On Thu, Nov 5, 2015 at 10:38 AM, Julian v. Bock <bock@openit.de> wrote:
> After doing some debugging it seems that the backend processes are
> spending most of their time in spinlocks in TransactionIdIsInProgress()
> trying to get a lock on ProcArrayLock. This function is called more
> often (when processing searches) the longer the replication transaction
> is running.
This is a known issue and is fixed for 9.6 in commit 8a7d0701814a4e.
I don' t know why you can't reproduce it without Slony, though. I
know that hot standby gets bit by this problem particularly hard
because it sometimes refrains from setting hint bits as aggressively
as it could. I would not think that Slony would have that problem,
though.
Cheers,
Jeff