Re: RI oddness

Jan Wieck <JanWieck(at)Yahoo(dot)com> writes:
> What'd be easy is this:

> - We already have two entry points for INSERT/UPDATE on FK
> table, but the one for UPDATE is fortunately unused.

> - We change analyze.c to install the RI_FKey_check_upd
> trigger if the constraint has an ON DELETE SET DEFAULT
> clause. Otherwise it uses RI_FKey_check_ins as it does
> now.

Unfortunately, such a fix really isn't going to fly as a patch release.
Not only does it not work for existing tables, but it won't work for
tables created by dump and reload from a prior version (since they
won't have the right set of triggers ... another illustration of why
the lack of an abstract representation of the RI constraints was a
Bad Move). In fact I'm afraid that your proposed change would actively
break tables imported from a prior version; wouldn't RI_FKey_check_ins
do the wrong thing if applied as an update trigger?

> I think the usage of ON DELETE SET DEFAULT is a very rare
> case out in the field. Thus the dump/reload requirement is
> limited to a small number of databases (if any).

But dump/reload won't fix the tables' triggers.

Given that ON DELETE SET DEFAULT isn't used much, I think we should
not waste time creating an incomplete hack solution for 7.1.*, but
just write it off as a known bug and move forward with a real solution
for 7.2.

regards, tom lane