Slow window function query with big table

Question

I'm doing some performance testing on a new DB design on PostgreSQL 9.4rc1 and I'm seeing some pretty slow queries using window functions. Here is my table setup:

CREATE TABLE player_stat (
 player_id VARCHAR(200) NOT NULL,
 stat_id BIGINT NOT NULL,
 value BIGINT NOT NULL DEFAULT 0,
 last_updated TIMESTAMP WITH TIME ZONE NOT NULL,
 last_active TIMESTAMP WITH TIME ZONE DEFAULT NULL,
 CONSTRAINT player_stat_pk PRIMARY KEY (player_id, stat_id),
 CONSTRAINT player_stat_fk1 FOREIGN KEY(stat_id) REFERENCES stat (id)
);
CREATE INDEX player_stat_stat_value_player_desc
 ON player_stat (stat_id, value DESC, player_id ASC);

I've inserted 30 million rows into this table split among 3 stats:

INSERT INTO player_stat (player_id, stat_id, value, last_updated) SELECT x.id, 1, x.v, now() FROM (SELECT generate_series(1,10000000) as id, trunc(random() * (1900-1200) + 1200) as v) AS x;
INSERT INTO player_stat (player_id, stat_id, value, last_updated) SELECT x.id, 2, x.v, now() FROM (SELECT generate_series(1,10000000) as id, trunc(random() * (1900-1200) + 1200) as v) AS x;
INSERT INTO player_stat (player_id, stat_id, value, last_updated) SELECT x.id, 3, x.v, now() FROM (SELECT generate_series(1,10000000) as id, trunc(random() * (1900-1200) + 1200) as v) AS x;

Then I try to rank the players for a given stat (EDIT):

SELECT * FROM 
( SELECT player_id
 , rank() OVER (ORDER BY value DESC, player_id ASC) as rank 
 FROM player_stat 
 WHERE stat_id = 1
) as t 
WHERE rank <= 20 
ORDER BY rank ASC;

This query takes about 5.5 seconds to return. Running Explain on it returns the following:

"Sort (cost=1167612.28..1176082.26 rows=3387993 width=15) (actual time=9726.132..9726.135 rows=20 loops=1)"
" Sort Key: t.rank"
" Sort Method: quicksort Memory: 25kB"
" -> Subquery Scan on t (cost=0.56..684349.57 rows=3387993 width=15) (actual time=0.080..9726.116 rows=20 loops=1)"
" Filter: (t.rank <= 20)"
" Rows Removed by Filter: 9999980"
" -> WindowAgg (cost=0.56..557299.83 rows=10163979 width=15) (actual time=0.077..8351.124 rows=10000000 loops=1)"
" -> Index Only Scan using player_stat_stat_value_player_desc on player_stat (cost=0.56..379430.20 rows=10163979 width=15) (actual time=0.054..2319.007 rows=10000000 loops=1)"
" Index Cond: (stat_id = 1)"
" Heap Fetches: 0"
"Planning time: 0.187 ms"
"Execution time: 9726.172 ms"

Is there any way can I speed this up? The time it takes seems to be growing linearly with the number of players in the table.

• user1822: Most probably unrelated, but why do you use a RC1 if there are already 4 bugfix release for 9.4?

Answer 1

Is there any way I can speed this up?

Yes. Don't use a varchar column for an integer number. Use integer or bigint if you burn that many IDs - much smaller in table and index and faster to process. Since you are ranking 10 million rows in your test, this is going to make a substantial difference.

(削除) player_id VARCHAR(200) NOT NULL, (削除ここまで)
player_id int NOT NULL,

Or a uuid if you must (I doubt that):

• Postgres uuid: Use as primary key, or in addition to SERIAL -for disconnected app-

Your query ranks 10 million rows. This is going to take some time, even when read from the index directly and no sort step.

Side note: If you bulk-insert rows first and add index and PK constraint (and FK constraint) after, that's going to be much faster, plus you get perfect indexes without bloat without running REINDEX or VACUUM FULL. Do make sure ANALYZE has been run on the table before testing performance, though.

What you didn't ask

.. but, going out on a limb here, what are probably looking for.

The EXPLAIN output reveals that you filter the top 20 rows: (t.rank <= 20). Your presented query does not show that. The query actually matching your EXPLAIN output would be:

SELECT * FROM (
 SELECT player_id
 , rank() OVER (ORDER BY value DESC, player_id ASC) AS rank
 FROM player_stat
 WHERE stat_id = 1
 ) t
WHERE t.rank <= 20;

Which can be improved dramatically:

SELECT row_number() OVER (ORDER BY value DESC, player_id ASC) AS rank
 , player_id
FROM player_stat
WHERE stat_id = 1
ORDER BY value DESC, player_id
LIMIT 20;

Explanation

• The important part for performance is the LIMIT clause in combination with ORDER BY matching the index: now the query reads exactly 20 rows from the top to the index, where it had to read 10000000 in your original version. We only use player_id and value, so we can still have an index-only scan. The rest is peanuts.

• That's all due to the sequence of events in a SELECT query: window functions are applied before LIMIT. Only if the sort order agrees, we don't have to consider the rest of the applicable 10000000 rows.

• We can use LIMIT 20 because the top 20 ranks are guaranteed to span no more than 20 rows. The PK on (player_id, stat_id) guarantees unique player_id per stat_id and since that is included in the ORDER BY, each rank is only assigned once - which also means we can use the slightly cheaper row_number() instead.

• Kyle: Thanks! You were right I must have copied the query over incorrectly. I fixed it in the original post. I didn't realize that window functions were applied before limits, That was the key to speed this up significantly.
• ypercubeᵀᴹ: @Erwin can't the last query be written without a derived table, using both window function and LIMIT in the same level? (and with the same efficiency I mean)
• Erwin Brandstetter: @ypercube: You are right, in this case we don't even need the subquery and still get the performance benefit. Simplified accordingly.
• ypercubeᵀᴹ: OK, I thought so but hadn't time to test. The crucial difference is I guess that the ORDER BY in the window and ORDER BY are identical. (also edit the "We need the subselect...")
• Erwin Brandstetter: @Kyle: You may be interested in the further simplification. As long as the outer sort order agrees with sort order needed for the index (and the one in the window function), we don't need the subquery.