Subject: Re: Searching for the cause of a bad plan
From: Csaba Nagy <nagy@ecircle-ag.com>
To: Simon Riggs <simon@2ndquadrant.com>
Cc: postgres performance list <pgsql-performance@postgresql.org>
In-Reply-To: <1190640429.4181.165.camel@ebony.site>
References: <1190369024.4661.173.camel@PCD12478>
	<6175.1190390913@sss.pgh.pa.us> <1190396352.4202.109.camel@ebony.site>
	<13644.1190397220@sss.pgh.pa.us> <1190399879.4202.127.camel@ebony.site>
	<27857.1190417434@sss.pgh.pa.us> <1190640429.4181.165.camel@ebony.site>
Content-Type: text/plain
Date: Mon, 24 Sep 2007 16:04:42 +0200
Message-Id: <1190642682.4661.247.camel@PCD12478>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit

On Mon, 2007-09-24 at 14:27 +0100, Simon Riggs wrote:
> Csaba, please can you copy that data into fresh tables, re-ANALYZE and
> then re-post the EXPLAINs, with stats data.

Well, I can of course. I actually tried to generate some random data
with similar record count and relations between the tables (which I'm
not sure I succeeded at), without the extra columns, but it was happily
yielding the nested loop plan. So I guess I really have to copy the
whole data (several tens of GB).

But from my very limited understanding of what information is available
for the planner, I thought that the record count estimated for the join
between table_a and table_b1 on column b should be something like

(estimated record count in table_a for value "a") * (weight of "b" range
covered by table_b1 and table_a in common) / (weight of "b" range
covered by table_a)

This is if the "b" values in table_a wouldn't be correlated at all with
the content of table_b2. The reality is that they are, but the planner
has no information about that.

I have no idea how the planner works though, so this might be totally
off...

I will copy the data and send the results (not promising though that it
will be today).

Cheers,
Csaba.