Re: Auto-tuning work_mem and maintenance_work_mem

* Andres Freund (andres(at)2ndquadrant(dot)com) wrote:
> On 2014-02-16 21:26:47 -0500, Robert Haas wrote:
> > I don't think anyone objected to increasing the defaults for work_mem
> > and maintenance_work_mem by 4x, and a number of people were in favor,
> > so I think we should go ahead and do that. If you'd like to do the
> > honors, by all means!
>
> Actually, I object to increasing work_mem by default. In my experience
> most of the untuned servers are backing some kind of web application and
> often run with far too many connections. Increasing work_mem for those
> is dangerous.

And I still disagree with this- even in those cases. Those same untuned
servers are running dirt-simple queries 90% of the time and they won't
use any more memory from this, while the 10% of the queries which are
more complicated will greatly improve.

> > I don't really know about cpu_tuple_cost. Kevin's often advocated
> > raising it, but I haven't heard anyone else advocate for that. I
> > think we need data points from more people to know whether or not
> > that's a good idea in general.
>
> FWIW It's a good idea in my experience.

I'm in favor of this also but I'm also in the camp of "gee, more data
would be nice".

Thanks,

Stephen