On Tue, Dec 03, 2013 at 04:37:58PM +0100, Andres Freund wrote:
> On 2013-12-03 10:29:54 -0500, Noah Misch wrote:
> > Sorry, my original report was rather terse. I speak of the scenario exercised
> > by the second permutation in that isolationtester spec. The multixact is
> > later than VACUUM's cutoff_multi, so 9.3.1 does not freeze it at all. 9.3.2
> > does freeze it to InvalidTransactionId per the code I cited in my first
> > response on this thread, which wrongly removes a key lock.
>
> That one is clear, I was only confused about the Assert() you
> reported. But I think I've explained that elsewhere.
>
> I currently don't see fixing the errorneous freezing of lockers (not the
> updater though) without changing the wal format or synchronously waiting
> for all lockers to end. Which both see like a no-go?
Not fixing it at all is the real no-go. We'd take both of those undesirables
before just tolerating the lost locks in 9.3.
The attached patch illustrates the approach I was describing earlier. It
fixes the test case discussed above. I haven't verified that everything else
in the system is ready for it, so this is just for illustration purposes.
--
Noah Misch
EnterpriseDB http://www.enterprisedb.com