Tom,
> I think you are probably looking at the same problem previously reported
> by Josh Berkus among others. Does the rate of context swaps shown by
> vmstat go through the roof when this happens? If you strace or ktrace
> one of the backends, do you see lots of semop()s and little else?
That would be interesting. Previously we've only demonstrated the problem on
long-running queries, but I suppose it could also affect massive concurrent
query access.
--
Josh Berkus
Aglio Database Solutions
San Francisco