Question

This problem is causing me a headache for a while. I’ve a PostgreSQL 8.4 database that consists of only one table containing more than 4.000.000 records. This table is structured as follows:

CREATE TABLE metadata (
  id serial NOT NULL,
  "value" text NOT NULL DEFAULT ''::text,
  segment_box box NOT NULL DEFAULT box(point(((-9223372036854775808)::bigint)::double precision, (1)::double precision), point((9223372036854775807::bigint)::double precision, ((-1))::double precision)),
  CONSTRAINT metadata_pk PRIMARY KEY (id)
)

CREATE INDEX metadata_segment_box_ix
  ON metadata
  USING gist
  (segment_box);

CREATE INDEX metadata_tag_value_ix
  ON metadata
  USING btree
  (value);

The table contains segments (in time), represented as rectangular boxes. These segments are annotated using the “value” column.

The kind of queries I would like to perform on the database try to find all segments with a specified value that is contained within a certain window. A query that successfully achieves this is:

SELECT * FROM (SELECT * FROM metadata WHERE value='X') a, 
(SELECT * FROM metadata WHERE AND value='Y') b 
WHERE a.segment_box <-> b.segment_box <= 3000

But, as you probably noticed, this query cannot be performed efficiently by the database. The cartesian product of sub-queries a and b is becoming really large. Is there a way to perform these queries more efficiently? I can imagine some sort of sliding window approach would do the trick. Maybe something like the following:

SELECT *, rank() OVER (
PARTITION BY "value" ORDER BY (segment_box[1])[0], (segment_box[0])[0]
) FROM metadata WHERE value='X' OR value='Y'

Update:
One of the things I tried after posting this question is creating a custom function in Postgres. I tried:

CREATE OR REPLACE FUNCTION within_window(size bigint DEFAULT 0)
  RETURNS setof metadata AS
$BODY$DECLARE
  segment RECORD;
  neighbour RECORD;
  newwindow box;
BEGIN
  FOR segment IN (
    SELECT * FROM metadata WHERE value='X' OR value='Y' 
      ORDER BY (segment_box[1])[0], (segment_box[0])[0]
  ) LOOP
    newwindow := box(segment.segment_box[0], 
      point((((segment.segment_box[1])[0]) + size), (segment.segment_box[1])[1]));
    FOR neighbour IN (
      SELECT DISTINCT ON (metadata_id) * FROM metadata WHERE value='X' OR value='Y') 
        AND segment_box &< newwindow
        AND segment_box &> newwindow 
    ) LOOP
      RETURN NEXT neighbour;
    END LOOP;
  END LOOP;
END;$BODY$
  LANGUAGE plpgsql;

However, this function is as slow as the basic solution I described above because of the subquery that must be performed many times. Any other thoughts on this??

Answer 1

I solved the problem myself with a kind of sweep line algorithm. Only one query is performed. I use a cursor to sweep back and forth over the query’s resultset. The resulting algorithm works as follows:

CREATE OR REPLACE FUNCTION within_window(size bigint DEFAULT 0)
  RETURNS setof metadata AS
$BODY$DECLARE 
crsr SCROLL CURSOR FOR (SELECT * FROM metadata WHERE value='X' OR value='Y' ORDER BY (segment_box[1])[0], (segment_box[0])[0]);
rc RECORD;
rcc RECORD;
crsr_position int;
last_crsr int;
BEGIN
    OPEN crsr;
    crsr_position := 0;
    LOOP FETCH NEXT FROM crsr INTO rc;
        IF NOT FOUND THEN
            EXIT;
        END IF;
        last_crsr := crsr_position;
        LOOP FETCH NEXT FROM crsr INTO rcc;
            IF NOT FOUND THEN
                EXIT;
            ELSEIF 
                rcc.segment_box &< box(rc.segment_box[0], point((((rc.segment_box[1])[0]) + size), (rc.segment_box[1])[1])) AND
                rcc.segment_box &> box(rc.segment_box[0], point((((rc.segment_box[1])[0]) + size), (rc.segment_box[1])[1]))
            THEN
                RETURN NEXT rcc;
            ELSE 
                EXIT;
            END IF;
        END LOOP;
        crsr_position := last_crsr + 1;
        MOVE ABSOLUTE crsr_position FROM crsr;
    END LOOP;
    CLOSE crsr;
END;$BODY$
  LANGUAGE plpgsql;

Using this function the query only needs 476 ms instead of 6+ minutes (on the 4+ million row database)!