Message-ID: <1097.85.160.40.238.1275639203.squirrel@sq.gransy.com>
In-Reply-To: <AANLkTinIM12lzk5oM0Aq2yYHdzr-ESJk7LANAeqW2Itu@mail.gmail.com>
References: <AANLkTinIM12lzk5oM0Aq2yYHdzr-ESJk7LANAeqW2Itu@mail.gmail.com>
Date: Fri, 4 Jun 2010 10:13:23 +0200 (CEST)
Subject: Re: slow query
From: tv@fuzzy.cz
To: "Anj Adu" <fotographs@gmail.com>
Cc: pgsql-performance@postgresql.org
User-Agent: SquirrelMail/1.4.10a
MIME-Version: 1.0
Content-Type: text/plain;charset=iso-8859-2
Content-Transfer-Encoding: 8bit
Importance: Normal

> I am reposting as my original query was mangled
>
> The link to the explain plan is here as it does not paste well into
> the email body.
>
> http://explain.depesz.com/s/kHa
>
>
> The machine is a 2 cpu quad core 5430 with 32G RAM and 6x450G 15K
> single raid-10 array
>
> 1G work_mem
> default_statistics_target=1000
> random_page_cost=1

Are you sure it's wise to set the work_mem to 1G? Do you really need it?
Don't forget this is not a 'total' or 'per query' - each query may
allocate multiple work areas (and occupy multiple GB). But I guess this
does not cause the original problem.

The last row 'random_page_cost=1' - this basically says that reading data
by random is just as cheap as reading data sequentially. Which may result
in poor performance due to bad plans. Why have you set this value?

Sure, there are rare cases where 'random_page_cost=1' is OK.

>
> I am curious why the hash join takes so long. The main table
> dev4_act_dy_fact_2010_05_t has 25 million rows. The table is
> partitioned into 3 parts per month. Remaining tables are very small (
> < 1000 rows)

Well, the real cause that makes your query slow is the 'index scan' part.

Index Scan using dev4_act_dy_fact_2010_05_t3_thedate on
dev4_act_dy_fact_2010_05_t3 a (cost=0.00..94041.89 rows=204276 width=60)
(actual time=164533.725..164533.725 rows=0 loops=1)

The first thing to note here is the difference in expected and actual
number of rows - the planner expects 204276 but gets 0 rows. How large is
this partition?

Try to analyze it, set the random_page_cost to something reasonable (e.g.
4) and try to run the query again.

Tomas