On Wed, Jan 29, 2014 at 6:06 AM, Andrew Dunstan <andrew@dunslane.net> wrote:
> mportance is in the eye of the beholder. As far as I'm concerned, min and
> max are of FAR less value than stddev. If stddev gets left out I'm going to
> be pretty darned annoyed, especially since the benchmarks seem to show the
> marginal cost as being virtually unmeasurable. If those aren't enough for
> you, perhaps you'd like to state what sort of tests would satisfy you.
I certainly agree that stddev is far more valuable than min and max.
I'd be inclined to not accept the latter on the grounds that they
aren't that useful.
So, am I correct in my understanding: The benchmark performed here [1]
was of a standard pgbench SELECT workload, with no other factor
influencing the general overhead (unlike the later test that queried
pg_stat_statements as well)? I'd prefer to have seen the impact on a
32 core system, where spinlock contention would naturally be more
likely to appear, but even still I'm duly satisfied.
There was no testing of the performance impact prior to 6 days ago,
and I'm surprised it took Mitsumasa that long to come up with
something like this, since I raised this concern about 3 months ago.
[1] http://www.postgresql.org/message-id/52E10E6A.5060708@lab.ntt.co.jp
--
Peter Geoghegan