On 2015-09-30 19:36:22 +0300, Heikki Linnakangas wrote:
> On 09/30/2015 04:31 AM, Chris Winslett wrote:
> >I'm orchestrating Postgres to behave as a leader-follower cluster. I've run
> >into issues when I am scaling down a connection count for a cluster
> >(scaling up is fine -- scaling down results in fatal errors). I use an
> >open source tool I've written to orchestrate the cluster called Governor (
> >http://github.com/compose/governor). It would work if I weren't running
> >hot_standby, but alas…I'm running with it.
> >
> >I found the code which throws the fatal is actually a pre-flight test for
> >hot_standby written in 2009 (i.e. battle tested):
> >
>
>https://github.com/postgres/postgres/blob/efc16ea520679d713d98a2c7bf1453c4ff7b91ec/src/backend/access/transam/xlog.c#L5312-L5321
> >
> >I've tested changing this value from a FATAL to a WARN. I've compiled and
> >tested my scenario and all appears to be correct:
s/correct/not fail directly/
> >https://github.com/compose/postgres/commit/2bdf6b36821987aadb401e1b8590ecc5b02126d8
The new phrasing certainly isn't accurate since it can actually prevent
replay from continuing.
> It will work, until you actually have more active transactions in the master
> than you have slots available in the standby.
There's a bunch of additional cases where it'll possibly fail. As the
actual maximum number of locks is determined by both max_locks_per_xact
and max_connections (and some more, but whatever) you can even run into
troubles with just one transaction on the primary. A case where that
frequently happens is when you restore a (partial) database dump because
that can easily cause a large number of exclusive locks to be acquired -
which will be WAL logged.
Greetings,
Andres Freund
