Re: [bug] Table not have typarray when created by single user mode

I wrote:
> However, if we're going to go this far, I think there's a good
> case to be made for going all the way and eliminating the policy
> of not making array types for system catalogs. That was never
> anything but a wart justified by space savings in pg_type, and
> this patch already kills most of the space savings. If we
> drop the system-state test in heap_create_with_catalog altogether,
> we end up with 601 initial pg_type entries. That still leaves
> the four bootstrap catalogs without array types, because they are
> not created by heap_create_with_catalog; but we can manually add
> those too for a total of 605 initial entries. (That brings initial
> pg_type to 14 pages as I show above; I think it was 13 with the
> original version of the patch.)
> In short, if we're gonna do this, I think we should do it like
> the attached. Or we could do nothing, but there is some appeal
> to removing this old inconsistency.

I pushed that, but while working on it I had a further thought:
why is it that we create composite types but not arrays over those
types for *any* relkinds? That is, we could create even more
consistency, as well as buying back some of the pg_type bloat added
here, by not creating pg_type entries at all for toast tables or
sequences. A little bit of hacking later, I have the attached.

One could argue it either way as to whether sequences should have
composite types. It's possible to demonstrate queries that will
fail without one:

regression=# create sequence seq1;
CREATE SEQUENCE
regression=# select s from seq1 s;
ERROR: relation "seq1" does not have a composite type

but it's pretty hard to believe anyone's using that in practice.
Also, we've talked more than once about changing the implementation
of sequences to not have a relation per sequence, in which case the
ability to do something like the above would go away anyway.

Comments?

regards, tom lane