On Thu, Jun 16, 2011 at 11:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> In typical cases where both versions of the row are on the same page,
> the window for the concurrent commit to happen is very narrow --- that's
> why you need so many clients to make it happen easily. With enough
> clients there's a good chance of losing the CPU between tuple visits.
> But of course Murphy's Law says this will happen in production
> situations even if the load isn't so high.
Thanks for doing the research on this. Much better we know now than
enter production like this.
> I believe that this is fundamentally unavoidable so long as we use
> SnapshotNow to read catalogs --- which is something we've talked about
> changing, but it will require a pretty major R&D effort to make it
> happen. In the meantime, we have to go back to using
> AccessExclusiveLock for table alterations. It doesn't help to have
> a lower lock level if that means that concurrent transactions will
> unpredictably fail instead of waiting.
If we were to change ALTER TABLE so it takes a session lock rather
than a normal lock, then we can commit the change and then wait until
the relcache invalidations have been received before we release the
lock. That way we would be able to avoid the concurrent issues you
describe.
Or alternatively we could just re-scan if we can't find a valid row
when building the cache. We have time in the failure path...
Do you have stomach for any this in 9.1?
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
