Re: Strategy for doing number-crunching

On Wed, Jan 4, 2012 at 3:11 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Matthew Foster <matthew(dot)foster(at)noaa(dot)gov> writes:
> > On Wed, Jan 4, 2012 at 10:48 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> >> Matthew Foster <matthew(dot)foster(at)noaa(dot)gov> writes:
> >>> We have a database with approximately 130M rows, and we need to produce
> >>> statistics (e.g. mean, standard deviation, etc.) on the data. Right
> now,
> >>> we're generating these stats via a single SELECT, and it is extremely
> >>> slow...like it can take hours to return results.
>
> >> What datatype are the columns being averaged? If "numeric", consider
> >> casting to float8 before applying the aggregates. You'll lose some
> >> precision but it'll likely be orders of magnitude faster.
>
> > The data are type double.
>
> Hmm. In that case I think you have some other problem that's hidden in
> details you didn't show us. It should not take "hours" to process only
> 130M rows. This would best be taken up on pgsql-performance; please see
> http://wiki.postgresql.org/wiki/Slow_Query_Questions
>
> regards, tom lane
>

Tom,

I think you are absolutely right. Some additional testing, with the
arithmetic removed from the queries, still shows very slow performance.

I'll do some more digging, and perhaps take this to the performance list.
Thanks for your advice!

Matt