3. InnoDB table format
• Use InnoDB by default
• InnoDB vs MyISAM pros:
• Row locking, allows concurrent writes
• ACID
• Non-blocking backups
• Better data recovery after a crash
• InnoDB cons:
• Lack of instant row count
3
4. What about MyISAM?
• Never use MyISAM for concurrent applications in
real time
• Possible use cases for MyISAM:
• No writes
• Batch writes only
• Mostly full scan selects
4
6. Table segmentation
• For very big tables or tables that will grow forever,
find some criteria to segment data into smaller
tables
• For instance:
• one table per month for log recording
• one table per country for users, so you can
shard them and put them on different servers
closer to the users
6
7. Table specialisation
• Don’t use the same table for heterogeneous
records that don’t share the same fields, it will
increase table size and affect performance when
using indexes
• For instance:
• In a database for classified ads for homes,
cars and jobs, use one table for each type,
because they don’t share fields like number of
rooms, engine power or salary.
7
9. Clustered index
1 1 2 A B … … …
Primary key
1 1 1 A A … … …
1 1 1A A
Row data
Primary keySecondary key 1
1 2 2A A
1 1 2A B
1 2 1 A B … … …
9
1 2 2 A A … … …
1 2 1A B
10. Clustered index
• You must always define a primary key, if there’s no
natural PK for the table, define an auto incremental PK
• Records are physically ordered in table by the PK
• Accessing a row using the PK is the fastest way,
because the row data is on the same page where the
index search leads
• Don’t include fields in the PK that could be modified
after insertion, it will delete and insert the record again at
the right position if you update the PK, affecting
performance
10
11. Secondary indexes
• All indexes other than the clustered index are known as secondary
indexes
• Each record in a secondary index contains the primary key columns
for the row, as well as the columns specified for the secondary index
• InnoDB uses this primary key value to search for the row in the
clustered index
• You could take profit from this design for paging by PK in selects
that use a secondary index
• If the primary key is long, the secondary indexes use more space,
so it is better to have a small auto increment PK and define a
unique key with the fields that would be the natural PK
11
15. Pros of indexes
1. Filter: access only the records you need, without
considering more records than necessary. Applies
to SELECT, UPDATE, REPLACE and DELETE
2. Sor/group: avoid using temporary tables
3. Cover: if all fields in a SELECT are included in the
index used (despite its order), data is retrieved
directly from this index, saving extra reads
15
16. Cons of indexes
1. Writes: the more indexes a table has, the slower
writes are going to be
2. Size: each index is going to increase total table
size
16
17. Rules for using indexes
• Only one index is used for each table in a query
• In case of using “OR” in a WHERE condition, it works as many different
queries, and each one could use its own index
• Fields use order in index is from left to right, always beginning with the first
one
• Its not mandatory to use all fields in an index, but every field used must be
consecutive
• Fields in WHERE condition go first, next GROUP BY and ORDER BY last
• Order of fields in WHERE condition doesn’t matter
• A range condition in WHERE or using GROUP BY or ORDER BY will prevent
using the next fields in the index
17
18. Rules for using indexes
fields use order
18
You can’t skip previous fields if you want to filter using the index for
rightmost fields
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT * FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND d_date = ‘2014-‐10-‐10’
AND fk_i_id_tbl_vertical = 1;
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT * FROM table
WHERE d_date = ‘2014-‐10-‐10’
AND fk_i_id_tbl_vertical = 1;
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT * FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND fk_i_id_tbl_vertical = 1;
19. Rules for using indexes
fields order in ranges, groups and sorts
19
None of this examples could filter using the index for the field
fk_i_id_tbl_vertical
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT * FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND d_date < ‘2014-‐10-‐10’
AND fk_i_id_tbl_vertical = 1;
SELECT * FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND fk_i_id_tbl_vertical = 1,
GROUP BY d_date;
SELECT * FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND fk_i_id_tbl_vertical = 1
ORDER BY d_date;
20. Rules for using indexes
covering indexes and fields order
20
This query will use the covering index for all the requested fields, also uses
the index for filtering the first field, but can’t use it for filtering the next fields
because the WHERE condition is skipping the second field in the index
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT fk_c_id_tbl_countries, d_date, fk_i_id_tbl_vertical FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND fk_i_id_tbl_vertical = 1;
This query will use the covering index for all the requested fields, but can’t
use it for filtering because the WHERE condition is skipping the first field in
the index
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT fk_c_id_tbl_countries, d_date, fk_i_id_tbl_vertical FROM table
WHERE d_date < ‘2014-‐10-‐10’
AND fk_i_id_tbl_vertical = 1;
21. Rules for using indexes
covering indexes and not indexed fields
21
You can’t use covering index optimization if any of the
requested fields is not included in the index
KEY `country_date_vertical` (`fk_c_id_tbl_countries`,`d_date`,`fk_i_id_tbl_vertical`)
SELECT fk_c_id_tbl_countries, d_date, fk_i_id_tbl_vertical, s_query FROM table
WHERE fk_c_id_tbl_countries = ‘es’
AND d_date = ‘2014-‐10-‐10’
AND fk_i_id_tbl_vertical = 1;
23. Duplicated indexes
• Avoid duplicating fields in different indexes, it
will affect write performance and increase table
size
• Think about the most frequent uses of the table, so
you can design the table itself and order the fields
in indexes smartly to avoid duplicate fields
23
24. Promote covering indexes
• Consider adding frequently requested fields at
the end of an index, even if they aren’t used in
WHERE, GROUP BY or ORDER BY
• If the major part of queries on a table are optimised
to use use covering indexes, is the most
important performance boost you can get
24
25. Index cardinality
• Cardinality is how many unique values an index has
• The more cardinality, the more efficient an index is
filtering records
• MySQL maintains approximate statistics about
cardinality
• Each MySQL version gives different cardinality
values, and they can become out of date under high
write load
25
26. Index cardinality
26
mysql> analyze table users;
mysql> show index from users;
+-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+
| Table | Non_unique | Key_name | Seq | Column_name | Cardinality |
+-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+
| users | 0 | PRIMARY | 1 | id | 9728224 |
| users | 1 | age_sex | 1 | age | 192 |
| users | 1 | age_sex | 2 | sex | 406 |
| users | 1 | sex_age | 1 | sex | 2 |
| users | 1 | sex_age | 2 | age | 406 |
| users | 1 | name | 1 | name | 38149 |
| users | 1 | active | 1 | active | 2 |
+-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐-‐+
27. Cardinality and distribution
27
Cardinality
version 5.5
Cardinality
version 5.6
count(distinct) Distribution
Filter
efficiency
PRIMARY 10.000.267 9.728.224 10.000.000 única optimum
name 18,868 38,149 10,001 ±0,01% c/u very good
age 18 192 101 ±1% c/u good
sex 18 2 2 50% c/u fair
active 18 2 2
‘0’: 0,1%
‘1’: 99,9%
‘0’ very good
‘1’ very bad
30. EXPLAIN fields
id Id of the SELECT, if there are more than one
select_type Query type (SIMPLE, UNION, SUBQUERY…)
table Table name
type Record search strategy
possible_keys Indexes to consider
key Index used
key_len Used length of the index
ref Fields matched with the index
rows Approximate record number to consider
filtered Percentage of filtered records by WHERE
Extra Additional info
30
31. Search strategies
from most to less optimal
• const: just one record, by primary key or unique index
• eq_ref: just one record for each other record in a JOIN
• ref: multiple records filtering by index
• index_merge: multiple records filtering by more than one index
• range: multiple records filtering by range using index
• index: read all records in the index file (index scan)
• ALL: read all records in the data file (full scan)
32. Information in the “Extra” field
• Using where: records are filtered after reading them, using the
WHERE condition (an index wasn’t able to filter all of them)
• Using index: all field data is read directly from the index, without
accessing the data file (covering index)
• Using where, Using index: as with “Using where”, records are
filtered after reading them, but data comes from an index
• Using filesort: extra step to sort after filtering records, when
records were not read in order from an index, using a temporary file
• Using temporary: temporary tables are needed to complete some
steps and satisfy the query (in memory or disk)
33. Guide to understand EXPLAIN
• “Using index” will boost query performance (covering index),
specially with lots of results
• Avoid “Using filesort” and “Using temporary”, specially with lots
of results
• “Using where” in queries of the types “ALL” o “index” means
that is not possible to discard any record directly from the index,
and they will be filtered while reading all the records
• Watch how may records are going to be read approximately
according to the field “rows”
• Watch the effective used length of the index according to the field
“key_len”
33
34. Just one record by Primary Key
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats
-‐> WHERE s_query = 'account executive'
-‐> AND fk_i_id_tbl_type_dates = 1
-‐> AND d_date = '2009-‐09-‐28'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: const
possible_keys: PRIMARY,s_query_i_num_ads_salary_d_date,d_date
key: PRIMARY
key_len: 774
ref: const,const,const
rows: 1
34
35. Just one record by UNIQUE KEY
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats_NEW
-‐> WHERE s_query = 'account executive'
-‐> AND fk_i_id_tbl_type_dates = 1
-‐> AND d_date = '2009-‐09-‐28'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats_NEW
type: const
possible_keys: unique_key,s_query_i_num_ads_salary_d_date,d_date
key: unique_key
key_len: 774
ref: const,const,const
rows: 1
35
36. JOIN of just one record
mysql> EXPLAIN SELECT * FROM tbl_users LEFT JOIN tbl_countries
-‐> ON tbl_users.fk_c_id_tbl_countries = tbl_countries.c_idG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: tbl_users
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 20986861
*************************** 2. row ***************************
id: 1
select_type: SIMPLE
table: tbl_countries
type: eq_ref
possible_keys: PRIMARY
key: PRIMARY
key_len: 6
ref: trovit_global.tbl_users.fk_c_id_tbl_countries
rows: 1
36
37. Many records using a partial
Unique Key
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats_NEW
-‐> WHERE s_query = 'account executive'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats_NEW
type: ref
possible_keys: unique_key,s_query_i_num_ads_salary_d_date
key: unique_key
key_len: 767
ref: const
rows: 1073
37
38. Many records using index
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats
-‐> WHERE d_date = '2012-‐10-‐15'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: ref
possible_keys: d_date
key: d_date
key_len: 3
ref: const
rows: 10908
38
39. OR condition and indexes
mysql> EXPLAIN SELECT a FROM t
-‐> WHERE b = 1 OR c = 1G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: t
type: index_merge
possible_keys: b,c,b_a
key: b,c
key_len: 4,4
ref: NULL
rows: 4863128
Extra: Using union(b,c); Using where
39
40. Range query using index
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats_NEW
-‐> WHERE s_query = 'account executive’
-‐> AND fk_i_id_tbl_type_dates = 1
-‐> AND d_date < ‘2011-‐04-‐25'G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_bl_trovit_stats
type: range
possible_keys: unique_key,s_query_i_num_ads_salary_d_date,d_date
key: unique_key
key_len: 774
ref: NULL
rows: 539
40
41. Condition is not using index,
but using covering index
mysql> EXPLAIN SELECT s_query FROM jobs_tbl_trovit_stats
-‐> WHERE i_num_ads_salary = 10G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: index
possible_keys: NULL
key: s_query_i_num_ads_salary_d_date
key_len: 775
ref: NULL
rows: 5543853
Extra: Using where; Using index
41
42. Full scan query, filter condition
doesn’t use index
mysql> EXPLAIN SELECT i_num_ads FROM jobs_tbl_trovit_stats
-‐> WHERE i_num_ads_salary = 10G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 5543853
Extra: Using where
42
43. Full scan query, retrieve all
records
mysql> EXPLAIN SELECT i_num_ads FROM jobs_tbl_trovit_statsG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 5543853
Extra: NULL
43
45. 45
KEY `a_b_c_d` (`a`,`b`,`c`,`d`) + id
KEY `a_b_c_d` (`a`,`b`,`c`,`d`) + id
mysql> EXPLAIN SELECT a, b, d FROM t
-‐> WHERE a = 1 ORDER BY idG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: t
type: ref
possible_keys: a,a_b,a_b_c_d
key: a_b_c_d
key_len: 4
ref: const
rows: 199704
Extra: Using where; Using index; Using filesort
200000 rows in set (0.28 sec)
The importance of “using index”
query optimizer choses wisely and favours the performance boost of
“using index” even if it’s forced to “using filesort”
46. 46
KEY `a` (`a`) + id
mysql> EXPLAIN SELECT a, b, d FROM t
-‐> FORCE KEY (a)
-‐> WHERE a = 1 ORDER BY idG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: t
type: ref
possible_keys: a
key: a
key_len: 4
ref: const
rows: 199704
Extra: Using where
200000 rows in set (0.53 sec)
The importance of “using index”
if we choose to force an index we think it would be better for performance (in this case the
bad extra “using filesort” is not used), we might be mistaken and the query could be slower
47. 47
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats
-‐> ORDER BY d_dateG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: jobs_tbl_trovit_stats
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 5543853
Extra: Using filesort
5066959 rows in set (3 min 27.96 sec)
But sometimes query optimizer is wrong
full scan query, it prefers to read only the data file and do a filesort, instead of using an
index to first: read keys directly in order and second: access the data file to retrieve fields
48. 48
mysql> EXPLAIN SELECT * FROM jobs_tbl_trovit_stats
-‐> FORCE KEY (d_date)
-‐> ORDER BY d_dateG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: homes_tbl_my_searches
type: index
possible_keys: NULL
key: d_date
key_len: 3
ref: NULL
rows: 5543853
5066959 rows in set (18.11 sec)
But sometimes query optimizer is wrong
same query, but this time is an index scan, with millions of records, is much better to force
the use of an index that satisfies the sorting of records, even if we must do a second read
from the data file to retrieve the fields for every record
50. Denormalize dates
don’t use DATE_FORMAT in WHERE, GROUP BY or ORDER BY
mysql> EXPLAIN SELECT SUM(f_revenue_in_euros*f_revenue_share/100) AS revenue,
-‐> DATE_FORMAT(d_date, “%Y-‐%m”) FROM tbl_publishers_stats
-‐> WHERE fk_i_id_tbl_publishers = ‘1658'
-‐> GROUP BY YEAR(d_date), MONTH(d_date)G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: tbl_publishers_stats
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 323826
Extra: Using where; Using temporary; Using filesort
50
51. Denormalize dates
use specific indexed fields for year, month, day, etc
mysql> EXPLAIN SELECT SUM(f_revenue_in_euros*f_revenue_share/100) AS revenue,
-‐> i_year, i_month FROM tbl_publishers_stats
-‐> WHERE fk_i_id_tbl_publishers = ‘1658'
-‐> GROUP BY i_year, i_monthG
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: tbl_publishers_stats_NEW
type: ref
possible_keys: publisher_year_month
key: publisher_year_month
key_len: 5
ref: const
rows: 236
Extra: Using where
51
52. Avoid using “*”
• Avoid using “*” in the field list of a SELECT
• Put only the fields you need in order to save unneeded disk reads
• Big extra optimisation if all the fields belong to the index used in
the query (covering index)
• Exception for using “*”: count(*) must always be used instead of
count(field_name):
• To avoid causing confusion to the query optimiser
• Avoids future errors if the query is modified and the field
inside the count() function is no longer indexed
52
53. Page queries
• Avoid running any long executing query
• Long queries block the execution of MySQL internal
maintenance processes, like the purge history
growing several gigabytes that would never shrink
again
• Long INSERT, UPDATE and DELETE, in addition to the
above, will delay replication to slaves
• Use LIMIT in queries of any type to page in smaller
and faster blocks
54. Using LIMIT the right way
• LIMIT filters final results of the query, only after
WHERE, GROUP BY and ORDER BY are
processed
• Beware of GROUP BY and ORDER BY “using
filesort”, all records are going to be file sorted
before LIMIT could take effect
• Even when using an index, the LIMIT
<offset>,<row_count> syntax will run slower
and slower as the offset increments
54
55. Using LIMIT the right way
always avoid “using filesort”
55
mysql> EXPLAIN SELECT * FROM homes_tbl_my_searches
-‐> ORDER BY s_where
-‐> LIMIT 10G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: homes_tbl_my_searches
type: ALL
possible_keys: NULL
key: NULL
key_len: NULL
ref: NULL
rows: 746475
Extra: Using filesort
10 rows in set (4.00 sec)
56. Using LIMIT the right way
always avoid “using filesort”
56
mysql> EXPLAIN SELECT * FROM homes_tbl_my_searches
-‐> ORDER BY s_what
-‐> LIMIT 10G
*************************** 1. row ***************************
id: 1
select_type: SIMPLE
table: homes_tbl_my_searches
type: index
possible_keys: NULL
key: what_where
key_len: NULL
ref: NULL
rows: 10
10 rows in set (0.00 sec)
57. # Slower as offset increments
mysql> SELECT i_id FROM jobs_tbl_trovit_stats_NEW
-‐> LIMIT 5000000,100;
100 rows in set (0.92 sec)
# Always fast
mysql> SELECT i_id FROM jobs_tbl_trovit_stats_NEW
-‐> WHERE i_id > $last_id
-‐> LIMIT 100;
100 rows in set (0.00 sec)
Using LIMIT the right way
• When paging, avoid LIMIT <offset>,<row_count>
it’s better to filter using a primary or unique key
57
58. LIMIT for long UPDATE and DELETE
• Divide long UPDATE and DELETE queries in several
shorter executions using LIMIT
• If possible, use indexed fields to find records
58
# Crontab to purge old records
Do
mysql> DELETE FROM homes_tbl_my_searches
-‐> WHERE dt_date < ‘2013-‐12-‐31’
-‐> LIMIT 1000;
While(rows affected > 0)
# One-‐time UPDATE, not worth to create an index just for this time
Do
mysql> UPDATE homes_tbl_my_searches SET i_active = 0
-‐> WHERE i_active != 0 AND s_what = ‘offensive stopword’
-‐> LIMIT 1000;
While(rows changed > 0)