On 2016-04-26 12:39:37 +0900, Michael Paquier wrote:
> Thinking about the logging of smgr invalidations, this is quite
> interesting. But what would we actually gain in doing that? Do you
> foresee any advantages in doing so? The only case where this would be
> useful now is for vm_extend by looking at the code.
Well, it'd make vm_extend actually correct, which replacing the
invalidation with a relcache one would not. Relcache invalidations are
transactional, whereas smgr ones are not (intentionally so!). I don't
think it's currently a big problem, but it does make me rather wary.
> >> As the invalidation patch is aimed at being backpatched, this may be
> >> something to do as well in back-branches.
> >
> > I'm a bit split here. I think forcing processing of invalidations at
> > moments they've previously never been processed is a bit risky for the
> > back branches. But on the other hand relcache invalidations are only
> > processed at end-of-xact, which isn't really correct for the code at
> > hand :/
>
> Oh, OK. So you mean that this patch is not aimed for back-branches
> with this new record type, but that's only for HEAD.
No, I think we got to do this in all branches. I was just wondering
about how to fix vm_extend(). Which I do think we got to fix, even in
the back-branches.
- Andres
