Odds are if you develop database applications you have been asked to make a large table of textual data searchable. How can we do this most reliably and efficiently? Bill compares a range of techniques to search text data with MySQL, with a focus on performance. - Ad hoc searches with the LIKE predicate or regular expressions. - MySQL FULLTEXT index. - InnoDB FULLTEXT index in MySQL 5.6. - Apache Solr search engine - Sphinx search engine. - Trigraphs.

1. Full Text Search
Throwdown
Bill Karwin, Percona Inc.

2. In a full text
search, the search
engine examines
all of the words in
every stored
document as it
tries to match
search words
supplied by the
user.
http://www.flickr.com/photos/tryingyouth/

3. StackOverflow Test Data
• Data dump, exported December 2011
• 7.4 million Posts = 8.18 GB
www.percona.com

4. StackOverflow ER diagram
searchable
text
www.percona.com

5. The Baseline:
Naive Search Predicates

6. Some people, when confronted with a problem, think
“I know, I’ll use regular expressions.”
Now they have two problems.
— Jamie Zawinsky
www.percona.com

7. Accuracy issue
• Irrelevant or false matching words ‘one’, ‘money’,
‘prone’, etc.:
SELECT * FROM Posts
WHERE Body LIKE '%one%'
• Regular expressions in MySQL support escapes
for word boundaries:
SELECT * FROM Posts
WHERE Body RLIKE '[[:<:]]one[[:>:]]'
www.percona.com

8. Performance issue
• LIKE with wildcards:
! SELECT * FROM Posts
49 sec
WHERE title LIKE '%performance%'
! OR body LIKE '%performance%'
! OR tags LIKE '%performance%';
• POSIX regular expressions:
! SELECT * FROM Posts 7 min 57 sec
WHERE title RLIKE '[[:<:]]performance[[:>:]]'
! OR body RLIKE '[[:<:]]performance[[:>:]]'
! OR tags RLIKE '[[:<:]]performance[[:>:]]';
www.percona.com

9. Why so slow?
CREATE TABLE TelephoneBook (
! FullName VARCHAR(50));
CREATE INDEX name_idx ON TelephoneBook
! (FullName);
INSERT INTO TelephoneBook VALUES
! ('Riddle, Thomas'),
! ('Thomas, Dean');
www.percona.com

10. Why so slow?
• Search for all with last name “Thomas”
uses index
SELECT * FROM telephone_book
WHERE full_name LIKE 'Thomas%'
• Search for all with first name “Thomas”
SELECT * FROM telephone_book
WHERE full_name LIKE '%Thomas'
can’t use index
www.percona.com

11. Because:
B-Tree indexes can’t
☞ search for substrings
www.percona.com

12. • FULLTEXT in MyISAM
• FULLTEXT in InnoDB
• Apache Solr
• Sphinx Search
• Trigraphs

13. FULLTEXT
in MyISAM

14. FULLTEXT Index with MyISAM
• Special index type for MyISAM
• Integrated with SQL queries
• Indexes always in sync with data
• Balances features vs. speed vs. space
www.percona.com

15. Insert Data into Index (MyISAM)
mysql> INSERT INTO Posts
SELECT * FROM PostsSource;
time: 33 min, 34 sec
www.percona.com

16. Build Index on Data (MyISAM)
mysql> CREATE FULLTEXT INDEX PostText
! ON Posts(title, body, tags);
time: 31 min, 18 sec
www.percona.com

17. Querying
SELECT * FROM Posts
WHERE MATCH(column(s))
AGAINST('query pattern');
must include all columns
of your index, in the
order you defined
www.percona.com

18. Natural Language Mode (MyISAM)
• Searches concepts with free text queries:
! SELECT * FROM Posts
WHERE MATCH(title, body, tags )
AGAINST('mysql performance'
IN NATURAL LANGUAGE MODE)
LIMIT 100;
time with index:
200 milliseconds
www.percona.com

19. Query Profile:
Natural Language Mode (MyISAM)
+-------------------------+----------+
| Status | Duration |
+-------------------------+----------+
| starting | 0.000068 |
| checking permissions | 0.000006 |
| Opening tables | 0.000017 |
| init | 0.000032 |
| System lock | 0.000007 |
| optimizing | 0.000007 |
| statistics | 0.000018 |
| preparing | 0.000006 |
| FULLTEXT initialization | 0.198358 |
| executing | 0.000012 |
| Sending data | 0.001921 |
| end | 0.000005 |
| query end | 0.000003 |
| closing tables | 0.000018 |
| freeing items | 0.000341 |
| cleaning up | 0.000012 |
+-------------------------+----------+
www.percona.com

20. Boolean Mode (MyISAM)
• Searches words using mini-language:
! SELECT * FROM Posts
WHERE MATCH(title, body, tags)
AGAINST('+mysql +performance'
IN BOOLEAN MODE)
LIMIT 100;
time with index:
16 milliseconds
www.percona.com

21. Query Profile:
Boolean Mode (MyISAM)
+-------------------------+----------+
| Status | Duration |
+-------------------------+----------+
| starting | 0.000031 |
| checking permissions | 0.000003 |
| Opening tables | 0.000008 |
| init | 0.000017 |
| System lock | 0.000004 |
| optimizing | 0.000004 |
| statistics | 0.000008 |
| preparing | 0.000003 |
| FULLTEXT initialization | 0.000008 |
| executing | 0.000001 |
| Sending data | 0.015703 |
| end | 0.000004 |
| query end | 0.000002 |
| closing tables | 0.000007 |
| freeing items | 0.000381 |
| cleaning up | 0.000007 |
+-------------------------+----------+
www.percona.com

22. FULLTEXT
in InnoDB

23. FULLTEXT Index with InnoDB
• Under development in MySQL 5.6
• I’m testing 5.6.6 m1
• Usage very similar to FULLTEXT in MyISAM
• Integrated with SQL queries
• Indexes always* in sync with data
• Read the blogs for more details:
• http://blogs.innodb.com/wp/2011/07/overview-and-getting-started-with-innodb-fts/
• http://blogs.innodb.com/wp/2011/07/innodb-full-text-search-tutorial/
• http://blogs.innodb.com/wp/2011/07/innodb-fts-performance/
• http://blogs.innodb.com/wp/2011/07/difference-between-innodb-fts-and-myisam-fts/
www.percona.com

24. Insert Data into Index (InnoDB)
mysql> INSERT INTO Posts
SELECT * FROM PostsSource;
time: 55 min 46 sec
www.percona.com

25. Build Index on Data (InnoDB)
• Still under development; you might see problems:
mysql> CREATE FULLTEXT INDEX PostText
! ON Posts(title, body, tags);
ERROR 2013 (HY000): Lost connection to
MySQL server during query
www.percona.com

26. Build Index on Data (InnoDB)
• Solution: make sure you define a primary key
column `FTS_DOC_ID` explicitly:
mysql> ALTER TABLE Posts
CHANGE COLUMN PostId
`FTS_DOC_ID` BIGINT UNSIGNED;
mysql> CREATE FULLTEXT INDEX PostText
! ON Posts(title, body, tags);
time: 25 min 27 sec
www.percona.com

27. Natural Language Mode (InnoDB)
• Searches concepts with free text queries:
! SELECT * FROM Posts
WHERE MATCH(title, body, tags)
AGAINST('mysql performance'
IN NATURAL LANGUAGE MODE)
LIMIT 100;
time with index:
740 milliseconds
www.percona.com

28. Query Profile:
Natural Language Mode (InnoDB)
+-------------------------+----------+
| Status | Duration |
+-------------------------+----------+
| starting | 0.000074 |
| checking permissions | 0.000007 |
| Opening tables | 0.000020 |
| init | 0.000034 |
| System lock | 0.000007 |
| optimizing | 0.000009 |
| statistics | 0.000020 |
| preparing | 0.000008 |
| FULLTEXT initialization | 0.577257 |
| executing | 0.000013 |
| Sending data | 0.106279 |
| end | 0.000018 |
| query end | 0.000012 |
| closing tables | 0.000018 |
| freeing items | 0.055584 |
| cleaning up | 0.000039 |
+-------------------------+----------+
www.percona.com

29. Boolean Mode (InnoDB)
• Searches words using mini-language:
! SELECT * FROM Posts
WHERE MATCH(title, body, tags)
AGAINST('+mysql +performance'
IN BOOLEAN MODE)
LIMIT 100;
time with index:
350 milliseconds
www.percona.com

30. Query Profile:
Boolean Mode (InnoDB)
+-------------------------+----------+
| Status | Duration |
+-------------------------+----------+
| starting | 0.000064 |
| checking permissions | 0.000005 |
| Opening tables | 0.000017 |
| init | 0.000047 |
| System lock | 0.000007 |
| optimizing | 0.000009 |
| statistics | 0.000019 |
| preparing | 0.000008 |
| FULLTEXT initialization | 0.347172 |
| executing | 0.000014 |
| Sending data | 0.008089 |
| end | 0.000011 |
| query end | 0.000012 |
| closing tables | 0.000015 |
| freeing items | 0.001570 |
| cleaning up | 0.000023 |
+-------------------------+----------+
www.percona.com

31. Apache Solr

32. Apache Solr
• http://lucene.apache.org/solr/
• Formerly known as Lucene, started 2001
• Apache License
• Java implementation
• Web service architecture
• Many sophisticated search features
www.percona.com

33. DataImportHandler
• conf/solrconfig.xml:
. . .
<requestHandler name="/dataimport"
class="org.apache.solr.handler.dataimport.DataImportHandler">
<lst name="defaults">
<str name="config">data-config.xml</str>
</lst>
</requestHandler>
. . .
www.percona.com

34. DataImportHandler
• conf/data-config.xml:
<dataConfig>
<dataSource type="JdbcDataSource"
driver="com.mysql.jdbc.Driver"
url="jdbc:mysql://localhost/testpattern?useUnicode=true"
batchSize="-1"
user="xxxx"
password="xxxx"/>
<document>
<entity name="id"
query="SELECT PostId, ParentId, Title, Body, Tags FROM Posts">
</entity>
</document>
</dataConfig> extremely important
to avoid buffering the
whole query result!
www.percona.com

35. DataImportHandler
•conf/schema.xml:
. . .
<fields>
<field name="PostId" type="string" indexed="true" stored="true" required="true" />
<field name="ParentId" type="string" indexed="true" stored="true" required="false" />
<field name="Title" type="text_general" indexed="false" stored="false"
required="false" />
<field name="Body" type="text_general" indexed="false" stored="false" required="false" /
>
<field name="Tags" type="text_general" indexed="false" stored="false" required="false" /
>
<field name="text" type="text_general" indexed="true" stored="false" multiValued="true"/
>
<fields>
<uniqueKey>PostId</uniqueKey>
<defaultSearchField>text</defaultSearchField>
<copyField source="Title" dest="text"/>
<copyField source="Body" dest="text"/>
<copyField source="Tags" dest="text"/>
. . .
www.percona.com

36. Insert Data into Index (Solr)
• http://localhost:8983/solr/dataimport?
command=full-import
time: 14 min 28 sec
www.percona.com

37. Searching Solr
• http://localhost:8983/solr/select/?q=mysql+AND
+performance
time: 79ms
Query results are cached (like MySQL Query Cache),
so they return much faster on subsequent execution
www.percona.com

38. Sphinx Search

39. Sphinx Search
• http://sphinxsearch.com/
• Started in 2001
• GPLv2 license
• C++ implementation
• SphinxSE storage engine for MySQL
• Supports MySQL protocol, SQL-like queries
• Many sophisticated search features
www.percona.com

40. sphinx.conf
source src1
{
! type = mysql
! sql_host = localhost
! sql_user = xxxx
! sql_pass = xxxx
! sql_db = testpattern
! sql_query = SELECT PostId, ParentId, Title,
! ! Body, Tags FROM Posts
! sql_query_info = SELECT * FROM Posts
! ! WHERE PostId=$id
}
www.percona.com

41. sphinx.conf
! index test1
{
! source = src1
! path = C:Sphinxdata
}
www.percona.com

42. Insert Data into Index (Sphinx)
! C:Sphinx> indexer.exe -c sphinx.conf.in --verbose test1
Sphinx 2.0.5-release (r3309)
using config file 'sphinx.conf'...
indexing index 'test1'...
collected 7397507 docs, 5731.8 MB time: 8 min 20 sec
sorted 920.3 Mhits, 100.0% done
total 7397507 docs, 5731776959 bytes
total 500.149 sec, 11460138 bytes/sec, 14790.60 docs/sec
total 11 reads, 15.898 sec, 314584.8 kb/call avg, 1445.3
msec/call avg
total 542 writes, 3.129 sec, 12723.3 kb/call avg, 5.7 msec/
call avg
Execution time: 500.196 s
www.percona.com

43. Querying index
$ mysql --port 9306
Server version: 2.0.5-release (r3309)
mysql> SELECT * FROM test1 WHERE MATCH('mysql performance');
+---------+--------+
| id | weight |
+---------+--------+
| 6016856 | 6600 |
| 4207641 | 6595 |
| 2656325 | 6593 |
| 7192928 | 5605 |
| 8118235 | 5598 |
. . .
20 rows in set (0.02 sec)
www.percona.com

44. Querying index
mysql> SHOW META;
+---------------+-------------+
| Variable_name | Value |
+---------------+-------------+
| total | 1000 |
| total_found | 7672 |
| time | 0.013 |
| keyword[0] | mysql |
| docs[0] | 162287 | time: 13ms
| hits[0] | 363694 |
| keyword[1] | performance |
| docs[1] | 147249 |
| hits[1] | 210895 |
+---------------+-------------+
www.percona.com

45. Trigraphs

46. Trigraphs Overview
• Not very fast, but still better than LIKE / RLIKE
• Generic, portable SQL solution
• No dependency on version, storage engine, third-
party technology
www.percona.com

47. Three-Letter Sequences
! CREATE TABLE AtoZ (
! c! ! CHAR(1),
! PRIMARY KEY (c));
! INSERT INTO AtoZ (c)
VALUES ('a'), ('b'), ('c'), ...
! CREATE TABLE Trigraphs (
! Tri!
! CHAR(3),
! PRIMARY KEY (Tri));
! INSERT INTO Trigraphs (Tri)
SELECT CONCAT(t1.c, t2.c, t3.c)
FROM AtoZ t1 JOIN AtoZ t2 JOIN AtoZ t3;
www.percona.com

48. Insert Data Into Index
my $sth = $dbh1->prepare("SELECT * FROM Posts") or die $dbh1->errstr;
$sth->execute() or die $dbh1->errstr;
$dbh2->begin_work;
my $i = 0;
while (my $row = $sth->fetchrow_hashref ) {
my $text = lc(join('|', ($row->{title}, $row->{body}, $row->{tags})));
my %tri;
map($tri{$_}=1, ( $text =~ m/[[:alpha:]]{3}/g ));
next unless %tri;
my $tuple_list = join(",", map("('$_',$row->{postid})", keys %tri));
my $sql = "INSERT IGNORE INTO PostsTrigraph (tri, PostId) VALUES
$tuple_list";
$dbh2->do($sql) or die "SQL = $sql, ".$dbh2->errstr;
if (++$i % 1000 == 0) {
print ".";
$dbh2->commit;
$dbh2->begin_work;
time: 116 min 50 sec
}
}
space: 16.2GiB
print ".n";
$dbh2->commit; rows: 519 million
www.percona.com

49. Indexed Lookups
! SELECT p.* time: 46 sec
FROM Posts p
JOIN PostsTrigraph t1 ON
! t1.PostId = p.PostId AND t1.Tri = 'mys'
!
www.percona.com

50. Search Among Fewer Matches
! SELECT p.* time: 19 sec
FROM Posts p
JOIN PostsTrigraph t1 ON
! t1.PostId = p.PostId AND t1.Tri = 'mys'
JOIN PostsTrigraph t2 ON
! t2.PostId = p.PostId AND t2.Tri = 'per'
www.percona.com

51. Search Among Fewer Matches
! SELECT p.* time: 22 sec
FROM Posts p
JOIN PostsTrigraph t1 ON
! t1.PostId = p.PostId AND t1.Tri = 'mys'
JOIN PostsTrigraph t2 ON
! t2.PostId = p.PostId AND t2.Tri = 'per'
JOIN PostsTrigraph t3 ON
! t3.PostId = p.PostId AND t3.Tri = 'for'
www.percona.com

52. Search Among Fewer Matches
! SELECT p.* time: 13.6 sec
FROM Posts p
JOIN PostsTrigraph t1 ON
! t1.PostId = p.PostId AND t1.Tri = 'mys'
JOIN PostsTrigraph t2 ON
! t2.PostId = p.PostId AND t2.Tri = 'per'
JOIN PostsTrigraph t3 ON
! t3.PostId = p.PostId AND t3.Tri = 'for'
JOIN PostsTrigraph t4 ON
! t4.PostId = p.PostId AND t4.Tri = 'man'
www.percona.com

53. Narrow Down Further
! SELECT p.* time: 13.8 sec
FROM Posts p
JOIN PostsTrigraph t1 ON
! t1.PostId = p.PostId AND t1.Tri = 'mys'
JOIN PostsTrigraph t2 ON
! t2.PostId = p.PostId AND t2.Tri = 'per'
JOIN PostsTrigraph t3 ON
! t3.PostId = p.PostId AND t3.Tri = 'for'
JOIN PostsTrigraph t4 ON
! t4.PostId = p.PostId AND t4.Tri = 'man'
WHERE CONCAT(p.title,p.body,p.tags) LIKE '%mysql%'
! AND CONCAT(p.title,p.body,p.tags) LIKE '%performance%';
www.percona.com

54. And the winner is...
Jarrett Campbell
http://www.flickr.com/people/77744839@N00

55. Time to Insert Data into Index
LIKE expression n/a
FULLTEXT MyISAM 33 min, 34 sec
FULLTEXT InnoDB 55 min, 46 sec
Apache Solr 14 min, 28 sec
Sphinx Search 8 min, 20 sec
Trigraphs 116 min, 50 sec
www.percona.com

56. Insert Data into Index (sec)
8000
6000
4000
2000
0
LIKE MyISAM InnoDB Solr Sphinx Trigraph
www.percona.com

57. Time to Build Index on Data
LIKE expression n/a
FULLTEXT MyISAM 31 min, 18 sec
FULLTEXT InnoDB 25 min, 27 sec
Apache Solr n/a
Sphinx Search n/a
Trigraphs n/a
www.percona.com

58. Build Index on Data (sec)
4000
3000
2000
1000
n/a n/a n/a
0
LIKE MyISAM InnoDB Solr Sphinx Trigraph
www.percona.com

59. Index Storage
LIKE expression n/a
FULLTEXT MyISAM 2382 MiB
FULLTEXT InnoDB ? MiB
Apache Solr 2766 MiB
Sphinx Search 3355 MiB
Trigraphs 16589 MiB
www.percona.com

60. Index Storage (MiB)
20000
15000
10000
5000
0
LIKE MyISAM InnoDB Solr Sphinx Trigraph
www.percona.com

61. Query Speed
LIKE expression 49,000ms - 399,000ms
FULLTEXT MyISAM 16-200ms
FULLTEXT InnoDB 350-740ms
Apache Solr 79ms
Sphinx Search 13ms
Trigraphs 13800ms
www.percona.com

62. Query Speed (ms)
400000
350000
300000
250000
200000
150000
100000
50000
0
LIKE MyISAM InnoDB Solr Sphinx Trigraph
www.percona.com

63. 1000
Query Speed (ms)
750
500
250
0
LIKE MyISAM InnoDB Solr Sphinx Trigraph
www.percona.com

64. Bottom Line
build insert storage query solution
49k-399k
LIKE expression 0 0 0
ms
SQL
FULLTEXT MyISAM 31:18 33:28 2382MiB 16-200ms MySQL
FULLTEXT InnoDB 25:27 55:46 ? 350-740ms MySQL 5.6
Apache Solr n/a 14:28 2766MiB 79ms Java
Sphinx Search n/a 8:20 3487MiB 13ms C++
Trigraphs n/a 116:50 16.2 GiB 13,800ms SQL
www.percona.com

65. Final Thoughts
• Third-party search engines are complex to keep in
sync with data, and adding another type of
server adds more operations work for you.
• Built-in FULLTEXT indexes are therefore useful
even if they are not absolutely the fastest.
• Different search implementations may return
different results, so you should evaluate what
works best for your project.
• Any indexed search solution is orders of
magnitude better than LIKE!
www.percona.com

66. New York, October 1-2, 2012
London, December 3-4, 2012
Santa Clara, April 22-25, 2013
www.percona.com/live

67. Expert instructors
In-person training
Custom onsite training
Live virtual training
http://www.percona.com/training
www.percona.com

68. http://www.pragprog.com/titles/bksqla/
www.percona.com

69. Copyright 2012 Bill Karwin
www.slideshare.net/billkarwin
Released under a Creative Commons 3.0 License:
http://creativecommons.org/licenses/by-nc-nd/3.0/
You are free to share - to copy, distribute and
transmit this work, under the following conditions:
Attribution. Noncommercial. No Derivative Works.
You must attribute this You may not use this work You may not alter,
work to Bill Karwin. for commercial purposes. transform, or build
upon this work.
www.percona.com