On Thu, Nov 14, 2013 at 6:28 PM, KONDO Mitsumasa
<kondo.mitsumasa@lab.ntt.co.jp> wrote:
> It is confirmation just to make sure, does "this patch" mean my patch? I
> agree with you about not adding another lock implementation. It will becomes
> overhead.
Yes, I referred to your patch. I don't want to go down this road,
because aggregation and constructing a timeline feels like the job of
another tool. I am concerned about local minima and maxima. Even with
the ability to reset min/max independently, you can't do so for each
entry individually. And this approach won't scale to a histogram or
more sophisticated ways of characterizing distribution, particularly
not multiplicatively for things other than execution time (blocks hit
and so on) - that spinlock needs to be held for very little time
indeed to preserve pg_stat_statements current low overhead.
As I said above, lets figure out how to have your tool or a similar
tool acquire snapshots inexpensively and frequently instead. That is a
discussion I'd be happy to have. IMO pg_stat_statements ought to be as
cheap as possible, and do one thing well - aggregate fixed-unit costs
cumulatively.
--
Peter Geoghegan