Automated master failover

Automated, Non-Stop
MySQL Operations and Failover

Yoshinori Matsunobu

Principal Infrastructure Architect, DeNA
Former APAC Lead MySQL Consultant at MySQL/Sun/Oracle
Yoshinori.Matsunobu@gmail.com
http://yoshinorimatsunobu.blogspot.com/

Table of contents

 Automating master failover (main topic)

 Minimizing downtime at master maintenance

Company Introduction: DeNA and Mobage

 One of the largest social game providers in Japan
 Both social game platform and social games themselves
 Subsidiary ngmoco:) in SF
 Japan localized phone, Smart Phone, and PC games
 2-3 billion page views per day
 25+ million users
 700+ MySQL servers
 1.3B$ revenue in 2010

HA Requirements for social games

 Requirements about high availability and integrity are
quite high
 Paid service dramatically raises expectations from users
– “I haven’t received a virtual item I paid for”
– “My HP/MP fell after I used non-free recovery item”
 Long downtime causes huge negative impacts on revenue
 Planned maintenance is not impossible, if properly planned
and announced
– Traffic at 5 am is less than 1/5 compared to 11 pm
– Much better than unplanned downtime

The goal is “No Single Point of Failure”
 We operate 700+ MySQL servers at DeNA
 More than 150 {master, slaves} pairs
 Mainly MySQL 5.0 and 5.1
 Statistically MySQL master went down once per a few months
 In many times caused by hangs on Linux or H/W failures
 Manual failover should be avoided if possible, to minimize downtime
 It is easy to make slaves not single point of failure
 Just running two or more slaves
 It is not trivial to make masters not single point of failure

 We want to automate master failover and slave promotion
 On regular MySQL 5.0/5.1, and 5.5+
– We don’t want to spend time for significant architecture changes on legacy
running services
 Without losing performance significantly
 Without spending too much money

Master Failover: What is the problem?
Writer IP Writer IP

master
slave1-> New Master

slave1 slave2 slave3 slave2: slave3:
Problem: CHANGE MASTER CHANGE MASTER
When a master goes down, the system also goes down until *manual* master
failover completes (you can’t do writes). It is not uncommon to take one hour or
even more to recover.

Objective:
Automate master failover. That is, pick one of the appropriate slaves as a new
master, making applications send write traffics to the new master,
then starting replication again.

Failure Example (1)
Writer IP All slaves have received all binlog events
from the crashed master.
master
Any slave can be a new master,
id=99 without recovering any data
id=100
id=101
Example: picking slave 1 as a new master
Get current binlog position (file1,pos1)
Grant write access
Slave 2 and 3 should execute
Activate writer IP address
CHANGE MASTER MASTER_HOST=
‘slave1’ …;
START SLAVE;
slave1 slave2 slave3
This is the easiest scenario.
id=99 id=99 id=99 But not all times it is so lucky.
id=100 id=100 id=100
id=101 id=101 id=101
Execute CHANGE MASTER TO MASTER_HOST=‘slave1’,
MASTER_LOG_FILE=‘file1’, MASTER_LOG_POS=pos1;

Failure Example (2)
All slaves have received same binlog events
from the crashed master.
master
But the crashed master has some events
id=99 that have not been sent to slaves yet.
id=100
id=101
id=102 id=102 will be lost if you promote
Copy and apply events (id=102) one of slaves to a new master.

Start Master
If the crashed master is reachable (via SSH)
and binlog file is readable, you should
save binlog (id=102) before promoting
slave1 slave2 slave3 a slave to a new master.
id=99 id=99 id=99
id=100 id=100 id=100 Using Semi-Synchronous replication
id=101 id=101 id=101 greatly reduces the risk of this scenario.
id=102 id=102 id=102
CHANGE MASTER

Failure Example (3)
Writer IP
Some slaves have events which
other slaves have not received yet.
master
id=99 You need to pick events from the
id=100 latest slave (slave 2), and apply to
id=101
other slaves so that all slaves will be
Start Master
consistent.
(Sending id=101 to slave 1, sending
id=100 and 101 to slave 3)
id=99 id=99 The issues are:
id=99
id=100 id=100 id=100 - How can we identify which binlog
id=101 id=101 id=101 events are not sent?
Identify which events are not sent - How can we make all slaves
Apply lost events eventually consistent?
CHANGE MASTER

Master Failover: What makes it difficult?
Writer IP MySQL replication is asynchronous.

master It is likely that some (or none of) slaves have
not received all binary log events from the
id=99 crashed master.
id=100
id=101
id=102 It is also likely that only some slaves have
received the latest events.
Save binlog events that
exist on master only
In the left example, id=102 is not replicated to
any slave.

slave1 slave2 slave3 slave 2 is the latest between slaves, but
id=99 id=99 id=99
slave 1 and slave 3 have lost some events.
id=100 id=100 id=100 It is necessary to do the following:
id=101 id=101 id=101 - Copy id=102 from master (if possible)
id=102 id=102 id=102
Identify which events are not sent - Apply all differential events, otherwise data
Apply lost events inconsistency happens.

Current HA solutions and problems
 Heartbeat + DRBD
 Cost: Additional passive master server (not handing any application traffic) is needed
 Performance: To make HA really work on DRBD replication environments, innodb-
flush-log-at-trx-commit and sync-binlog must be 1. But these kill write performance
 Otherwise necessary binlog events might be lost on the master. Then slaves can’t
continue replication, and data consistency issues happen

 MySQL Cluster
 MySQL Cluster is really Highly Available, but unfortunately we use InnoDB

 Semi-Synchronous Replication (5.5+)
 Semi-Sync replication greatly minimizes the risk of “binlog events exist only on the
crashed master” problem
 It guarantees that *at least one* (not all) slaves receive binlog events at commit. Some
of slaves might not receive all binlog events at commit.

 Global Transaction ID
 On mysql side, it’s not supported yet. Adding global transaction Id within binary logs
require binlog format change, which can’t be done in 5.1/5.5.
– Check Google’s Global Transaction ID patch if you’re interested
 There are ways to implement global tx ID on application side, but it’s not possible
without accepting complexity, performance, data loss, and/or consistency problems

More concrete objective
 Make master failover and slave promotion work
 Saving binary log events from the crashed master (if possible)
– Semi-synchronous replication helps too
 Identifying the latest slave
 Applying differential relay log events to other slaves
 Applying saved binary log events from master
 Promoting one of the slaves to a new master
 Making other slaves replicate from the new master

 Automate the above procedure
 Master failure should also be detected automatically

 Do the above
 Without introducing too much complexity on application side
 With 5.0/5.1 InnoDB

Saving binlog events from (crashed) master
Dead Master Latest Slave Other Slaves

Lost events {Master_Log_File, Read_Master_Log_Pos} from
SHOW SLAVE STATUS (mysqld-bin.000013, 12345)
mysqlbinlog --start-position=12345 mysqld-bin.000013 mysqld-bin.000014….

 If the dead master is reachable via SSH, and binary logs are
accessible (Not H/W failure, i.e. InnoDB data file corruption on the
master), binlog events can be saved.
 Lost events can be identified by checking {Master_Log_File,
Read_Master_Log_Pos} on the latest slave + mysqlbinlog
 Using Semi-Synchronous replication greatly reduces the risk of
events loss

Understanding SHOW SLAVE STATUS
mysql> show slave status¥G
Slave_IO_State: Waiting for master to send event
Master_Host: master_host  {Master_Log_File,
Master_User: repl Read_Master_Log_Pos} :
Master_Port: 3306 The position in the current
Connect_Retry: 60 master binary log file up to
Master_Log_File: mysqld-bin.000980 which the I/O thread has read.
Read_Master_Log_Pos: 629290122
Relay_Log_File: mysqld-relay-bin.000005  {Relay_Master_Log_File,
Relay_Log_Pos: 26087338 Exec_Master_Log_Pos} :
Relay_Master_Log_File: mysqld-bin.000980 The position in the current
Slave_IO_Running: Yes master binary log file up to
Slave_SQL_Running: Yes which the SQL thread has read
Replicate_Do_DB: db1 and executed.
…
Last_Errno: 0  {Relay_Log_File,
Relay_Log_Pos} :
Last_Error:
Exec_Master_Log_Pos: 629290122 The position in the current
Seconds_Behind_Master: 0 relay log file up to which the
Last_IO_Errno: 0 SQL thread has read and
executed.
Last_IO_Error:
Last SQL Errno: 0

Identifying the latest slave
Slave 1 Slave 2 Slave 3

Relay log name slave1-relay.003300 slave2-relay.003123 slave3-relay.001234

{Master_Log_File, mysqld-bin.001221 mysqld-bin.001221 mysqld-bin.001221
Read_Master_ pos 102067 pos 102238 pos 101719
Log_Pos}

 Relay log name/position is not helpful to identify the latest
slave, because relay log name/position is independent from
slaves
 By comparing {Master_Log_File, Read_Master_Log_Pos},
you can identify the latest slave
 Slave 2 is the latest

Next issue: Applying diffs to other slaves

 How can we identify
which binlog events
need to be applied to
slave1 slave2 slave3 each slave?
Id=99 Id=99 Id=99
Id=100 Id=100 Id=100
Id=101 Id=101 Id=101
Identify which events are not sent
Apply lost events

Identifying what events need to be applied
Slave 1 Slave 2 Slave 3

slave2-relay.003123
slave3-relay.001234
slave1-relay.003300

{Master_Log_File, mysqld-bin.001221 mysqld-bin.001221 mysqld-bin.001221
Read_Master_ pos 102067 pos 102238 pos 101719
Log_Pos}

 Since we know all slave’s master position, by comparing these
positions, generating differential relay log events should be
possible
 There is no simple way to generate differential relay log events
based on master’s log file/position

Relay log internals: “at” and “end_log_pos”
[user@slave2] mysqlbinlog slave2-relay-bin.003123
# at 106
#101210 4:19:03 server id 1384 end_log_pos 0  “# at xxx” corresponds to relay log
Rotate to mysqld-bin.001221 pos: 4 position of the slave. This is not
… master’s binlog position. Each
# at 101835 slave might have different relay log
#110207 15:43:42 server id 1384 end_log_pos 101764 position for the same binary log
Query thread_id=1784 exec_time=0 error_code=0 event.
SET TIMESTAMP=1297061022/*!*/;
BEGIN /*!*/;  end_log_pos corresponds to the
# at 101910 master’s binary log position. This
#110207 15:43:42 server id 1384 end_log_pos 102067 is unique between slaves.
Query thread_id=1784 exec_time=0 error_code=0
SET TIMESTAMP=1297061022/*!*/;  At the beginning of the relay log
update ………………….. /*!*/; file, normally master’s binary log
# at 102213 file name is written.
#110207 15:43:42 server id 1384 end_log_pos 102211
Query thread_id=1784 exec_time=0 error_code=0  end_log_pos of the tail of the last
SET TIMESTAMP=1297061022/*!*/; relay log should be equal to
update ………………….. /*!*/; {Master_Log_File,
# at 102357 Read_Master_Log_Pos} from
#110207 15:43:42 server id 1384 end_log_pos 102238 SHOW SLAVE STATUS.
Xid = 12951490691
COMMIT/*!*/;
EOF

Relay log internals: How to identify diffs
[user@slave2] mysqlbinlog slave2-relay-bin.003123 [user@slave3] mysqlbinlog slave3-relay-bin.001234
… …
# at 101807 # at 234567
#110207 15:43:42 server id 1384 end_log_pos 101719 #110207 15:43:42 server id 1384 end_log_pos 101719
Xid = 12951490655 Xid = 12951490655
COMMIT/*!*/; COMMIT/*!*/;
# at 101835 EOF
Query thread_id=1784 exec_time=0 error_code=0  Slave 2 has received more binlog
SET TIMESTAMP=1297061022/*!*/; events than Slave 3
BEGIN /*!*/;
# at 101910
#110207 15:43:42 server id 1384 end_log_pos 102067  Check the last end_log_pos on the
Query thread_id=1784 exec_time=0 error_code=0 behind slave (101719 at Slave 3)
update ………………….. /*!*/;
# at 102213  Search Slave 2’s relay log where
#110207 15:43:42 server id 1384 end_log_pos 102211 end_log_pos == 101719
update ………………….. /*!*/;  Events from relay log position 101835
# at 102357 are lost on slave 3
 mysqlbinlog --start-position=101835
Xid = 12951490691
should be applied to slave 3
COMMIT/*!*/;
EOF

Relay log and “Partial Transaction”
Massive transactions
{Relay_Master_Log_File, Exec_Master_Log_Pos}

Slave
Binary Logs Relay Logs
Master …
…
BEGIN; BEGIN;
UPDATE… UPDATE… These events are
INSERT… INSERT… NOT executed forever
UPDATE… (EOF)
COMMIT;
(EOF)
{Master_Log_File, Read_Master_Log_Pos}

 Alive slave IO thread writes valid relay log events, so invalid (can’t read)
events should not be written to the relay log
 But if master crashes while sending binary logs, it is likely that only some
parts of the events are sent and written on slaves.
 In this case, slave does not execute the last (incomplete) transaction.
 {Master_Log_File, Read_Master_Log_Pos} points to the end of the relay log,
but {Relay_Master_Log_File, Exec_Master_Log_Pos} will point to the last
transaction commit.

Lost transactions
Relay_Log_Pos  In some unusual cases, relay
Exec_Master_Log_Pos logs are not ended with
(Current slave1’s data)
transaction commits
[user@slave1] mysqlbinlog mysqld-relay-bin.003300  i.e. running very long
# at 91807 transactions
Xid = 12951490655  Read_Master_Log_Pos always
COMMIT/*!*/; points to the end of the relay
# at 91835 log’s end_log_pos
#110207 15:43:42 server id 1384 end_log_pos 101764  Exec_Master_Log_Pos points
Query thread_id=1784 exec_time=0 error_code=0 to the end of the transaction’s
end_log_pos (COMMIT)
BEGIN
/*!*/;  In the left case,
# at 91910 Exec_Master_Log_Pos ==
Read_Master_Log_Pos is never
true
update …………………..  Slave 1’s SQL thread will never
/*!*/; execute BEGIN and UPDATE
statements
(EOF)
 Unapplied events can be
generated by mysqlbinlog –
Read_Master_Log_Pos start-position=91835

Recovering lost transactions
[user@slave2] mysqlbinlog mysqld-relay-bin.003123 [user@slave1] mysqlbinlog mysqld-relay-bin.003300
# at 106 # at 106
Rotate to mysqld-bin.001221 pos: 4 Rotate to mysqld-bin.001221 pos: 4
… …
# at 101807 # at 91807
Xid = 12951490655 Xid = 12951490655
COMMIT/*!*/; COMMIT/*!*/;
Relay_Log_Pos
# at 101835 # at 91835 (current slave1’s pos)
Query thread_id=1784 exec_time=0 error_code=0 Query thread_id=1784 exec_time=0 error_code=0
SET TIMESTAMP=1297061022/*!*/; SET TIMESTAMP=1297061022/*!*/;
BEGIN BEGIN
/*!*/; /*!*/; (A)
# at 101910 # at 91910
Query thread_id=1784 exec_time=0 error_code=0 Query thread_id=1784 exec_time=0 error_code=0
SET TIMESTAMP=1297061022/*!*/; SET TIMESTAMP=1297061022/*!*/;
update 1………………….. update 1…………………..
/*!*/; /*!*/; (EOF)
# at 102213
Query thread_id=1784 exec_time=0 error_code=0  The second update event is lost on
SET TIMESTAMP=1297061022/*!*/; slave 1, which can be sent from slave 2
update 2…………………..  The first update event is not executed
/*!*/; (B) on slave 1’s SQL thread
# at 102357
#110207 15:43:42 server id 1384 end_log_pos 102238  (A) + (B) should be applied on slave
Xid = 12951490691 1, wichin the same transaction
COMMIT/*!*/; (EOF)

Steps for recovery
Dead Master Latest Slave Slave(i)

Wait until SQL thread
executes all events

Final Relay_Log_File,
Relay_Log_Pos

(i1) Partial Transaction
Master_Log_File
Read_Master_Log_Pos

(i2) Differential relay logs from each slave’s read pos to
the latest slave’s read pos
(X) Differential binary logs from the latest slave’s read pos
to the dead master’s tail of the binary log

 On slave(i),
 Wait until the SQL thread executes events
 Apply i1 -> i2 -> X
– On the latest slave, i2 is empty

Design notes

 Trimming ROLLBACK events from mysqlbinlog
 Purging relay logs
 Identifying whether SQL thread has really executed
all events
 Handling malicious queries
 Parallel recovery on multiple slaves
 Row based format

mysqlbinlog and ROLLBACK events
[user@slave1] mysqlbinlog slave1-relay.003300 --position=91835  mysqlbinlog adds a ROLLBACK
# at 91835 statement at the end of the
generated file
SET TIMESTAMP=1297061022;
BEGIN
# at 91910
 mysqlbinlog may add a
#110207 15:43:42 server id 1384 end_log_pos 102067 ROLLBACK statement and/or an
Query thread_id=1784 exec_time=0 error_code=0 equivalent BINLOG event at the
SET TIMESTAMP=1297061022; beginning of the generated file
update ………………….. (included in the START event)
ROLLBACK; /* added by mysqlbinlog */

[user@slave2] mysqlbinlog slave2-relay.003123
# at 4
 If ROLLBACK is executed in the
#101221 20:48:00 server id 1071 end_log_pos 107 Start: binlog middle of the transaction,
v 4, server v 5.5.8-log created 101221 20:48:00 database will be inconsistent
ROLLBACK;
BINLOG '
8JMQTQ8vBAAAZwAAAGsAAAAAAAQANS41LjgtbG9nAAAAAAAAA  Trimming these ROLLBACK
AAAAAAAAAAAAAAAAAAAAAAAAAAA queries/events from mysqlbinlog
AAAAAAAAAAAAAAAAAAAAAAAAEzgNAAgAEgAEBAQEEgAAVAA outputs is needed
EGggAAAAICAgCAA== '/*!*/;
# at 102213
#110207 15:43:42 server id 1384 end_log_pos 102211 …  Do not trim necessary rollback
update …………………..
statements (i.e. BEGIN; updating
# at 102357 non-trans table, updating trans
#110207 15:43:42 server id 1384 end_log_pos 102238 table, ROLLBACK)
Xid = 12951490691
COMMIT/*!*/;
ROLLBACK; /* added by mysqlbinlog */

Purging relay logs
 By default, when SQL thread has read and executed the whole
relay log file, SQL thread automatically removes it.
 Because it is not needed by the SQL thread anymore
 But for recovering other slaves, the old relay logs might be needed

 SET GLOBAL relay_log_purge=0, and set it in my.cnf
 Side effect:
 Relay log files will sooner or later occupy the whole disk space
– No similar features like expire_logs_days for binary logs
– Schedule the following batch job will help
* SET GLOBAL relay_log_purge=1;
* FLUSH LOGS;
* Waiting for a while so that SQL thread switches the log file (old logs are removed)
* SET GLOBAL relay_log_purge=0;
 When SQL thread reaches the end of the relay log file and if
relay_log_purge equals to 1, the SQL thread removes all of the relay logs
it has executed so far
– No way to remove “all relay logs before yesterday”
– Invoking cron jobs at the same time on the all slaves will cause “no relay log
found for recovery” situation

Tips: Removing lots of large files

 Another serious side effect:
 SQL thread removes all relay log files when it reaches the end of the
relay log
 When you set relay_log_purge=1 per day, the total relay log file size
might reach 10GB or (much) more
 Dropping lots of large files take very long time on ext3
 SQL thread stops until removing all relay logs
– Might take 90 seconds to drop 30*1GB files

 Solution: Creating hard links
 foreach (relay_logs)
– ln /path/to/relay_log /path/to/archive_dir/
 SET GLOBAL relay_log_purge=1; FLUSH LOGS; SET GLOBAL
relay_log_purge=0;
 rm –f /path/to/archive_dir/*

How to identify whether SQL thread
has executed all events
 You need wait until SQL thread has executed all events
 SELECT MASTER_POS_WAIT (<Master_Log_File>,<Read_Master_Log_Pos>)
may not work
 MASTER_POS_WAIT() blocks until the slave has read and applied all
updates up to the specified position in the master log.
 If only part of the transactions are sent to the slave, SQL thread will never
execute up to Read_Master_Log_Pos.

 Check SHOW PROCESSLIST outputs
 If there is a thread of “system user” that has “^Has read all relay log; waiting for the
slave I/O thread to update it” state, the SQL thread has executed all events.

mysql> show processlist¥G
Id: 14
User: system user
Host:
db: NULL
Command: Connect
Time: 5769
State: Has read all relay log; waiting for the slave I/O thread
to update it
Info: NULL

Malicious queries
 Some malicious queries might cause recovery problems
 insert into t1 values(0,0,"ROLLBACK);
# at 15465
#110204 17:02:33 server id 1306 end_log_pos 1662 Query thread_id=30069 exec_time=0
error_code=0
ROLLBACK");
 Problems happen if end_log_pos value matches the target position
 Use mysqlbinlog --base64-output=always to identify starting position
– Query events are converted to row format. Base64 row format never contains
malicious strings
– Supported in mysqlbinlog from MySQL 5.1 or higher, but can work with
MySQL 5.0 server, too
– After identifying starting relay log position, generate events by normal
mysqlbinlog arguments (printing query events don’t cause problems here)
 5.1 mysqlbinlog can read 5.0/5.1 binlog format
 Use --base64-output=never for 5.0 mysqld to suppress printing BINLOG
events

Parallel Recovery

 In some cases many (10 or more) slaves are deployed

 Each slave can be recovered in parallel

 Relay logs are deleted once the slave executes
CHANGE MASTER.
 You must not execute CHANGE MASTER on the
latest slave until you generate diff relay logs for all
the rest slaves

Recovery procedure
Manager Dead Master Latest Slave newM Slaves

1. Saving Master Binlog Phase
Generate binlog

2. Diff Log Generation on the New Master Phase
Generate diff relay log

3. Master Log Apply Phase

Generate non-executed relay logs
Apply all logs

4. Parallel Slave Diff Log Generation Phase
Generate diff relay log

5. Parallel Slave Log Apply Phase
Generate non-executed relay logs
Apply all logs
Change Master, Start Slave

Row based format
# at 2642668
# at 2642713
#110411 16:14:00 server id 1306 end_log_pos 2642713 Table_map: `db1`.`t1`
mapped to number 16
#110411 16:14:00 server id 1306 end_log_pos 2642764 Write_rows: table id 16
flags: STMT_END_F

BINLOG '
OKqiTRMaBQAALQAAABlTKAAAABAAAAAAAAEABWdhbWVfAAJ0MQADAwP8
AQIG
OKqiTRcaBQAAMwAAAExTKAAAABAAAAAAAAEAA//4CmgAAApoAAALAGFhY
WFhYTI2NjM0
'/*!*/;
 Multiple “#at” entries + same number of “end_log_pos” entries (when parsed by
mysqlbinlog)
 “Table_map” event + “Write_rows (or others)” event + STMT_END
 Write_rows events can be many when using LOAD DATA, Bulk INSERT, etc
 mysqlbinlog prints out when valid “Table Map .. STMT End” events are written
 If slave A has only partial events, it is needed to send complete “Table Map ..
STMT End” events from the latest slave

Automating failover
 Common HA tasks
 Detecting master failure
 Node Fencing (Power off the dead master, to avoid split brain)
 Updating writer IP address

 Writing a script to do failover, based on what I have covered so far

 Running master failover scripts automatically
 Make sure not to stop by stupid errors
– Creating working/logging directory if not exists
– Check SSH public key authentication and MySQL privileges at the beginning of starting
the monitoring script
 Decide failover criteria
– Not starting failover if one or more slave servers are not alive (or SQL thread can’t be
started)
– Not starting failover if the last failover has happened recently (within 8 hours)

 Notification/Operation
 Sending mails
 Disabling scheduled backup jobs on the new master
 Updating internal administration tool status, master/slave ip address mappings, etc

Tool: Master High Availability Toolkit

Manager master

MySQL-MasterHA-Manager
- master_monitor
- master_switch
- masterha_manager slave1 slave2 slave3
master MySQL-MasterHA-Node
- save_binary_logs
- apply_diff_relay_logs
- filter_mysqlbinlog
- purge_relay_logs
 Manager
 master_monitor: Detecting master failure
 master_switch: Doing failover (manual, or automatic failover invoked by masterha_manager)
 Node : Deploying on all MySQL servers
 save_binary_logs: Copying master’s binary logs if accessible
 apply_diff_relay_logs: Generating differential relay logs from the latest slave, and applying all
differential binlog events
 filter_mysqlbinlog: Trimming unnecessary ROLLBACK events
 purge_relay_logs: Deleting relay logs without stopping SQL thread
 We have started using this tool internally. Will publish as OSS soon

One Manager per Datacenter
 Each Manager monitors multiple
MySQL masters within the same
datacenter

 If managers at DC2 and DC3 are
DC1 DC2 reachable from the manager at
master master DC1, and if a master is not
Mgr Mgr master reachable from none of the
master
managers, the master failover
master master
procedure starts
 Main purpose is to avoid split
Mgr
brain
master master
master  If any catastrophic failure
DC3 (datacenter crash) happens, we do
manual failover

Case
 Kernel panic happened on the master

 Checking whether the master is really dead (10 sec)
 Checking SSH reachability (to check saving binlog is possible or not)
 Check connectivity through other datacenters (secondary networks)

 STONITH (Forcing power off)
 To make sure the master is really not active
 Power off time highly depends on H/W
– Dell PowerEdge R610: 5-10 seconds (via telnet+DRAC)
– HP DL360: 4-5 seconds (via ipmitool+iLO)

 Master recovery
 Finished in less than 1 second

 Parallel slave recovery
 Finished in less than 1 second

Current limitations & tips
 Three or more-tier replication is not supported (i.e. Master->Master2->Slave)
 Check Global Transaction ID project
– Tracing differential relay log events becomes much easier
– Binlog format needs to be changed (It doesn’t work with -5.5)

 LOAD DATA [LOCAL] INFILE with SBR is not supported
 It’s deprecated actually, and it causes significant replication delay.
 SET sql_log_bin=0; LOAD DATA … ; SET sql_log_bin=1; is recommended
approach

 Replication filtering rules (binlog-do-db, replicate-ignore-db, etc) must be
same on all MySQL servers

 Do not use MySQL 5.0.45 or lower version
 end_log_pos is incorrect (not absolute): http://bugs.mysql.com/bug.php?id=22540
– I did a bit hack to make the tool work with 5.0.45 since we still have some legacy
servers, but generally upgrades should be done
 When replication network failure happens, a bogus byte stream might be sent to
slaves, which will stop SQL threads: http://bugs.mysql.com/bug.php?id=26489

Table of contents

 Automating master failover

 Minimizing downtime at master maintenance

Minimizing downtime at master maintenance

 Operations that need switching master
 Upgrading MySQL
 Replacing H/W components (increasing RAM, etc)

 Operations that do NOT need switching master
 Adding/dropping index
 Adding columns
 oak-online-alter-table or Facebook OSC may help

 Adding/Changing shards
 Can be done without stopping service, if designed well
 Hash based sharding makes it difficult to re-shard without stopping
services
 Mapping table based sharding makes it much easier

Tentative three-tier replication
Writer App Writer App

master slave1 Orig master New master

slave2 slave3 slave2 slave3

 Applications gradually establish database connections to the
new master (or just moving writer IP address, if you can
accept burst errors)
 Writes on the orig master will be finally replicated to the new
master
 Destroying orig master when orig master has sent all binlog
events

Tentative three-tier replication (2)

master slave1 Orig master New master

slave2 slave3 slave2 slave3
 Cons: Consistency problems might happen
 AUTO_INCREMENT doesn’t work (ID conflict), unless carefully using
auto_increment_increment and auto_increment_offset
 When the current master is updated, the row on the slave 1 is not locked
 “#1. Updating cur master set value=500 where id=1, #2. Updating slave 1 set value=1000
where id=1, #3. Replicating #1” -> #2 is lost
 Works well for INSERT-only, non-auto-inc query patterns
 Other possible approaches
 Using Spider + VP storage engine on the orig master
– Synchronous updates to the new master
– Replicatoin channel must be disconnected between orig master and new master

Promoting one of slaves
Activating write IP
after slave1 promotes
master
slave1-> New Master

slave1 slave2 slave3 slave2: slave3:
CHANGE MASTER CHANGE MASTER

 Cons: A few seconds of write downtime happens
 Until slave 1 is activated as a new master
 Master switch should be done as quickly as possible
– Otherwise applications can not execute updates for a long time

Is a few seconds of downtime acceptable?

 In some cases it is acceptable for a few seconds of
downtime on master
 500+ connections per second regularly
 100 connections at 3am
 2 seconds downtime will make 200 connections get
tentative errors
– Pushing reload button will be fine

Graceful master switch
 FLUSH TABLES WITH READ LOCK is not a silver bullet
 Does not return errors immediately
 Applications are kept waiting in orig master forever, unless read_timeout is set
 Response time and number of connections are highly increased

 Updating multiple mysql instances (multiple shards) is not uncommon
 “COMMIT Successful on node 1 -> COMMIT failure on node 2” results in data
inconsistency
 At least transaction commit should not be aborted

 More graceful approach
 Rejecting new database connections (DROP USER app_user)
 Waiting for 1-2 seconds so that almost all database connections are disconnected
 Rejecting all updates except SUPER by SET GLOBAL read_only=1;
 Waiting for .N second
 Rejecting all updates by FLUSH TABLES WITH READ LOCK

Part of Master High Availability Toolkit
 “Fast master switch” functionality is included, mostly
based on master failover tool
 master_switch --master_state=alive
 Master switch in 2 seconds (2 seconds for graceful writer
blocks)

 Differences from master failover are:
 Not automatic (interactive)
 All servers must be alive
 Replication delay must be zero
 Freezing updates on the current master is needed
 No power off
 No binlog/relay log recovery (Just using
MASTER_POS_WAIT() is fine)

Conclusion

 Automating master failover is possible
 Without introducing too much complexity on application
side
 With 5.0/5.1 InnoDB
 Works perfectly with Semi Synchronous Replication

 Our tool will soon be released as an open source
software

Automated master failover

More Related Content

Automated master failover