Re: can't get rid of unnesesary SORT step in explain plan for hash join - Mailing list pgsql-sql

From Alexey Nalbat
Subject Re: can't get rid of unnesesary SORT step in explain plan for hash join
Date
Msg-id 01051413453300.13379@workshop.price.ru
Whole thread Raw
In response to Re: can't get rid of unnesesary SORT step in explain plan for hash join  (Tom Lane <tgl@sss.pgh.pa.us>)
List pgsql-sql
On Sun, 13 May 2001, Tom Lane wrote:
> Alexey Nalbat <alexey@price.ru> writes:
> > So, my question is: how can I get rid of this unnesesary "Sort" step
> > in the execution plan for hash join?
> 
> You can't, because it's not unnecessary.  Hash join doesn't promise
> to produce its outputs in any particular order.  But the Unique
> filter needs to see its inputs in order by the fields being made
> unique.

Hello.

Tom, thanks for your answer. But I am not agree with it.

Here is the query of my interest:
+ select distinct c_id, m_id from products, ( select r_id from resellers where r_name like 'a%' ) as temp where m_id =
123and products.r_id = temp.r_id order by m_id, c_id;
 
+ 
+ Unique  (cost=16.83..16.83 rows=1 width=12)
+   ->  Sort  (cost=16.83..16.83 rows=1 width=12)
+     ->  Hash Join  (cost=8.16..16.82 rows=1 width=12)
+       ->  Index Scan using products_mcr on products  (cost=0.00..8.14 rows=10 width=8)
+       ->  Hash  (cost=8.14..8.14 rows=10 width=4)
+         ->  Index Scan using resellers_n on resellers  (cost=0.00..8.14 rows=10 width=4)

While executing this query postgres at first creates hash on table "resellers",
then get from index "products_mcr" for rows with "m_id=123" already ordered (!!!)
pairs "c_id,r_id", for each that pair it checks join condition using hash. If postgers
behaves like this then the result of this hash join is already sorted, and extra
"Sort" is not needed.

In fact query without "distinct" and "order by" clauses returns ordered values
of "c_id". I mean this query:
+ select c_id from products, ( select r_id from resellers where r_name like 'a%' ) as temp where m_id = 123 and
products.r_id= temp.r_id" returns ordered values of "c_id;
 
+ 
+ Hash Join  (cost=8.16..16.82 rows=1 width=12)
+   ->  Index Scan using products_mcr on products  (cost=0.00..8.14 rows=10 width=8)
+   ->  Hash  (cost=8.14..8.14 rows=10 width=4)
+     ->  Index Scan using resellers_n on resellers  (cost=0.00..8.14 rows=10 width=4)

And one more argument. We are now transferring our DB from Oracle to Postgres,
and in oracle this query does not have "Sort" in explain_plan, it has only "Unique":
+ select distinct c_id from products where m_id = 123 and r_id in ( select r_id from resellers where r_name like 'a%' )
orderby c_id
 
+ 
+ SORT, UNIQUE
+    HASH JOIN,
+       INDEX, RANGE SCAN PRODUCTS_MCR
+       TABLE ACCESS, FULL RESELLERS

So I suppose, that this "Sort" step can be removed from the execution plan. What
do you think about this?

Thanks a lot.

-- 

WBR, Alexey Nalbat


pgsql-sql by date:

Previous
From: "David D. Kilzer"
Date:
Subject: Re: Using ORDER BY with AGGREGATE/GROUP BY in a SELECT statement
Next
From: "Sylte"
Date:
Subject: Auto incrementing an integer