You are here

Possible optimization for sort_merge and UNION ORDER BY LIMIT

MySQL Performance Blog - Tue, 18/09/2007 - 3:19pm

Every so often you need to perform sort results retrieved from MySQL when your WHERE clause goes beyound col=const values which would allow MySQL to still use second portion of the index for the order by. Ranges as well as IN lists make this optimization impossible, not even speaking about index merge optimization. Lets look at this example:

PLAIN TEXT SQL:
  1. CREATE TABLE `utest` (
  2.   `c1` int(10) UNSIGNED NOT NULL,
  3.   `c2` int(10) UNSIGNED NOT NULL,
  4.   `ord` int(10) UNSIGNED NOT NULL,
  5.   KEY `c1` (`c1`,`ord`),
  6.   KEY `c2` (`c2`,`ord`)
  7. ) ENGINE=MyISAM DEFAULT CHARSET=latin1
  8.  
  9. mysql> EXPLAIN SELECT * FROM utest WHERE c1=5 OR c2=5 ORDER BY ord DESC LIMIT 10 \G
  10. *************************** 1. row ***************************
  11.            id: 1
  12.   select_type: SIMPLE
  13.         TABLE: utest
  14.          type: index_merge
  15. possible_keys: c1,c2
  16.           KEY: c1,c2
  17.       key_len: 4,4
  18.           ref: NULL
  19.          rows: 162380
  20.         Extra: USING sort_union(c1,c2); USING WHERE; USING filesort
  21. 1 row IN SET (0.00 sec)

As you can see MySQL 5.1.21 uses sort_merge to access the rows but it can't use it together with order by efficiently.
I'd say this is limitation as for this case you well could be retrieving sorted streams by the indexes and doing merge sort to get resulted rows in sorted order. Once this is implemented similar approach could be used for optimizing ORDER BY with IN.

This original query (in memory data) takes 1sec. Let's see how classic pre MySQL 5.0 solution - using UNION instead of single query works in this case:

PLAIN TEXT SQL:
  1. mysql> EXPLAIN (SELECT * FROM utest WHERE c1=5) union (SELECT * FROM utest WHERE c2=5) ORDER BY ord DESC LIMIT 10 \G
  2. *************************** 1. row ***************************
  3.            id: 1
  4.   select_type: PRIMARY
  5.         TABLE: utest
  6.          type: ref
  7. possible_keys: c1
  8.           KEY: c1
  9.       key_len: 4
  10.           ref: const
  11.          rows: 108112
  12.         Extra:
  13. *************************** 2. row ***************************
  14.            id: 2
  15.   select_type: UNION
  16.         TABLE: utest
  17.          type: ref
  18. possible_keys: c2
  19.           KEY: c2
  20.       key_len: 4
  21.           ref: const
  22.          rows: 54268
  23.         Extra:
  24. *************************** 3. row ***************************
  25.            id: NULL
  26.   select_type: UNION RESULT
  27.         TABLE: <union1,2>
  28.          type: ALL
  29. possible_keys: NULL
  30.           KEY: NULL
  31.       key_len: NULL
  32.           ref: NULL
  33.          rows: NULL
  34.         Extra: USING filesort
  35. 3 rows IN SET (0.00 sec)

The query looks scare but in fact in completes in 1.05 sec not significantly faster than sort index merge in this case.

As the query time implies MySQL is not smart enough in this case to "dive into" the union and add ORDER BY ORD LIMIT 10 to individual queries.
What if we do int manually ?

PLAIN TEXT SQL:
  1. mysql> EXPLAIN (SELECT * FROM utest WHERE c1=5 ORDER BY ord DESC LIMIT 10) union (SELECT * FROM utest WHERE c2=5 ORDER BY ord DESC LIMIT 10) ORDER BY ord DESC LIMIT 10 \G
  2. *************************** 1. row ***************************
  3.            id: 1
  4.   select_type: PRIMARY
  5.         TABLE: utest
  6.          type: ref
  7. possible_keys: c1
  8.           KEY: c1
  9.       key_len: 4
  10.           ref: const
  11.          rows: 108112
  12.         Extra: USING WHERE
  13. *************************** 2. row ***************************
  14.            id: 2
  15.   select_type: UNION
  16.         TABLE: utest
  17.          type: ref
  18. possible_keys: c2
  19.           KEY: c2
  20.       key_len: 4
  21.           ref: const
  22.          rows: 54268
  23.         Extra: USING WHERE
  24. *************************** 3. row ***************************
  25.            id: NULL
  26.   select_type: UNION RESULT
  27.         TABLE: <union1,2>
  28.          type: ALL
  29. possible_keys: NULL
  30.           KEY: NULL
  31.       key_len: NULL
  32.           ref: NULL
  33.          rows: NULL
  34.         Extra: USING filesort
  35. 3 rows IN SET (0.00 sec)

As you can see explain does not really change (it does not show if index is used for finding rows or sorting as well) but query speed goes to less than
0.01 sec.

So it is great optimization which you need to do manuall for the time being.

What is also interesting is the fact MySQL is unable to handle even basic UNION with limit (without order by) optimally - in creates result set for the union fully and when only takes 10 rows from it:

PLAIN TEXT SQL:
  1. mysql> EXPLAIN (SELECT * FROM utest WHERE c1=5 ) union (SELECT * FROM utest WHERE c2=5 ) LIMIT 10 \G
  2. *************************** 1. row ***************************
  3.            id: 1
  4.   select_type: PRIMARY
  5.         TABLE: utest
  6.          type: ref
  7. possible_keys: c1
  8.           KEY: c1
  9.       key_len: 4
  10.           ref: const
  11.          rows: 108112
  12.         Extra:
  13. *************************** 2. row ***************************
  14.            id: 2
  15.   select_type: UNION
  16.         TABLE: utest
  17.          type: ref
  18. possible_keys: c2
  19.           KEY: c2
  20.       key_len: 4
  21.           ref: const
  22.          rows: 54268
  23.         Extra:
  24. *************************** 3. row ***************************
  25.            id: NULL
  26.   select_type: UNION RESULT
  27.         TABLE: <union1,2>
  28.          type: ALL
  29. possible_keys: NULL
  30.           KEY: NULL
  31.       key_len: NULL
  32.           ref: NULL
  33.          rows: NULL
  34.         Extra:
  35. 3 rows IN SET (0.00 sec)

Such query takes about 0.9 sec on the test data. Another possible optimizer improvement to do.

P.S This post is inspired by Does MySQL Optimize UNION with LIMIT clause topic on our MySQL Forums.