MySQL replication
Usage
- slurm1 and slurm2 make use of master master replication to provide high availability.
Requirements
Semisynchronous replication requires
MySQL 5.5 or newer. SL6 ships with
MySQL 5.1.x as such you will have do download the newer RPMs from the
MySQL website.
These RPMs have been copied to the Phoenix repo and are available through yum. Note you will have to first install
MySQL-shared-compat package as otherwise you will receive an error about conflicts.
Setup
- slurm1 and slurm2 are both masters
- Both can be written to and changes will be replicated to the other node.
- slurm1 and slurm2 are both semisynchronous masters and slaves
- When a transaction is performed on either node the change is replicated to the other machine
Config
Install
Download version 5.5 or newer RPMs from
http://dev.mysql.com/downloads/mysql/ (version 5.6 is available in the phoenix repo)
Install compatibility libs
yum install MySQL-shared-compat
Install the client and server
yum install MySQL-client MySQL-server
</verba
%ICON{TODO}%tim>
The first thing we have to do is change the password. The installation of MySQL-server will create the following file with a random password to log into MySQL with
<verbatim>
/root/.mysql_secret
</verbatim>
Log into MySQL and change the password
<verbatim>
mysql -p
SET PASSWORD FOR 'root'@'localhost' = PASSWORD('newpwd');
</verbatim>
Open another terminal and ensure you are able to log in with the new password.
Now lets edit the config to define our replication. The below is from slurm2, on slurm1 the only values that would change are the server-id and the bind-address, these would be 2 and 148.187.64.42 respectively.
<verbatim>
vim /etc/my.cnf
[mysqld]
datadir=/var/lib/mysql
socket=/var/lib/mysql/mysql.sock
user=mysql
# Disabling symbolic-links is recommended to prevent assorted security risks
symbolic-links=0
# Master-Slave config
bind-address=148.187.64.43
server-id=1 # master
log_bin=/var/log/mysql/mysql-bin.log
binlog_do_db=slurm_acct_db # the slurmdbd database
binlog_ignore_db = mysql
binlog_ignore_db = test
log-bin=mysql-bin
[mysqld_safe]
log-error=/var/log/mysqld.log
pid-file=/var/run/mysqld/mysqld.pid
</verbatim>
Restart MySQL for the changes to take affect
<verba
%ICON{TODO}%tim>
service mysql restart
Master slave replication
Now we need to log in and create the database if it does not exist. Note I upgraded from
MySQL 5.1 to 5.6 after the slurm RPMs had been installed.
WARNING There is a
bug in 5.6 that breaks replication when a grant command is issued with incorrect syntax.
mysql -p
CREATE DATABASE slurm_acct_db;
# Note replace 'password' in the below command with the actual password
GRANT REPLICATION SLAVE ON *.* TO 'root'@'%' IDENTIFIED BY 'password';
FLUSH PRIVILEGES;
show master status;ver for the primary slurmctld if it ever fails. The backup controller should be hosted on a node different from the node hosting the primary slurmctld. However, both hosts should mount a c
+------------------+----------+---------------+------------------+-------------------+
| File | Position | Binlog_Do_DB | Binlog_Ignore_DB | Executed_Gtid_Set |
+------------------+----------+---------------+------------------+-------------------+
| mysql-bin.000018 | 120 | slurm_acct_db | mysql,test | |
+------------------+----------+---------------+------------------+-------------------+
1 row in set (0.00 sec)
Now we need to install the packages on the other server, edited the my.cnf and set the password as previously described.
Lets now set up this server as a slave, first create the database if it does not exist.
mysql -p
CREATE DATABASE slurm_acct_db;
Define the other machine as the master, again change 'password' to the actual password. Note the MASTER_LOG_FILE and MASTER_LOG_POS values are taken from the 'show master status;' output on the other node.
CHANGE MASTER TO MASTER_HOST='148.187.64.43', MASTER_USER='root', MASTER_PASSWORD='password', MASTER_LOG_FILE='mysql-bin.000018', MASTER_LOG_POS=120;
Now we have defined our master start the slave and ensure that the values are correct. Master_Host, Master_Log_File and Read_Master_Log_Pos are the entries of main interest.
START SLAVE;
show slave status \G
Before enabling the other node as a master and defining slave replication in the other direction lets test the current replication.
# On slurm2
use slurm_acct_db;
create table testmaster21 (mid int(11) auto_increment, PRIMARY KEY (mid)) Engine=MyISAM;
# On slurm1 check the table exists
use slurm_acct_db;
show tables;
Master master replication
If all is good define master slave replication in the other direction.
On slurm1 display the statussacctmgr ping
ver for the primary slurmctld if it ever fails. The backup controller should be hosted on a node different from the node hosting the primary slurmctld. However, both hosts should mount a c
show master status;
+------------------+----------+---------------+------------------+-------------------+
| File | Position | Binlog_Do_DB | Binlog_Ignore_DB | Executed_Gtid_Set |
+------------------+----------+---------------+------------------+-------------------+
| mysql-bin.000010 | 120 | slurm_acct_db | mysql,test | |
+------------------+----------+---------------+------------------+-------------------+
1 row in set (0.00 sec)
Grant replication permission on the slave
GRANT REPLICATION SLAVE ON *.* TO 'root'@'%' IDENTIFIED BY 'password';
FLUSH PRIVILEGES;
show master status;
On slurm2 define slurm1 as a master, again using the values from the other node for MASTER_LOG_FILE and MASTER_LOG_POS
CHANGE MASTER TO MASTER_HOST='148.187.64.42', MASTER_USER='root', MASTER_PASSWORD='password', MASTER_LOG_FILE='mysql-bin.000010', MASTER_LOG_POS=120;
Again start the slave and check the status
START SLAVE;
show slave status \G
Ensure replication is working.
# On slurm1
use slurm_acct_db;
DROP TABLE testmaster21;
# On slurm2 check the table no longer exists
use slurm_acct_db;
show tables;
Enabling semisynchronous replication
WARNING MySQL 5.6 does not work with slurmdb at time of writting (19/08/2013)
Semisynchronous replication is provided by plugins. As both slurm1 and slurm2 are masters and slaves to one another install both master and slave plugins
INSTALL PLUGIN rpl_semi_sync_master SONAME 'semisync_master.so';
INSTALL PLUGIN rpl_semi_sync_slave SONAME 'semisync_slave.so';
By default these plugins are not automatically enabled so enable them on both systems.
SET GLOBAL rpl_semi_sync_master_enabled = 1;
SET GLOBAL rpl_semi_sync_slave_enabled = 1;
The replication between master and slave will need to be restarted for the nodes to be recognised as semisynchronous slaves.
STOP SLAVE IO_THREAD; START SLAVE IO_THREAD;
To ensure the semisynchronous plugin is enabled across service restarts of mysql add the following lines to each severs under the [mysqld] section of /etc/my.cnf
vim /etc/my.cnf
rpl_semi_sync_master_enabled=1
rpl_semi_sync_master_timeout=1000 # 1 second
rpl_semi_sync_slave_enabled=1
Testing
TBA
Monitoring
There is a check_mysql plugin available that makes our lives easier here's how to implement it.
yum install nagios-plugins-mysql.x86_64 --enablerepo=epel
- We now need to create a unprivileged user to log into the database to perform these checks with.
#mysql shell
#log in as root
#change password to an actual password
GRANT SELECT, REPLICATION CLIENT ON *.* TO 'nagios'@'localhost' IDENTIFIED BY 'password';
- Then update the nrpe and services files in cfengine and we're good to go.. Below is an example, the -S flag checks the slave status.
#Change password to the real password
command[check_mysql]=/usr/libexec/nagios-plugins/check_mysql -u nagios -p password -S -d slurm_acct_db
HA testing
Tested killing
MySQL processes with -s 9 signal on primary DB host, sacctmgr commands continued to return correct info on both slurm1 and slurm2
Submitted a job, waited until it was running stopped
MySQL gracefully on primary DB host job finished ok, other squeue returned successfully throughout
Brought down public interface on slurm2 after job started to run, caused sacctmgr commands to fail. It appears the slurmdbd deamon is required to be reachable in order for fail over to occur.
- Killed slurmctrld (init script would not stop the service), started again and sacctmgr and sinfo returned, however job did not complete
Brought down public interface on slurm1 to test fail over of slurmctld: Note is appears
StateSaveLocation in the slurm.conf needs to refer to a shared filesystem location for this to work.
- squeue and sinfo ran but took a longer time to return
- when a job was submitted it did not complete
- Added shared location for the StateSaveLocation using the NAS mounted on /slurm. Jobs completed and I could submit jobs with the primary controller down but sinfo and sacctmgr were still slow to return.
Issues
- Upon initial install mysql failed to start due to the fact /var/log/mysql was not owned by the mysql user.
- If the slave status shows that it is having trouble connecting to the master ensure the firewall is not blocking communication on port 3306 and grant permissions have been given. If it is open the port and restart the slave (STOP SLAVE; START SLAVE;)
- The syntax slurmdb uses when initially creating the tables in MySQL is not valid in MySQL 5.6 (works in 5.1). As such we cannot take advantage of semi synchronous replication at this time. TBA bug report with SchedMD.
Further reading
--
GeorgeBrown - 2013-08-12