Re: BUG #5165: Poor performance with Left-join where right side does not exist

I did not not check in 8.4 and maybe it is solved there, but let me explain it better.
My particular case is a Join of two tables where the other table result should be "is null".
Using explain, I found out that nested-loop and hash-join merge algrothims are bad for me [both tables are quite big].
The merge-join algorithm looked like a great way to solve the problem , as both the table has the same ordered index ,
whichis just the thing the merge-join needs. 
The two tables and the slow join are:
A
category | idB | multiple-values

B
category | idB | multiple-values

--------------
select B.idB , A.idB from B left join A on B.idB = A.idB and
A.category=B.category
where A.idB is null [and A.category=202] limit 10
--------------

Mind you that if the last where is switched from "A.idB is null" to any regular check which can be done on the one row
(likecolC=5) , the merge join works just fine (and fast). 
In the "is null" case however, the performance is very poor. I would guess that there is a special case in the postgres
source-codefor the "is null" which has a different behavior. It can be, because unlike a single-row condition, is-null
needsa different behaviour. 


assaf










________________________________
From: Robert Haas <robertmhaas@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: assaf <assaf_lehr@yahoo.com>; pgsql-bugs@postgresql.org
Sent: Thu, November 5, 2009 3:07:41 AM
Subject: Re: [BUGS] BUG #5165: Poor performance with Left-join where right  side does not exist

On Wed, Nov 4, 2009 at 3:35 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> "assaf" <assaf_lehr@yahoo.com> writes:
>> PostgreSQL version: 8.37
>> Description:        Poor performance with Left-join where right side does
>> not exist
>
> 8.4 might be smarter about this case for you.  It's hard to tell for sure
> with so few details.

EXPLAIN output would be a good place to start, and EXPLAIN ANALYZE for
the queries that run quickly enough that you can let them run to
completion.

...Robert