MariaDB workshop conducted May 2017

1. MariaDB workshop
Alex Chistyakov, Git in Sky

2. Outline
- Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

3. Outline
- Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

4. What’s in a box?
- Ubuntu 16.04.2
- Python 2.7.12
- MariaDB 10.0.29
- Sakila DB, Employees DB
- Percona Toolkit 2.2.16
- Anemometer

5. How to use Vagrant
- Create an empty folder
- Download https://goo.gl/ap6r6E there (rename it to
‘Vagrantfile’)
- Run ‘vagrant up’ in that folder
- Wait until a VM starts
- Run ‘vagrant ssh’ to get in
- My .mysql_history: https://goo.gl/AyrTW7

6. - Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

7. What is a table?
- A collection of related data

8. What is a table?
- A collection of related data
- Consists of columns and rows

9. What is a table?

10. The DDL
- Manipulates the database structure (also called schema)

11. DDL statements
- CREATE
- ALTER
- DROP
- TRUNCATE
- RENAME

12. How to create a table?
CREATE TABLE language (
language_id TINYINT UNSIGNED NOT NULL AUTO_INCREMENT,
name CHAR(20) NOT NULL,
last_update TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP,
PRIMARY KEY (language_id)
)ENGINE=InnoDB DEFAULT CHARSET=utf8;

13. Primary keys
- Identify a record uniquely
- So, adding two equal keys is not possible
- Can be natural like “passport number”
- Or surrogate
- Surrogate keys are auto-generated on the DB side

14. Natural PKs can be composite
CREATE TABLE film_actor (
actor_id SMALLINT UNSIGNED NOT NULL,
film_id SMALLINT UNSIGNED NOT NULL,
last_update TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP,
PRIMARY KEY (actor_id,film_id),
KEY idx_fk_film_id (`film_id`),
CONSTRAINT fk_film_actor_actor FOREIGN KEY (actor_id) REFERENCES actor (actor_id),
CONSTRAINT fk_film_actor_film FOREIGN KEY (film_id) REFERENCES film (film_id)
)ENGINE=InnoDB DEFAULT CHARSET=utf8;

15. Autoincrement primary keys
- Are surrogate
- Are 1,2,3,4 or 8 bytes long
- BTW, INT(10) is 4 bytes long
- Are incremented on every INSERT
- Should be always used
- BTW, InnoDB table is a clustered index* around its PK
- If no explicit PK exists 6-byte row ID will be used

16. Exercise #1
- Create a table

17. - Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

18. A trivial SELECT query
- SELECT * FROM employees WHERE hire_date='1986-06-26'
- Please, never use “SELECT *”, always select certain
columns!
- A slightly better version:
- SELECT emp_no, first_name, last_name FROM employees
WHERE hire_date='1986-06-26';

19. Using a single table is impractical*
- Four types of JOINs:
- INNER JOIN
- LEFT OUTER JOIN
- RIGHT OUTER JOIN
- CROSS JOIN
- Left and right outer joins are equivalent

20. Left outer join example
- SELECT e.emp_no, first_name, last_name, salary FROM
employees e LEFT OUTER JOIN salaries s on e.emp_no =
s.emp_no WHERE hire_date='1986-06-26';
- This query selects an employee even if no payment
records exist in the salaries table

21. Aggregate queries and GROUP BY
- SELECT e.emp_no, first_name, last_name, SUM(salary)
FROM employees e LEFT OUTER JOIN salaries s on
e.emp_no = s.emp_no WHERE hire_date='1986-06-26'
GROUP BY e.emp_no;

22. How to get people w/no salary recs
- INSERT INTO employees(emp_no, first_name, last_name)
VALUES(600000, 'Alex', 'Chistyakov');
- Let’s count number of salary records using COUNT()
aggregate function

23. HAVING is like WHERE
- SELECT e.emp_no, first_name, last_name,
COUNT(salary) FROM employees e LEFT OUTER JOIN
salaries s on e.emp_no = s.emp_no GROUP BY e.emp_no
HAVING COUNT(salary) = 0;

24. Another way to do the same
- SELECT e.emp_no, first_name, last_name,
COUNT(salary) FROM employees e LEFT OUTER JOIN
salaries s on e.emp_no = s.emp_no WHERE s.emp_no IS
NULL;
- This query is more optimal*

25. Exercise #2
- Write a SELECT query which get all employees with total
sum of all salary records greater than 40000

26. - Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

27. Why indexes?
- Latency Numbers Every Programmer Should Know:
https://goo.gl/v4CEWU
- Indexes helps to avoid unnecessary disk operations

28. How indexes work?
- Index is a data structure optimized for search
- There are several types of indexes: hash indexes, B-tree
indexes
- Hash indexes allow to find exact rows
- B-tree indexes allow to find ranges
- InnoDB and Aria support B-tree indexes only

29. B-tree index
- “B” stands for “balanced”, not for “binary”

30. “SQL Tuning” by Dan Tow
- https://goo.gl/jRbD5H
- A must read for every DBA!
- Discusses how to build effective
indexes in great details
- Unfortunately does not cover
aggregate functions and sorting

31. Column cardinality
- Cardinality is a measure of data uniqueness
- Columns with more unique values have higher cardinality
- Columns with few unique values have lower cardinality

32. A composite index
- Covers two or more columns
- Allows to find rows by subsequently applying a filter
column-by-column
- Order of columns in a composite index matters!

33. Index selectivity
- An ability of a certain condition to filter
- Is expressed as a number of columns after filtering
divided by a total number of columns
- Lower values mean greater selectivity
- Some authors define selectivity as a total number of
columns divided by a resulting number of columns

34. Building a good composite index
- Columns with higher individual selectivity should go first
in a composite index
- Non-selective columns should be the latest

35. Functional indexes
- Original MySQL does not have functional indexes
- MariaDB adds support for virtual columns
- Functional indexes can be created over virtual columns

36. Virtual column example
- ALTER TABLE employees ADD lower_last_name
varchar(16) GENERATED ALWAYS AS (lower(last_name))
PERSISTENT;
- CREATE INDEX lower_last_name ON
employees(lower_last_name);
- SELECT e.emp_no, first_name, last_name FROM
employees e WHERE lower_last_name LIKE 'chistya%';

37. Let’s add %
- SELECT e.emp_no, first_name, last_name FROM
employees e WHERE lower_last_name LIKE '%chistya%';
- This will always lead to a full scan in current MariaDB
and MySQL implementations
- Full Text Search engine should be used instead
- I recommend Sphinx or Solr

38. Using ORDER BY
- In most real life cases can’t be covered by an index
- Dan Tow doesn’t consider these cases at all
- No good solution exists

39. Exercise #3
- Write a select which gets all salary records for the
employee w/emp_no = 10001 ordered by amount of the
salary record
- Create a covering index for this query

40. Things not to do in your life
- Please never ever do ORDER BY RAND()!
- How to do it properly: get a good random number on the
client side
- LIMIT 50 OFFSET 5000000 is the next thing not to do
- How to do it properly: “emp_no > $last_emp_no LIMIT
50”

41. - Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

42. A bit of history
- MySQL supported pluggable storage engines for years
- Two most notable were MyISAM and InnoDB
- MyISAM did not support transactions in any way
- InnoDB was transactional

43. MariaDB engines
- Many mysql.* tables are still in MyISAM format
- Aria storage engine emerged and is optionally
transactional in a crash-proof sense (does not support
explicit transactions though)
- InnoDB fully supports transactions
- I recommend to use InnoDB

44. A bit of InnoDB internals
- /var/lib/mysql/ib_logfile[01] are InnoDB redo logs
- The redo log works as a circular buffer
- It’s not practical to set the InnoDB log size
(innodb_log_file_size) to more than 128M
- This change requires restart

45. Generic recovery process
- Works the same way for any engine with WAL/redo
log/intent log/whatever
- The service starts after crash
- Log records are examined
- Finished transactions are applied to their final
destinations, unfinished ones are thrown out
- Aria performs these steps when in transaction mode too

46. COMMIT and auto-commit
- Every query starts and commits an implicit transaction
by default
- SET autocommit = 0; disables this
- START TRANSACTION or BEGIN should be used then to
start a transaction
- And COMMIT to finish it
- DDL statements perform COMMIT implicitly

47. ROLLBACK and savepoints
- ROLLBACK is used to abort a transaction
- Transactions can’t be nested but this behavior can be
emulated using savepoints
- SAVEPOINT label
- ROLLBACK TO label
- RELEASE SAVEPOINT label

48. A bit of InnoDB internals - MVCC
- MVCC stands for “Multiversion concurrency control”
- Records are declared dead but still occupy disk space
- InnoDB storage file never shrinks
- InnoDB uses a single file for everything by default and
this file can’t be compacted

49. It’s possible to overcome this
- innodb_file_per_table=1
- Every table will occupy a separate file (two separate
files in fact)
- Beware of Unix file descriptors limits!
- ulimit -n 65535 somewhere before starting mysqld_safe

50. Long transactions can be evil
- DDL statements require an exclusive lock on table
metadata
- An explicit transaction holds a read lock on every table it
uses
- If number of transactions per second is high enough the
DDL statement will wait forever

51. Transactions: logical perspective
- The SQL standard defines 4 transaction isolation levels
- READ UNCOMMITTED
- READ COMMITTED
- REPEATABLE READ
- SERIALIZABLE

52. READ UNCOMMITTED
- The weakest level
- Allows dirty reads
- A transaction can get non-committed data of other
transaction

53. READ COMMITTED
- Non-repeatable reads are possible
- Phantom reads are possible

54. REPEATABLE READ
- The default isolation level
- Non-repeatable reads are not possible
- Phantom reads are possible

55. SERIALIZABLE
- The strongest level
- Non-repeatable reads and phantom reads are impossible

56. Exercise #4
- Open two different connections to the employees DB, set
autocommit to 0;
- Set isolation level to READ COMMITED in both windows,
select total number of employees whose names started
with Alex in the 1st session, delete the employee with ID
499559 in the 2nd session (don’t forget to COMMIT),
repeat the query in the 1st session

57. Exercise #4
- Set isolation level to REPEATABLE READ in both windows,
select total number of employees whose names started
with Alex in the 1st session, delete the employee with ID
499517 in the 2nd session (don’t forget to COMMIT),
repeat the query in the 1st session

58. Exercise #4
- Set isolation level to REPEATABLE READ in both windows,
select total number of employees whose names started
with Alex in the 1st session, insert an employee called
Alexis Doe in the 2nd session (don’t forget to COMMIT),
repeat the query in the 1st session

59. Exercise #4
- Set isolation level to SERIALIZABLE in both windows,
select total number of employees whose names started
with Alex in the 1st session, insert an employee called
Alex Didnotfail in the 2nd session (don’t forget to
COMMIT), repeat the query in the 1st session

60. - Tables and DDL
- Queries and DML
- Indexes and compound indexes
- Transactions and how they work, isolation levels
- Authorization and authentication, client protocol

61. mysql.user table
- Stores user privileges
- Can (but should not) be manipulated directly
- FLUSH PRIVILEGES rereads effective rights from it
- Uses MyISAM storage

62. GRANT statement
- Creates user accounts
- Grants privileges to them
- Is documented at https://goo.gl/zBHTd4

63. A superuser
- Has ALL PRIVILEGES ON *.*
- Has a number of SUPER privileges

64. A list of privileges
- Privileges can be global, database level, table level,
column level, function level and procedure level
- A list is available in GRANT command documentation

65. Default client credentials
- Can be set in ~/.my.cnf file like this:
[client]
user = root
password = Pheexaigee8a

66. Using views to limit rights
- Create a view using a privileged table columns
- Grant privileges to that view

67. Using stored procedures
- Create a stored procedure to perform AAA tasks
- Grant privileges to that stored procedure

68. MySQL wire protocol
- Is encrypted using a session key
- Can’t be easily proxied on L3 because of that

69. Exercise #5
- Grant all privileges on the employees.salaries table to a
user called “manager” with password da5ca9aeNgee%, a
user can connect from any host
- Create a view on a table employees consisting of emp_no
and the first and last names and grant a read privilege on
it to a user called “reader” with password eLegah0aez8a

70. - Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

71. MySQL slow queries log
- The simplest way to do performance tuning
- Should be enabled in the MariaDB config file
- Slow queries will be written to a file for subsequent
analysis

72. Slow queries log config vars
- slow_query_log = on
- slow_query_log_file = /var/log/mysql/mariadb-slow.log
- long_query_time = 0.1
- log-queries-not-using-indexes

73. Analyzing the log w/Percona Tools
- pt-query-digest
- Documented at https://goo.gl/YCv1ya
- In the simplest case produces a textual report on most
time-consuming queries

74. Analyzing the log w/Anemometer
- Anemometer is a web-based slow query monitor created
at Box (https://github.com/box/Anemometer)
- Anemometer uses pt-query-digest to process the slow
query log internally
- Anemometer requires PHP, a webserver and a number of
other tools
- So, we use an Ansible role to simplify its deployment

75. Ansible role for Anemometer
- Ansible is a popular Configuration Management tool
- Ansible is written in Python and uses YAML as a
configuration description language
- A role for Anemometer is at https://goo.gl/us6V82
- This role works for Ubuntu 14.04 hosts and does not work
for 16.04 yet
- This is trivial to correct, expect a fix in a week

76. Demo time!
- Let’s analyze live queries in our Vagrant box

77. Partitioning and sharding
- Partitioning is a process of splitting a big table in smaller
subset on the same server
- Partitioning works well for time-series data
- Sharding is a process of splitting a big table in a number
of unrelated tables on different servers
- Sharding requires serious modifications of the app code

78. Partitioning in MariaDB
- MariaDB inherits MySQL support for partitioning
- Partitioning is documented at https://goo.gl/1CwIKX
- Certain limitations apply:
- Queries are not parallelized
- Partitioned table can’t contain or be referenced by
foreign keys

79. Partitioning in the real life
- Is tricky to set up properly
- Is often misused (I personally have never seen MySQL
partitioning set up properly)
- I strongly recommend not to use partitioning

80. Exercise #6
- Get familiar with the Anemometer tool
- Read and explain a query plan

81. - Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

82. What is replication?
- Storing the same data on multiple MariaDB servers
- Establishing a master/slave relationship between the
original and the copies
- Distributing data modifications from a master node to
slave nodes

83. Master and slave nodes
- The master node gets data modification queries
(INSERTs, UPDATEs and DELETEs)
- The master node sends data changes to slaves
- Slave nodes are read-only and get updates from the
master
- Data modification on slave nodes is not prohibited in
MySQL/MariaDB world

84. Types of replication
- Replication can be synchronous or asynchronous
- Replication can also be master-slave or master-master
- All 4 options are possible: “synchronous master-slave”,
“asynchronous master-slave”, “synchronous
master-master” and “asynchronous master-master”
- Asynchronous master-slave is the default MariaDB setting

85. Master-slave and master-master
- There is only a single master in a MS replication topology
- There is more than one master in a MM setup
- A master should propagate data changes to all hosts in
the replication topology
- So, every master is also a slave in a MM setup

86. Sync or async
- Async: a transaction on a master is finished as soon as
it’s written to a transaction log on a master
- Semisync: a transaction on a master is finished only
after it’s written to a transaction log on one of slaves
- Sync: a transaction on a master is finished when it’s
acknowledged and committed on all slaves

87. Replication lag
- Replication lag is a delay between the same operations
on a master and on a slave
- Replication lag is meaningful for async replication only
- Replication lag should be minimized

88. Multi-master replication scalability
- Multi-master replication does not scale on writes!
- It’s a popular belief that it does (because there is more
than one master)
- But every master should perform exactly the same set of
write operations!

89. Multi-master tips and tricks
- Avoid writing to the same table on different masters!
- Split your schema to several non-related table sets
logically bound to different services if possible
- Work with these table sets on different masters
independently

90. - Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

91. The binary log
- The binary log stores data modification events (both DDL
and DML changes)
- The binary log is storage neutral (works for Aria, InnoDB,
etc.)
- The binary log is not a transaction log
- The binary log can store events in 3 different formats

92. Binary log formats
- SBR (statement-based replication)
- RBR (row-based replication)
- Mixed (stores statements or rows when appropriate)
- Mixed seems to be the best of both worlds
- But it is not, in fact (avoid using it)

93. Statement-based replication
- Stores INSERT/UPDATE/DELETE and
CREATE/DROP/TRUNCATE statements as is
- Requires less space in the log
- Is not 100% accurate for all statements

94. SBR non-determinism
- INSERT INTO t1(c1, mtime) VALUES(1, NOW())
- NOW() can be different on master and slave
- INSERT INTO t2(c1, c2) VALUES(1, RAND())
- RAND() is definitely different on master and slave
- Fixes are trivial - master should send exact values
- DELETE FROM t1 LIMIT 10; - fix is not trivial

95. SBR is broken (mixed is broken too)
- Error 1062 (Duplicate entry NNN for key X)
- But why?..I just inserted a bunch of rows!
- This is a bug somehow related to range locking on a
primary key on slave side
- There is a lot of instructions on the Internet, something
like “set slave-skip-errors to 1062”

96. Never trust random Internet guys
- Don’t do “slave-skip-errors”
- To fix this bug properly…
- NEVER USE SBR OR MIXED LOG FORMATS, USE RBR!
- The only problem is that RBR is broken too

97. The binary log concept is broken
- Correctly implemented binary log stores physical changes
to the storage layer (WAL records)
- MySQL historically used pluggable storage layers, some of
them were non-transactional
- The binary log is on the wrong abstraction layer
- This can’t be easily fixed

98. RBR is broken (much less than SBR)
- DELETE FROM t1; generates a lot of rows to be written to
the binary log
- The slave can begin lagging
- A slave SQL thread uses indexes to apply row deltas
- Having a primary key is inevitable!
- It’s better to use surrogate keys

99. libslave
- A library to mimic a MySQL slave
- https://github.com/tarantool/libslave
- Can be embedded to an app, allows an app to connect to
the MySQL master and read the binlog

100. Cascading replication topologies
- Replication can (and should be)
cascaded (5 slaves on a single master
is a bad idea)
- A slave can be a master for a slave
- Config should be tweaked:
log-slave-updates=1

101. Replication rings
- If you absolutely need
master-master, you can have one
- Every master should have its own
key space
- auto_increment_offset=1
auto_increment_increment=10

102. Semisync replication
- Added since MariaDB 5.5, declared stable since 10.1.3
- Documented at https://goo.gl/wuiKfJ
- If a slave fails to acknowledge before a certain timeout,
a master switches to async automatically and switches
back when a slave catches up

103. Parallel replication
- Traditional MariaDB replication uses a single SQL thread
on the slave side
- Starting with 10.0.5 it’s possible to use several threads
- Documented at https://goo.gl/0p4SH9

104. Delayed replication
- Replication is not a backup!
- Delayed replication is (well, can be)
- Introduced in MariaDB 10.2.3
- Documented at https://goo.gl/BZguD9
- Replication delay can be achieved using pt-slave-delay
tool from Percona Toolkit

105. GTID
- Globally unique binlog events identification
- Introduced in 10.0.2
- Documented at https://goo.gl/xgJ27M
- Has a number of significant benefits: slave server can be
easily reconnected to another master, slave log position
is saved in a transactional way

106. - Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

107. WSREP
- WSREP is a library for distributing working sets
- The Galera cluster is built around that library

108. The Galera cluster
- Is InnoDB-only
- Is semisync
- Does not use traditional replication at all

109. A common Galera cluster setup
- Two master nodes and one arbiter node
- The arbiter node does not store anything

110. - Basics of performance monitoring
- Notion of replication, types of replication
- Traditional replication in details
- Galera cluster and how it works
- MMM, PRM and query proxying

111. Questions?
- Please feel free to email me at alex@gitinsky.com
- My Skype ID is demeliorator

112. Thank you!
- Good luck in the wonderful world of MariaDB!