional pg_depend entries even if they weren't there before.
> Otherwise people wouldn't get the new behavior until they do a
> full dump/reload.

> I can see two ways we could do that:

> * add logic to pg_dump

> * modify ALTER EXTENSION ADD TYPE so that it automatically recurses
> from a base type to its array type (and I guess we'd need to add
> something for relation rowtypes and multiranges, too).

> I think I like the latter approach because it's like how we
> handle ownership: pg_dump doesn't emit any commands to explicitly
> change the ownership of dependent types, either. (But see [1].)
> We could presumably steal some logic from ALTER TYPE OWNER.
> I've not tried to code that here, though.

Now that the multirange issue is fixed (3e8235ba4), here's a
new version that includes the needed recursion in ALTER EXTENSION.
I spent some more effort on a proper test case, too.

regards, tom lane