Re: pg_upgrade and statistics

Bruce Momjian <bruce@momjian.us>

From: Bruce Momjian <bruce@momjian.us>

To: Tom Lane <tgl@sss.pgh.pa.us>

Cc: Kevin Grittner <Kevin.Grittner@wicourts.gov>, Peter Eisentraut <peter_e@gmx.net>, Daniel Farina <daniel@heroku.com>, Greg Stark <stark@mit.edu>, pgsql-hackers <pgsql-hackers@postgresql.org>

Date: 2012-03-15T01:21:24Z

Lists: pgsql-hackers

On Wed, Mar 14, 2012 at 08:26:06PM -0400, Tom Lane wrote:
> Bruce Momjian <bruce@momjian.us> writes:
> > Does anyone know how bad the queries will be with only one target?
> 
> Bad.  That cycle seems like largely a waste of time.  About the only
> thing it would do for you is ensure that relpages/reltuples are up to
> date, which seems like something we could possibly arrange for during
> the data import.

Well, it is also getting us the most common value, which seems useful.

> > I did see if vacuumdb --analyze-only was somehow being throttled by the
> > vacuum settings, but saw the drive at 100% utilization analying a 36GB
> > table on a 24GB RAM server, so it seems I/O bound.
> 
> I think it'd be good to explicitly set vacuum_cost_delay to 0 in the
> first pass, in the same way as you are forcing
> default_statistics_target, just in case somebody has a nondefault
> setting for that.  The second pass could probably be allowed to use some
> higher delay setting.

OK, I have now set vacuum_cost_delay=0 for the first vacuumdb
(target=1).

-- 
  Bruce Momjian  <bruce@momjian.us>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

  + It's impossible for everything to be true. +