Optimizer Trace and EXPLAIN FORMAT=JSON in 5.7

I accidentally stumbled upon this Stack Overflow question this morning:

I am wondering if there is any difference in regards to performance between the following:

SELECT ... FROM ... WHERE someFIELD IN(1,2,3,4);
SELECT ... FROM ... WHERE someFIELD between  0 AND 5;
SELECT ... FROM ... WHERE someFIELD = 1 OR someFIELD = 2 OR someFIELD = 3 ...;

It is an interesting question because there was no good way to answer it when it was asked in 2009. All of the queries resolve to the same output in EXPLAIN. Here is an example using the sakila schema:

mysql> EXPLAIN SELECT * FROM film WHERE film_id BETWEEN 1 AND 5\G
mysql> EXPLAIN SELECT * FROM film WHERE film_id IN (1,2,3,4,5)\G
mysql> EXPLAIN SELECT * FROM film WHERE film_id =1 or film_id=2 or film_id=3 or film_id=4 or film_id=5\G
********* 1. row *********
           id: 1
  select_type: SIMPLE
        table: film
   partitions: NULL
         type: range
possible_keys: PRIMARY
          key: PRIMARY
      key_len: 2
          ref: NULL
         rows: 5
     filtered: 100.00
        Extra: Using where

Times have changed though. There are now a couple of useful features to show the difference :)

Optimizer Trace

Optimizer trace is a new diagnostic tool introduced in MySQL 5.6 to show how the optimizer is working internally. It is similar to EXPLAIN, with a few notable differences:

  • It doesn't just show the intended execution plan, it shows the alternative choices.
  • You enable the optimizer trace, then you run the actual query.
  • It is far more verbose in its output.

Here are the outputs for the three versions of the query:

  1. SELECT * FROM film WHERE film_id BETWEEN 1 AND 5
  2. SELECT * FROM film WHERE film_id IN (1,2,3,4,5)
  3. SELECT * FROM film WHERE film_id =1 or film_id=2 or film_id=3 or film_id=4 or film_id=5

What is the difference?

The optimizer trace output shows that the first query executes as one range, while the second and third execute as 5 separate single-value ranges:

                  "chosen_range_access_summary": {
                    "range_access_plan": {
                      "type": "range_scan",
                      "index": "PRIMARY",
                      "rows": 5,
                      "ranges": [
                        "1 <= film_id <= 1",
                        "2 <= film_id <= 2",
                        "3 <= film_id <= 3",
                        "4 <= film_id <= 4",
                        "5 <= film_id <= 5"
                      ]
                    },
                    "rows_for_plan": 5,
                    "cost_for_plan": 6.0168,
                    "chosen": true
                  }

This can also be confirmed with the handler counts from SHOW STATUS:

BETWEEN 1 AND 5: 
 Handler_read_key: 1
 Handler_read_next: 5
IN (1,2,3,4,5):
 Handler_read_key: 5
film_id =1 or film_id=2 or film_id=3 or film_id=4 or film_id=5:
 Handler_read_key: 5

So I would say that BETWEEN 1 AND 5 is the cheapest query, because it finds one key and then says next, next, next until finished. The optimizer seems to agree with me. A single range access plus next five times costs 2.0168 instead of 6.0168:

                  "chosen_range_access_summary": {
                    "range_access_plan": {
                      "type": "range_scan",
                      "index": "PRIMARY",
                      "rows": 5,
                      "ranges": [
                        "1 <= film_id <= 5"
                      ]
                    },
                    "rows_for_plan": 5,
                    "cost_for_plan": 2.0168,
                    "chosen": true
                  }
                }

For context, a cost unit is a logical representation of approximately one random IO. It is stable to compare costs between different execution plans.

Ranges are not all equal

Perhaps a better example to demonstrate this, is the difference between these two ranges:

  • SELECT * FROM film WHERE film_id BETWEEN 1 and 20
  • SELECT * FROM film WHERE (film_id BETWEEN 1 and 10) or (film_id BETWEEN 911 and 920)

It's pretty obvious that the second one needs to execute in two separate ranges. EXPLAIN will not show this difference, and both queries appear the same:

********* 1. row *********
           id: 1
  select_type: SIMPLE
        table: film
   partitions: NULL
         type: range
possible_keys: PRIMARY
          key: PRIMARY
      key_len: 2
          ref: NULL
         rows: 20
     filtered: 100.00
        Extra: Using where

Two distinct separate ranges may be two separate pages, and thus have different cache efficiency on the buffer pool. It should be possible to distinguish between the two.

EXPLAIN FORMAT=JSON

EXPLAIN FORMAT=JSON was introduced in MySQL 5.6 along with OPTIMIZER TRACE, but where it really becomes useful is MySQL 5.7. The JSON output will now include cost information (as well as showing separate ranges as attached_condition):

********* 1. row *********
EXPLAIN: {
  "query_block": {
    "select_id": 1,
    "cost_info": {
      "query_cost": "10.04"
    },
    "table": {
      "table_name": "film",
      "access_type": "range",
      "possible_keys": [
        "PRIMARY"
      ],
      "key": "PRIMARY",
      "used_key_parts": [
        "film_id"
      ],
      "key_length": "2",
      "rows_examined_per_scan": 20,
      "rows_produced_per_join": 20,
      "filtered": "100.00",
      "cost_info": {
        "read_cost": "6.04",
        "eval_cost": "4.00",
        "prefix_cost": "10.04",
        "data_read_per_join": "15K"
      },
      "used_columns": [
        "film_id",
        "title",
        "description",
        "release_year",
        "language_id",
        "original_language_id",
        "rental_duration",
        "rental_rate",
        "length",
        "replacement_cost",
        "rating",
        "special_features",
        "last_update"
      ],
      "attached_condition": "((`film`.`film_id` between 1 and 10) or (`film`.`film_id` between 911 and 920))"
    }
  }
}

With the FORMAT=JSON output also showing cost, we can see that two ranges costs 10.04, versus one big range costing 9.04 (not shown). These queries are not identical in cost even though they are in EXPLAIN output.

Conclusion

I have heard many users say "joins are slow", but a broad statement like this misses magnitude. By including the cost information in EXPLAIN we get all users to speak the same language. We can now say "this join is expensive", which is a much better distinction :)

It is time to start using OPTIMIZER TRACE, and particularly in 5.7 ditch EXPLAIN for EXPLAIN FORMAT=JSON.

MySQL is 20 years old tomorrow!

According to Wikipedia, the initial release of MySQL was 23 May 1995. This places it at 20 years old young tomorrow.

2015 is actually a year of anniversaries. As well as MySQL reaching the big two-oh, we are also celebrating 10 years of InnoDB under the stewardship of Oracle, and five years of MySQL @ Oracle.

Please make sure toast to MySQL this weekend!

Percona Live

(Photo from Percona Live community dinner 2015.)

A quick update on our native Data Dictionary

In July 2014, I wrote that we were working on a new native InnoDB data dictionary to replace MySQL's legacy frm files.

This is quite possibly the largest internals change to MySQL in modern history, and will unlock a number of previous limitations, as well as simplify a number of failure states for both replication and crash recovery.

With MySQL 5.7 approaching release candidate (and large changes always coming with risk attached) we decided that the timing to try to merge in a new data dictionary was just too tight. The data dictionary development is still alive and well, but it will not ship as part of MySQL 5.7.

So please stay tuned for updates... and thank you for using MySQL!

MySQL 5.6.24 Community Release Notes

Thank you to the MySQL Community, on behalf of the MySQL team @ Oracle. Your bug reports, testcases and patches have helped create a better MySQL 5.6.24.

In particular:

  • Thank you to Simon Mudd for reporting that ALTER TABLE did not take advantage of fast alterations if the table contained temporal columns found to be in pre-5.6.4 format. Bug #72997.
  • Thank you to Doug Warner for reporting that a TRUNCATE TABLE operation on a temporary table raised an assertion. Bug #72080.
  • Thank you to Elena Stepanova for reporting that an InnoDB full-text phrase search returned incorrect results. Bug #75755.
  • Thank you to Miljenko Brkic for reporting that the exptime set with the memcached API set command was ignored. Bug #70055.
  • Thank you to Nilnandan Joshi for reporting a regression where the mysqld server could not shutdown while the memcached plugin was active. Bug #74956.
  • Thank you to Zhai Weixiang for reporting that several mutexes on InnoDB dummy tables caused contention, and could be relaxed safely. Bug #73361.
  • Thank you to Zhai Weixiang for reporting that append operations using the InnoDB memcached API could cause a segfault. Bug #75200.
  • Thank you to Ramesh Sivaraman and Roel Van de Paar for reporting that a number of ALTER TABLE statements that attempt to add partitions, columns, or indexes to a partitioned table while a write lock was in effect for this table were not handled correctly. Bug #74451, Bug #74478, Bug #74491, Bug #74560, Bug #74746, Bug #74841, Bug #74860, Bug #74869.
  • Thank you to Santosh Praneeth Banda for reporting that when gtid_mode=ON and slave_net_timeout was set to a low value, the slave I/O thread could appear to hang. Bug #74607.
  • Thank you to Tsubasa Tanaka and Daniël van Eeden for independently reporting that a server warning error referred to an obsolete table_cache system variable. Thank you also to Daniël for providing a patch. Bug #73373, Bug #75081.
  • Thank you to Simon Mudd for reporting that Performance Schema statement events tables incorrectly referenced system variables (@ @ instead of @@). Bug #71634.
  • Thank you to Yao Deng for reporting that mysql_real_connect() could close a file descriptor twice if the server was not running. Bug #69423.
  • Thank you to Olle Nilsson for reporting that notification of events for the general log were received by the audit log plugin only of the general query log was enabled. Bug #60782.
  • Thank you to Daniël van Eeden for reporting a security issue.

Thank you again to all the community contributors listed above. If I missed a name here, please let me know!

- Morgan

MySQL 5.6.23 Community Release Notes

Thank you to the MySQL Community, on behalf of the MySQL team @ Oracle. Your bug reports, testcases and patches have helped create a better MySQL 5.6.23.

In particular:

  • Thank you to Inaam Rana for reporting an issue with the InnoDB purge thread state and tablespace export operations. Bug #75298.
  • Thank you to Inaam Rana for reporting that the INNODB_METRICS adaptive_hash_searches_btree counter failed to report counter data. Bug #74511.
  • Thank you to Shahriyar Rzayev and David Bennett for suggesting improvements to error messages from InnoDB on non-Windows platforms David also provided a suggested patch. Bug #73365.
  • Thank you to Sadao Hiratsuka and Piotr Jurkiewicz for independently reporting that integer columns failed to increment and decrement correctly when using the memcached plugin. Piotr also provided a suggested patch to fix the issue. Bug #69415, Bug #74874.
  • Thank you to Roel Van de Paar for reporting a security issue. Bug #74292.
  • Thank you to Roel Van de Paar for reporting a security issue. Bug #74597.
  • Thank you to Santosh Praneeth Banda for reporting a situation where a slave could have more GTIDs than a master, and data inconsistency could occur. Bug #72635.
  • Thank you to Domas Mituzas for reporting that when using SHOW SLAVE STATUS to monitor replication performance, Seconds_Behind_Master sometimes displayed unexpected lag behind the master. Bug #72376.
  • Thank you to Matt Swanson for reporting that InnoDB full-text boolean searches incorrectly handled + combined with parentheses. Bug #74845.
  • Thank you to Roel Van de Paar for reporting a security issue. Bug #74447.
  • Thank you to Sean Pringle for reporting that an internal temporary table could cause problems if the file was orphaned and the name was used later for other queries. Bug #32917.

Thank you again to all the community contributors listed above. If I missed a name here, please let me know!

- Morgan