Message-ID: <492DB289.8020703@archonet.com>
Date: Wed, 26 Nov 2008 20:33:13 +0000
From: Richard Huxton <dev@archonet.com>
User-Agent: Thunderbird 2.0.0.18 (X11/20081105)
MIME-Version: 1.0
To: Andrus <kobruleht2@hot.ee>
CC: mweilguni@sime.com, pgsql-performance@postgresql.org
Subject: Re: Increasing pattern index query speed
References: <gg9pg5$1edv$1@news.hub.org> <492A7591.3000700@archonet.com>
	<C2F601D897194339A0D2327188C781D6@andrusnotebook>
	<492D27C8.7090206@archonet.com>
	<4936D68EF8C84087BBB9212C4019D5D8@andrusnotebook>
	<492D9E78.9080902@archonet.com>
	<534B7B28155C4275BD2B9AA8F3B656DD@andrusnotebook>
In-Reply-To: <534B7B28155C4275BD2B9AA8F3B656DD@andrusnotebook>
Content-Type: text/plain; charset=windows-1257
Content-Transfer-Encoding: 7bit

Andrus wrote:
> Richard,
> 
>>> Results are provided in bottom of the message to which you replied.
>>
>> No - the explains there were contrasting a date test BETWEEN versus =.
> 
> I changed rid.toode statitics target to 100:
> 
> ALTER TABLE firma2.rid ALTER COLUMN toode SET STATISTICS 100;
> analyze firma2.rid;
> 
> Analyze takes 3 seconds and testcase rans fast.
> I'm planning to monitor results by looking log file for queries which
> take longer than 10 seconds.

Sensible. I don't know if 10 seconds is the right value for your
database, but there will be a point that filters out most of your
traffic but still gives enough to find problems.

> Do you still need results ?
> If yes, which query and how many times should I run?

If changing the statistics seems to help, you're not going to want to go
back just to repeat tests.

>> Ah, I think I understand. The test case was *missing* this clause.
> 
> I added  this clause to testcase. Also added char(70) colums containing
> padding characters to all three tables. Cannot still reproduce this issue
> in testcase in fast devel 8.3 notebook.
> In testcase order_products contains product_id values in a very regular
> order, maybe this affects the results. No idea how to use random() to
> generate random
> products for every order.

Ideally you don't even want random products. You want a distribution of
products that matches the same "shape" as you have in your production
database.

-- 
  Richard Huxton
  Archonet Ltd