On 2022-Aug-23, Jehan-Guillaume de Rorthais wrote:
Hi,
[...]
> However, it seems get_relation_idx_constraint_oid(), introduced in eb7ed3f3063,
> assume there could be only ONE constraint depending to an index. But in fact,
> multiple constraints can rely on the same index, eg.: the PK and a self
> referencing FK. In consequence, when looking for a constraint depending on an
> index for the given relation, either the FK or a PK can appears first depending
> on various conditions. It is then possible to trick it make a FK constraint a
> parent of a PK...
Hmm, wow, that sounds extremely stupid. I think a sufficient fix might
be to have get_relation_idx_constraint_oid ignore any constraints that
are not unique or primary keys. I tried your scenario with the attached
and it seems to work correctly. Can you confirm? (I only ran the
pg_regress tests, not anything else for now.)
If this is OK, we should make this API quirkiness very explicit in the
comments, so the patch needs to be a few lines larger in order to be
committable. Also, perhaps the check should be that contype equals
either primary or unique, rather than it doesn't equal foreign.
--
Álvaro Herrera 48°01'N 7°57'E — https://www.EnterpriseDB.com/