Re: Avoid orphaned objects dependencies, take 3

Hi,

On Fri, Jun 21, 2024 at 01:22:43PM +0000, Bertrand Drouvot wrote:
> Another thought for the RelationRelationId class case: we could check if there
> is a lock first and if there is no lock then acquire one. That way that would
> ensure the relation is always locked (so no "risk" anymore), but OTOH it may
> add "unecessary" locking (see 2. mentioned previously).

Please find attached v12 implementing this idea for the RelationRelationId class
case. As mentioned, it may add unnecessary locking for 2. but I think that's
worth it to ensure that we are always on the safe side of thing. This idea is
implemented in LockNotPinnedObjectById().

A few remarks:

- there is one place where the relation is not visible (even if
CommandCounterIncrement() is used). That's in TypeCreate(), because the new
relation Oid is _not_ added to pg_class yet.
Indeed, in heap_create_with_catalog(), AddNewRelationType() is called before
AddNewRelationTuple()). I put a comment in this part of the code explaining why
it's not necessary to call LockRelationOid() here.

- some namespace related stuff is removed from "test_oat_hooks/expected/alter_table.out".
That's due to the logic in cachedNamespacePath() and the fact that the same
namespace related stuff is added prior in alter_table.out.

- the patch touches 37 .c files, but that's mainly due to the fact that
LockNotPinnedObjectById() has to be called in a lot of places.

Regards,

--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com