Robert Haas <robertmhaas@gmail.com> writes:
> I don't have any complaint about labelling some of the unique indexes
> as primary keys, but I think installing foreign keys that don't really
> enforce anything may lead to confusion.
I'm not sure if I buy the "confusion" argument, but after thinking
about this more I realized that there's a stronger roadblock for
treating catalog interrelationships as SQL foreign keys. Namely,
that we always represent no-reference situations with a zero OID,
whereas it'd have to be NULL to look like a valid foreign-key case.
Changing to real NULLs in those columns would of course break a ton
of C code, so I don't think that's gonna happen.
We could overlay an FK onto OID columns that should never have zero
entries, but that would miss enough of the interesting cases that
I don't find it a very attractive proposal.
So I think we should finish up the project of making real SQL
constraints for the catalog indexes, but the idea of making FK
constraints too looks like it'll end in failure.
The reason I got interested in this right now is the nearby
discussion [1] about why findoidjoins misses some catalog
relationships and whether we should fix that and/or make its results
more readily accessible. I'd thought perhaps FK constraint entries
could supersede the need for that tool altogether, but now it seems
like not. I still like the idea of marking OID relationships in the
catalog headers though. Perhaps we should take Joel's suggestion
of a new system catalog more seriously, and have genbki.pl populate
such a catalog from info in the catalog header files.
regards, tom lane
[1] https://www.postgresql.org/message-id/flat/d1f413ff-0e50-413d-910c-bdd9ee9752c1%40www.fastmail.com