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

The following bug has been logged online:

Bug reference:      5165
Logged by:          assaf
Email address:      assaf_lehr@yahoo.com
PostgreSQL version: 8.37
Operating system:   linux
Description:        Poor performance with Left-join where right side does
not exist
Details:

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
--------------
I have indexes on all columns and thier permutations.
I needed merge-join here and expected brief results ,as it is easy to find
not-nulls running on both indexes.
My DB is quite big (20M items) and the result was supposed to be 0.5M items.
I stopped waiting after 18 hours.

There is workaround , select B.idB from B where category=202 and idB not in
(select distinct idB from A where category=202);
It finishes in good time(10min) , but it`s totally a waste to use subplan
here.

"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.

            regards, tom lane

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: =A0 =A0 =A0 =A0Poor performance with Left-join where right =
side does
>> not exist
>
> 8.4 might be smarter about this case for you. =A0It's hard to tell for su=
re
> 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

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

assaf lehr <assaf_lehr@yahoo.com> wrote:
> From: Robert Haas <robertmhaas@gmail.com>

>> 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.

> let me explain it better.
> [hand-waving and pseudo-code]

People would like to help you, but you're not providing enough
information to let them.  If you can show the actual tables involved
in a slow query, the actual query, and the output you get when you run
the query with EXPLAIN ANALYZE in front of the query (or just EXPLAIN
if the EXPLAIN ANALYZE can't run to completion in an acceptable amount
of time); then people will probably be able to help more.

It's also good to include the output of SELECT version(); and
information on the OS, hardware, and the contents of your
postgresql.conf file (excluding comments).

-Kevin

"Kevin Grittner" <Kevin.Grittner@wicourts.gov> writes:
> assaf lehr <assaf_lehr@yahoo.com> wrote:
>> let me explain it better.
>> [hand-waving and pseudo-code]

> People would like to help you, but you're not providing enough
> information to let them.

Well, the really important point is that this query is using a
construction that was specifically improved in 8.4.  The advice
is still going to be the same: see if it isn't better with 8.4.

            regards, tom lane

ok. thanks, when my project will move to 8.4, I `ll re-do this test.




________________________________
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Kevin Grittner <Kevin.Grittner@wicourts.gov>
Cc: Robert Haas <robertmhaas@gmail.com>; assaf lehr <assaf_lehr@yahoo.com>; pgsql-bugs@postgresql.org
Sent: Thu, November 5, 2009 5:21:28 PM
Subject: Re: [BUGS] BUG #5165: Poor performance with Left-join where right side does not exist

"Kevin Grittner" <Kevin.Grittner@wicourts.gov> writes:
> assaf lehr <assaf_lehr@yahoo.com> wrote:
>> let me explain it better.
>> [hand-waving and pseudo-code]

> People would like to help you, but you're not providing enough
> information to let them.

Well, the really important point is that this query is using a
construction that was specifically improved in 8.4.  The advice
is still going to be the same: see if it isn't better with 8.4.

            regards, tom lane