On Sun, Jan 16, 2005 at 07:28:07PM -0600, Jim C. Nasby wrote:
> On Sun, Jan 16, 2005 at 08:01:36PM -0500, Tom Lane wrote:
> > "Jim C. Nasby" <decibel@decibel.org> writes:
> > > Wouldn't the original proposal that had a state machine handle this?
> > > IIRC the original idea was:
> >
> > > new tuple -> known good -> possibly dead -> known dead
> >
> > Only if you disallow the transition from possibly dead back to known
> > good, which strikes me as a rather large disadvantage. Failed UPDATEs
> > aren't so uncommon that it's okay to have one permanently disable the
> > optimization.
>
> Actually, I guess I wasn't understanding the problem to begin with.
> You'd never go from new tuple to known good while the transaction that
> created the tuple was in-flight, right? If that's the case, I'm not sure
> where there's a race condition. You can't delete a tuple that hasn't
> been committed, right?
Er, nevermind, I thought about it and realized the issue.
What changes when a delete is done on a tuple? It seems that's the
key... if a tuple has been marked as possibly being expired/deleted,
don't allow it to go into known_good unless you can verify that the
transaction that marked it as potentially deleted was rolled back.
--
Jim C. Nasby, Database Consultant decibel@decibel.org
Give your computer some brain candy! www.distributed.net Team #1828
Windows: "Where do you want to go today?"
Linux: "Where do you want to go tomorrow?"
FreeBSD: "Are you guys coming, or what?"