Speed Up Your Database 300 Times

Transcript

1. @afilina
   Speed Up Your Database
   300 Times
   OSCON - May 11, 2017
   

   View Slide

2. Objectives
   • What is slow?
   • Investigate performance issues.
   • Indexes.
   • Better queries.
   • Checklist.
   

   View Slide

3. Anna Filina
   • Project rescue expert
   • Dev, trainer, speaker
   • 1 company
   • 2 conferences
   

   View Slide

4. What is Slow?
   

   View Slide

5. Multitude of Queries
   • Loops:

   

   

   • Reduce number of queries.
   • Check for lazy loading in your ORM.

   foreach ($pictures as $id) {

   query_pic_details($id);

   }
   

   View Slide

6. Lazy Loading
   • Data loaded on demand.

   • Use JOIN to get all related data in one go.

   echo $picture->album->user->name;
   

   View Slide

7. Multitude of Queries
   • Deeply nested code.
   • Try get_first_thumb instead.

   

   View Slide

8. Count Queries
   • Manual: add custom code before queries.
   • Auto: ORMs have listeners or data
   collectors.

   

   View Slide

9. Slow Query log
   • Open my.cnf

   

   

   • Queries taking 0.2s or longer will be
   logged.
   • Open the log file and analyze the queries.
   slow_query_log=ON
   long_query_time=0.2
   slow_query_log_file=/path/to/file
   

   View Slide

10. Example Schema
    200

    users
    100

    albums each
    100

    pictures each
    2,000,000 pictures
    

    View Slide

11. Simple Query
    • Get all pictures from user #1

    

    

    

    • 38 seconds.
    SELECT picture.id, picture.title

    FROM picture

    LEFT JOIN album ON picture.album_id = album.id

    WHERE album.user_id = 1;
    

    View Slide

12. Why is it so Slow?
    

    View Slide

13. Explain
    • Prefix with EXPLAIN:

    

    

    

    • Each row = table scanned:

    

    

    EXPLAIN SELECT picture.id, picture.title

    FROM picture

    LEFT JOIN album ON picture.album_id = album.id

    WHERE album.user_id = 1;
    +----+-------------+---------+------+---------------+------+---------+------+---------+--------------------------------+
    | id | select_type | table | type | possible_keys | key | key_len | ref | rows | Extra |
    +----+-------------+---------+------+---------------+------+---------+------+---------+--------------------------------+
    | 1 | SIMPLE | album | ALL | NULL | NULL | NULL | NULL | 20000 | Using where |
    | 1 | SIMPLE | picture | ALL | NULL | NULL | NULL | NULL | 2000000 | Using where; Using join buffer |
    +----+-------------+---------+------+---------------+------+---------+------+---------+--------------------------------+
    

    View Slide

14. Explain
    • Details:

    

    

    

    • Because of the join,

    we’re scanning all albums for each
    picture.
    • Total 40 billion rows scanned.

    table : album
    key : NULL
    rows : 20,000
    table : picture
    key : NULL
    rows : 2,000,000
    

    View Slide

15. Indexes
    

    View Slide

16. How do They Work?
    Flip through all pages of an address book...
    ... or pull on the letter tab.
    

    View Slide

17. Add an Index
    • album.id

    • Same query down to 2.38 sec.
    • We just saved 36 sec on every request!

    ALTER TABLE album ADD PRIMARY KEY(id);
    

    View Slide

18. Under the Hood
    • Details:

    

    

    

    • With the index,

    we find our album right away.
    • Total 2,000,000 rows scanned.

    table : album
    key : PRIMARY
    rows : 1
    table : picture
    key : NULL
    rows : 2,000,000
    

    View Slide

19. Add a Better Index
    • picture.album_id

    

    • Now down to 0.12 sec.
    • 317 times faster than original.
    ALTER TABLE picture ADD INDEX(album_id);
    

    View Slide

20. Under the Hood
    • Details:

    

    

    

    • Total 200,000 rows scanned.

    table : album
    key : NULL
    rows : 20,000
    table : picture
    key : album_id
    rows : 100
    

    View Slide

21. Add a Better Index
    • album.user_id

    

    • Now down to 0.10 sec.
    • Gained another 17% in speed.

    ALTER TABLE album ADD INDEX(user_id);
    

    View Slide

22. Under the Hood
    • Details:

    

    

    

    • With the second index,

    we’re scanning 100 pics.
    • Total 10,000 rows scanned.
    table : album
    key : user_id
    rows : 100
    table : picture
    key : album_id
    rows : 100
    

    View Slide

23. Don't Index Everything!
    There is an index on each field
    And all selects take long to yield.
    Try finding which ones you need,
    Remove the rest and feel the speed.
    -- Anna Filina
    

    View Slide

24. Indexes Used
    • Primary and foreign keys make for great
    indexes.
    • Also whatever you use in JOIN, WHERE,
    ORDER BY.
    

    View Slide

25. Index Criteria
    • Change frequency
    • Selectivity: how many different values?
    • Good examples:
    ◦ user.id
    ◦ picture.date
    • Bad examples:
    ◦ user.is_active
    ◦ picture.views

    

    View Slide

26. Should We Optimize
    Further?
    We’re quite fast already.
    

    View Slide

27. Airport Customs
    • 5 customs officers, 1 min to process a
    traveller.
    • 5 travellers / min = fluid.
    • 6 travellers / min = queue slowly builds.
    • 1000 travellers / min = 3h wait if you
    arrive on 2nd minute.
    

    View Slide

28. Smaller Indexes
    • BIGINT as index is slower than TINYINT.
    • Use UNSIGNED to double the maximum
    value.
    

    View Slide

29. Tips for Better
    Queries
    

    View Slide

30. Functions
    • Avoid functions on an indexed column:
    • The index will be ignored here.

    SELECT id, title FROM picture

    WHERE YEAR(create_date) >= 2011;
    

    View Slide

31. Limit
    • Now down to 0.0009 sec.
    • 100 times faster than the last query.
    SELECT picture.id, picture.title

    FROM picture

    LEFT JOIN album ON picture.album.id = album.id

    WHERE album.user_id = 1

    LIMIT 25;
    

    View Slide

32. Other Constraints
    • Range:

    

    

    

    

    • Only makes sense if that column is
    indexed:

    

    

    SELECT picture.id, picture.title

    FROM picture

    LEFT JOIN album ON picture.album.id = album.id

    WHERE album.user_id = 1

    AND picture.id BETWEEN 26 AND 50; # 0.023 sec
    ALTER TABLE picture ADD INDEX(id);
    

    View Slide

33. Other Constraints
    • IN clause:

    

    

    

    

    

    

    

    

    

    SELECT picture.id, picture.title

    FROM picture

    LEFT JOIN album ON picture.album.id = album.id

    WHERE album.user_id = 1

    AND picture.id IN (15,26,29,32,45); # 0.048 sec
    

    View Slide

34. Story
    

    View Slide

35. Problem
    • New features released before sale
    • Index was not used
    • Creating too many tmp tables on disk
    • Slow queries
    • Long queue
    

    View Slide

36. How Did I Find the Problem?
    • Status:

    

    

    

    

    

    

    

    

    

    SHOW GLOBAL STATUS;
    +------------------------------------------+-------------+
    | Variable_name | Value |
    +------------------------------------------+-------------+
    | Aborted_clients | 0 |
    | Aborted_connects | 0 |
    | Binlog_cache_disk_use | 0 |
    | Binlog_cache_use | 0 |
    | Binlog_stmt_cache_disk_use | 0 |
    | Binlog_stmt_cache_use | 0 |
    | Bytes_received | 23378012 |
    | Bytes_sent | 2707328 |
    

    View Slide

37. How Did I Find the Problem?
    • Created_tmp_disk_tables
    ◦ Temporary tables created on order by, group
    by, etc.
    ◦ Stored on disk when can't fit in RAM
    ◦ Maybe too big resultset.
    ◦ Are we using index?
    • Handler_read_rnd_next
    ◦ Full or partial table scan
    

    View Slide

38. Split Tables
    • MySQL can partition your tables
    transparently.
    ◦ Creates multiple files on disk.
    ◦ Shows data as a single table.

    

    View Slide

39. Archives
    • MySQL ARCHIVE storage engine:
    ◦ Doesn’t support UPDATE or DELETE
    ◦ Very fast for SELECT
    • You can archive old transactions, news,
    etc.

    

    View Slide

40. Split Databases
    • Horizontal partitioning, sharding.
    • Users 1-1000 and their data in database
    #1
    • Users 1001-2000 and their data in
    database #2
    

    View Slide

41. Checklist
    ✓Count queries.
    ✓Log the slow ones.
    ✓Use EXPLAIN and
    check number of
    rows scanned.
    ✓Try different
    indexes, remove
    unused ones.
    ✓Use small indexes,
    preferably
    numeric.
    ✓Don’t use
    functions on
    indexed columns.
    ✓Make sure indexes
    are used.
    ✓Limit number of
    results.
    ✓Split tables and
    databases.
    

    View Slide

42. Further Reading
    • Storage engines: http://www.linux.org/
    article/view/an-introduction-to-mysql-
    storage-engines
    • Sharding: http://
    www.jurriaanpersyn.com/archives/
    2009/02/12/database-sharding-at-
    netlog-with-mysql-and-php/
    • Optimization manual: http://
    dev.mysql.com/doc/refman/5.7/en/
    optimization.html
    

    View Slide

43. Even More Fun Stuff
    • Setup read slaves
    • Denormalize where appropriate
    • SSD hard drive
    

    View Slide

44. @afilina afilina.com
    

    View Slide