On Fri, 2009-09-25 at 13:23 +0300, Heikki Linnakangas wrote:
> Simon Riggs wrote:
> > On Wed, 2009-09-23 at 17:45 +0300, Heikki Linnakangas wrote:
> >> XactClearRecoveryTransactions() when we see a shutdown checkpoint, which
> >> clears all recovery locks. But doesn't that prematurely clear all locks
> >> belonging to prepared transactions as well?
> >
> > Much better to read your second post(s). :-)
> >
> > Yes, you have found a(nother) issue. This was the first one that gave me
> > pause to think of the answer. The locks currently aren't tracked as to
> > whether they are 2PC or not, so we would need to store that info also so
> > that we can selectively release locks later.
> >
> > Question: is it possible to do a fast shutdown when we have a prepared
> > transaction?
>
> Yes.
>
> > Would it be better to take a different approach there for
> > prepared transactions? It seems strange to write a shutdown checkpoint
> > when the system isn't yet "clean".
>
> Hmm, I guess you could define prepared transactions as active backends
> from the shutdown point of view, and wait for them to finish. I can see
> one problem, though: Once you issue shutdown, fast or smart, we no
> longer accept new connections. So you can't connect to issue the
> ROLLBACK/COMMIT PREPARED anymore. Anyway, it would be a change from the
> current behavior, so it would be better to cope with prepared
> transactions in the standby.
Definitely need to cope with them for Hot Standby. My point was general
one to say that behaviour is very non-useful for users with prepared
transactions. It just causes manual effort by a DBA each time the system
is shutdown.
-- Simon Riggs www.2ndQuadrant.com
