On Fri, 3 Mar 2006, Tom Lane wrote:
> Stephan Szabo <sszabo@megazone.bigpanda.com> writes:
> > On Fri, 3 Mar 2006, Tom Lane wrote:
> >> BTW, I had another thought about this: if we go this way, then the plans
> >> associated with RI check queries would essentially always be trivial
> >> index lookups (for everything except RI_Initial_Check).
>
> > Would we have to do anything odd if we want to be testing only some of the
> > index columns and possibly not in the index order (like match partial
> > where some of the fk side is null)? I don't honestly see us doing match
> > partial any time soon, but I'd like to have an idea of what'd be involved.
>
> Match partial would be an index lookup with a subset of the keys, which
> btree at least is fine with. You could argue that a "sufficiently
> partial" match would be better done as a seqscan, but I think we could
> just bull ahead and do it as indexscans always ...
So at that level, not having leading columns is okay, so we wouldn't have
an issue with this (as opposed to if we planned such a statement, where it
wouldn't choose to use the index)?
> >> If we did this then RI checks would no longer be subvertible by rules or
> >> user triggers.
>
> > I don't think that it'd really help because it's the actions that are
> > generally subvertible not the checks and since those are looking at the
> > potentially not indexed fk side, I don't think the above would apply.
>
> Oh, right, we'd probably still need to do planning in that case. Unless
> we wanted to insist on having an FK-side index too for every FK, which
> is something I'm not for.
>
> Do you think it's worth redoing the implementation of just the PK checks?
I'd say it's worth trying. If it's probably faster, and it actually
sounds like it might be cleaner code than the way the statements are
built, even if it's another, different way of doing it.