Joshua D. Drake wrote:
>
> On 08/18/2015 09:41 AM, Alvaro Herrera wrote:
> >
> >Alvaro Herrera wrote:
> >
> >>One thing to look at is the rate of WAL generation for a set number of
> >>transactions. Maybe the later releases are generating more WAL due to
> >>multixacts, for instance (prior to 9.3 these weren't wal-logged.)
> >
> >Also try 9.5alpha2, wherein bug #8470 is fixed, which is a big
> >performance regression in 9.3 and 9.4.
>
> Is there a reason we didn't backport this? (A point to a thread is just
> fine).
There's a thread somewhere, yeah, but the meat of it is that there were
other fixes for multixacts going in at the same time and we didn't want
people to be unable to upgrade to a version with those fixes if it later
turned out that the fix for #8470 had bugs.
Some time later, I remembered that I had developed two fixes for #8470,
one of which was cleaner but more invasive, and got pushed to 9.5;
another fix was much more self-contained but uglier and perhaps we could
consider it for 9.3 and 9.4. I'm out of resources for further multixact
fixes myself (therefore I'm very grateful to Andres, Thomas and Robert
who shouldered the burden of the last round of fixes), so I didn't
pursue this. If anyone else wants to pursue it, I can provide the
patch, which evolved somewhat from the last version of it I posted.
(Then again, the last time I asked for testing help on the topic of
#8470 nobody volunteered even though several people said they were
affected by the bug. Go figure.)
Cheers
--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services