In our games we have a quartely league system. Every 3 months, the current season ends and all scorings of the current quarter are archived into a table called archive_league. Each entry represents a user and his score for a given league quarter. It comes with the following fields:

id - the id of an entry
uid - the id of a user
rounds - the number of games the user played in the given quarter
score - the score of the user for the given quarter
rank - the position of the player, descending by score, in the given quarter
date - the date of the quarter this entry represents

My goal is to go through all entries for a given date and assign the field "rank" for results. For example the player with the highest score should receive rank = 1, 2nd highest score rank = 2 and so on. If 2 players share the same score, they should receive the same rank (olympic scoring).


player x with score 528 receives rank 7
player y with score 528 receives rank 7
player z with score 529 receives rank 9

I am currently achieving this goal with this query:

UPDATE archive_league
        LEFT JOIN
                    COUNT(id) + 1
                    archive_league x
                    x.score > t.score
                        AND x.date = :quarter) AS new_rank
        archive_league t) AS temp USING (id) 
    rank = new_rank
    date = :quarter;

The problem is, however, with some hundreds of thousands of entries, this query runs for a couple of days, even though I have created indexes for the WHERE conditions. How can I optimize this query to run faster?


Materialize it

It may be more performant to make a real (materialized) temp table instead of the inline subquery.

-- Cleanup any old temporary table structure

-- Initialize a new temporary table. This will copy your same data types.
-- Structure only, no data.
  AS SELECT id, rank AS new_rank
     FROM archive_league WHERE 0=1;

-- Apply a PK index to the temp table for performance.

-- Compute ranks, store in temp table.
INSERT INTO temp_ranks
SELECT t.id,
      (SELECT COUNT(*) + 1
          FROM archive_league
          WHERE score > t.score
            AND date = :quarter) AS new_rank
    FROM archive_league t;

-- Apply to original table the materialized (temp) ranks.
UPDATE archive_league
LEFT JOIN temp_ranks USING (id) 
SET rank = new_rank
WHERE date = :quarter;

If we were using Oracle, I wouldn't expect this technique to make a difference (as Oracle's optimizer is pretty smart). However, MySQL's optimizer has some weak spots, one of which is that SELECT ... JOIN is pretty optimized, will choose best algorithm at run time (merge, hash, nested loop), yet UPDATE ... JOIN and DELETE ... JOIN lack the same optimization. Such an optimization is not impossible, but nobody has written the C code to make it so. If you're brilliant with programming databases in C, you are welcome to write an optimization and submit as a patch to MySQL (or MariaDB).

Reference from the manual https://dev.mysql.com/doc/refman/5.7/en/subquery-optimization.html:

A limitation on UPDATE and DELETE statements that use a subquery to modify a single table is that the optimizer does not use semi-join or materialization subquery optimizations. As a workaround, try rewriting them as multiple-table UPDATE and DELETE statements that use a join rather than a subquery.

  • Thanks! I would like to avoid temporary tables because they potentially cause problems with our replication. I remember using MySQL variables for updates, maybe there is a way to improve the query by storing the current rank in a variable and looping through the data? Alternatively, I could create a permanent sort table instead of a temprorary one to work around the replication problems. – rewb0rn Nov 28 '17 at 9:54
  • 1
    Main drawback to a variable is that it isn't an indexed structure, so if you have 100,000s of records, it may not be as fast as you would like. As far as your replication, what binlog_format are you running? If you are in ROW format, then the temp table will not be an issue, as the updates to the real table will be propagated as row change vectors. If you are in MIXED mode, you could SET SESSION binlog_format=ROW; just for this session, and the changes would propagate by row-based replication. Otherwise a non-temporary (permanent) intermediate table should work. – Joshua Huber Nov 28 '17 at 18:16
  • Yes we are using MIXED, so setting it for the current session only should work. Thanks for the suggestions! – rewb0rn Nov 29 '17 at 10:47

As an alternative to the accepted answer I found a solution with the use of @variables that performs much faster in comparison and does not need any temporary tables. The improvement comes from the fact that a single ORDER BY is much faster than a COUNT(*) ... WHERE ... statement per existing row. The variables are then updated in the order of the sort and a simple CASE can check if the score equals the previous player or not.

Note that this assumes that the player with the highest score has a score != 0.

UPDATE archive_league LEFT JOIN (

SELECT x.id, 
  FROM (SELECT t.id,
               @rank := @rank + 1,
               @lastRank := CASE WHEN t.score = @lastScore
                      THEN @lastRank
                      ELSE @rank
                END AS new_rank,
               @lastScore := t.score
          FROM archive_league t
          JOIN (SELECT @rank := 0, @lastRank := 0, @lastScore := 0) r
    WHERE date = :date
      ORDER BY t.score DESC) x) AS temp USING (id) SET rank = new_rank     
WHERE date = :date;

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.