Re: postgresql recommendation memory

On Mon, Nov 11, 2013 at 09:14:43AM -0700, Scott Marlowe wrote:
> On Mon, Nov 11, 2013 at 1:09 AM, Евгений Селявка <evg(dot)selyavka(at)gmail(dot)com> wrote:
> > Scott hi, i calculate all of my jdbc pool size. Maximum is 300 connections
> > from components wich use jdbc. I don't think that this is a good idea use
> > pgbouncer, because our application using spring framework which using jdbc
> > and prepared statement. I try to talk with our developer about disabling
> > prepared statement in this framework, they don't want do this. Thats why i
> > will try to upgrade HW and buy CPU with more core as you say based on
> > formula 3-4xcore. But most of this connection is idle. This is a web based
> > app not a datawarehouse, thats why all this connection is lightwear.
> >
> > About my db freeze i set this kernel parameter:
> > echo 1048576 > /proc/sys/vm/min_free_kbytes
> > echo 80 > /proc/sys/vm/vfs_cache_pressure
> >
> > And my freeze intervals is steel smaller. I try to dig deeper.
>
> well you can hopefully reduce connections from jdbc pooling then. The
> fact that the connections are idle is good.
>
> The problem you run into is what happens when things go into
> "overload" I.e. when the db server starts to slow down, more of those
> idle connections become not idle. If all 300 are then waiting on the
> db server, it will slow to a crawl and eventually fall over.
>
+1 I would definitely encourage the use of pgbouncer to map the 300 connections
to a saner number that your DB can actually handle. We had a similar problem
and very, very occasionally the server would "lockup". Once we put the
resource management pooler in place, performance has been the same best-case
and much, much better worse-case and NO lockups.

Regards,
Ken