Re: Postgres not using indexes

From: "Kevin Grittner" <Kevin(dot)Grittner(at)wicourts(dot)gov>
To: "pgsql-bugs(at)postgresql(dot)org" <pgsql-bugs(at)postgresql(dot)org>, "Lawrence Cohan" <LCohan(at)web(dot)com>
Subject: Re: Postgres not using indexes
Date: 2011-03-30 20:11:44
Message-ID: 4D934830020000250003BFB4@gw.wicourts.gov
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-bugs

Lawrence Cohan <LCohan(at)web(dot)com> wrote:

> looks like we will need to change at least the two values below
> and maybe play with work_mem to see if it solves our issues.

You will probably get better throughput by bumping up
shared_buffers to the recommended setting, but beware of "stalls" in
query processing at checkpoint time. If that happens you want to
make the background writer more aggressive and/or back off on
shared_memory, so that there isn't such a glut of dirty pages to
write during a checkpoint.

I think even the recommended setting for effective_cache size is on
the low side. This one affects how expensive the optimizer thinks
index usage will be, so given your current problem this is probably
important to raise. I add up shared_buffers and what free tells me
is cached space is after PostgreSQL has been running a while. That
usually winds up being 1GB to 2GB less than total memory on our
machines, so actually, I usually just start there.

We usually need to reduce random_page_cost to get good plans. For a
fully-cached database you may want to reduce both seq_page_cost and
random_page_cost to equal numbers around 0.05. With partial
caching, we often leave seq_page_cost alone and reduce
random_page_cost to 2. YMMV.

The setting for work_mem can be tricky, especially with 1200
connections configured. Each connection may be using one or more
allocations of work_mem at the same time. Which leads to the
question of why you have 1200 connections configured. You are
almost always better off using a connection pooler to limit this to
something on the order of twice your CPU cores plus your effective
spindle count. Tomcat has a very good connection pooler built in,
as do many other products. There are also good external poolers,
like pgpool and pgbouncer.

With a reasonable amount of RAM you're almost always better off
bumping wal_buffers to 32MB.

> The only issue is that we are running a 24/7 web site against the
> db and if we need to restart PG for the changes to take place we
> will need to wait for a downtime before any changes can be made.

Some of these can be set per user with ALTER ROLE. New connections
would then start using the new settings with no down time.

-Kevin

In response to

Responses

Browse pgsql-bugs by date

  From Date Subject
Next Message Lawrence Cohan 2011-03-30 20:49:48 Re: Postgres not using indexes
Previous Message Kevin Grittner 2011-03-30 19:49:07 Re: Postgres not using indexes