Re: Simply join in PostrgeSQL takes too long - Mailing list pgsql-performance

From Rod Taylor
Subject Re: Simply join in PostrgeSQL takes too long
Date
Msg-id 1083155014.30065.189.camel@jester
Whole thread Raw
In response to Re: Simply join in PostrgeSQL takes too long  (Manfred Koizar <mkoi-pg@aon.at>)
Responses Re: Simply join in PostrgeSQL takes too long  (Manfred Koizar <mkoi-pg@aon.at>)
List pgsql-performance
> Rod, he has a hierarchy of genres.  Genre 1 has 6379 child genres and a
> book can be in more than one of these.

     bookgenres.genre_id = genre_children.genre_child_id AND
     genre_children.genre_id = 1

I see, sorry. I didn't notice the genre_child_id in the where clause.
First glance had them all as genre_id.

When I run into this I usually create a 3rd table managed by triggers
that would relate the book to all genre entries. Insert takes a little
longer, but the selects can still be very quick.

The below plpgsql forces the kind of algorithm we wish the planner could
choose. It should be fairly quick irregardless of dataset.


CREATE OR REPLACE FUNCTION book_results(numeric) RETURNS SETOF numeric
AS
'
DECLARE
  v_genre ALIAS FOR $1;
  v_limit integer = 10;
  t_rows RECORD;
  v_transmitted integer = 0;

  v_transmitted_values numeric[] = ARRAY[1];

BEGIN
  FOR t_rows IN SELECT book_id
                  FROM bv_bookgenres AS b
                  JOIN bv_genre_children AS g ON (b.genre_id =
g.genre_child_id)
                 WHERE g.genre_id = v_genre
                  LOOP

    -- If this is a new value, transmit it to the end user
    IF NOT t_rows.book_id = ANY(v_transmitted_values) THEN
      v_transmitted_values := array_append(v_transmitted_values,
t_rows.book_id);
      v_transmitted := v_transmitted + 1;
      RETURN NEXT t_rows.book_id;
    END IF;

    EXIT WHEN v_transmitted >= v_limit;
  END LOOP;

  RETURN;
END;
' LANGUAGE plpgsql;

EXPLAIN ANALYZE SELECT * FROM book_results(1);
SELECT * FROM book_results(1);


pgsql-performance by date:

Previous
From: Bruno Wolff III
Date:
Subject: Re: [JDBC] is a good practice to create an index on the
Next
From: Tom Lane
Date:
Subject: Re: Join problem