On Fri, 2008-11-14 at 09:00 -0600, Dan Armbrust wrote:
> > There was concurrent access to the table during VACUUMing, so the long
> > delay is explainable as long waits for cleanup lock, plus probably
> > thrashing the cache with bloated indexes. The CPU overhead per row seems
> > OK. We should instrument the wait time during a VACUUM and report that
> > also.
> Is that a guess? Or something you can tell from the log above?
The number of row versions in each index was different after vacuuming.
That tells me some writes occurred and I inferred from that that other
read-only activity occurred as well. Reads or writes will slow down a
VACUUM.
Perhaps you have vacuum_cost_delay set also?
--
Simon Riggs www.2ndQuadrant.com
PostgreSQL Training, Services and Support