On Tue, 22 Apr 2008, David Wilson wrote:
> My guess at this point is that I'm just running into index update
> times and checkpoint IO. The only thing that still seems strange is
> the highly variable nature of the COPY times- anywhere from <1.0
> seconds to >20 seconds, with an average probably around 8ish.
Have you turned on log_checkpoints to see whether those are correlated
with the slow ones? Given that you've had an improvement by increasing
checkpoint_segments, it's not out of the question to think that maybe
you're still getting nailed sometimes during the more stressful portions
of the checkpoint cycle (usually right near the end). The 1 second ones
just might just happen to be ones that start just as the previous
checkpoint finished. To make lining those up easier, you might turn on
logging of long statements with log_min_duration_statement to see both
bits of data in the same log file. That might get you some other
accidental enlightenment as well (like if there's some other statement
going on that's colliding with this load badly).
This is a bit out of my area, but after reading the rest of this thread I
wonder whether raising the default_statistics_target parameter a bit might
reduce the instances of bad plans showing up.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD