MySQL+Keepalived主从高可用集群原理+实战
目录:
MySQL+Keepalived主从高可用集群原理+实战
背景
随着公司业务的发展,对数据库的要求越来越高,以前的单机MySQL肯定是玩不转了,亟需升级成MySQL集群,这是一个公司在业务发展时不得不面临的问题
就单机MySQL而言,自己玩玩可以,运用到实际项目中,那肯定要挨批的。一方面数据不安全,万一数据库的电脑磁盘坏了,就坑了。另一方面数据库的并发能力是有限的,一般并发数200~500就差不多了,当然你要继续往上增,也是可以的,那就影响整体Mysql的响应时间。
那么有哪些集群方案呢?
集群方案
集群方案之前已经讨论过了,可以参考之前的【MySQL主从集群原理+实战】,今天我们来实现这种Keepalived+双主热备级联复制集群
优点:
读写分离,增加整体性能
Master有故障转移(采用Keepalived实现),当Master宕机时,备用Master自动顶上去,且IP无变化(VIP)
Master和备用Master之间采用半同步机制,最大程度上保证数据一致性
缺点:
Master故障转移时,Slave需要用脚本手动切换,增加复杂度
数据存在一致性问题,因为异步,所以Slave的数据一定不是最新的,需要等待一个时间窗后才能读取
Slave过多时Slave对Master的负载以及网络带宽都会成为一个严重的问题
Keepalived双主热备级联复制集群
m1和m2作为双主,互为主从且同步方式为半同步,为什么呢?因为这样在Master主机宕机恢复,对Master备机写入的数据也能重新同步到Master主,此时的主备角色正好颠倒过来,所以要设置互为主从且半同步。
m3作为级联复制的中间节点,新增加m4和m5作为新的Slave,正常情况下架构如下
当发生故障时,VIP自动漂移到Master备机上面,且通过Keepalived的状态脚本自动切换Slave,即m3到Master备机上,当Master主机(m1)恢复时,VIP又会自动漂移到m1上并且m3也会自动切换到m1上!
Keepalived状态切换脚本需要ssh和mysql免密登录的支持,其原理就是m1或者m2检测到自己成为Master角色时,通过ssh登录到m3上执行MySQL的切换命令
环境准备
docker run -dit --name m1 --privileged mysql5.7.24 /usr/sbin/init
docker run -dit --name m2 --privileged mysql5.7.24 /usr/sbin/init
docker run -dit --name m3 --privileged mysql5.7.24 /usr/sbin/init
docker run -dit --name m4 --privileged mysql5.7.24 /usr/sbin/init
docker run -dit --name m5 --privileged mysql5.7.24 /usr/sbin/init
各个IP如下:
VIP 172.17.0.101
m1 172.17.0.10
m2 172.17.0.11
m3 172.17.0.15
m4 172.17.0.16
m5 172.17.0.17
MYSQL_HOME变量定义,每台都定义下,后面的命令都依赖MYSQL_HOME
# 定义变量
MYSQL_VERSION=5.7.24
MYSQL_HOME=/usr/local/mysql-$MYSQL_VERSION
需要注意的是每台上要重置一下server_uuid,否则后面的主从复制会出错,直接删掉auto.cnf让MySQL自动生成即可
rm -rf $MYSQL_HOME/data/auto.cnf
由于是在docker里面,可能没有ssh,每台都装上ssh服务,后面的从机自动切换需要用到
yum install -y openssh-server openssh-clients passwd
双主配置
重置配置
如果哪个步骤出问题了,需要重新来,可以通过下面的方法重置
删除所有的库,分开执行
mysql -uroot -p -e "show databases" | grep -vE 'mysql|information_schema|performance_schema|test|sys|Database' | awk '{ print "drop database "$1";" }' > drop.sql && cat drop.sql
mysql -uroot -p < drop.sql
重置Slave
STOP SLAVE;
RESET SLAVE;
重置Master,binlog将会重新开始
STOP MASTER;
Master主机
# 先用-15优雅杀掉进程
ps aux | grep -v grep | grep mysql | awk '{ print $2 }' | xargs kill -15
sleep 1
# 修改配置
cat > /etc/my.cnf <<EOF
[client]
user=root
password=root
[mysqld]
user=mysql
character-set-server=utf8mb4
collation-server=utf8mb4_general_ci
datadir=$MYSQL_HOME/data
log-bin=mysql-bin-master
expire_logs_days=5
server-id=1
log-slave-updates=1
plugin-load = "rpl_semi_sync_master=semisync_master.so;rpl_semi_sync_slave=semisync_slave.so"
rpl-semi-sync-master-enabled=1
rpl-semi-sync-slave-enabled=1
[mysqld_safe]
log-error=$MYSQL_HOME/logs/mysqld.log
pid-file=$MYSQL_HOME/pids/mysqld.pid
[client]
default-character-set=utf8mb4
EOF
# 启动
/etc/init.d/mysqld start
# 查看启动日志
tail -f $MYSQL_HOME/logs/mysqld.log
log-slave-updates=1:这个参数需要注意下,互为主从的时候两台都需要写入自己从对方取到的数据到binlog里面
plugin-load = "rpl_semi_sync_master=semisync_master.so;rpl_semi_sync_slave=semisync_slave.so" rpl-semi-sync-master-enabled=1 rpl-semi-sync-slave-enabled=1主从半同步都需开启,因为是互为主从
[client]:需要注意下,配置了可以免密执行SQL,方便后面实现自动切换从机脚本!
执行完毕后登陆MySQL创建同步用户
DROP USER 'repl'@'%' ;
CREATE USER 'repl'@'%' IDENTIFIED BY 'repl';
GRANT REPLICATION SLAVE ON *.* TO 'repl'@'%';
flush privileges;
看一下Master状态
SHOW MASTER STATUS;
+-------------------------+----------+--------------+------------------+-------------------+
| File | Position | Binlog_Do_DB | Binlog_Ignore_DB | Executed_Gtid_Set |
+-------------------------+----------+--------------+------------------+-------------------+
| mysql-bin-master.000001 | 154 | | | |
+-------------------------+----------+--------------+------------------+-------------------+
Master备机
# 先用-15优雅杀掉进程
ps aux | grep -v grep | grep mysql | awk '{ print $2 }' | xargs kill -15
sleep 1
# 修改配置
cat > /etc/my.cnf <<EOF
[client]
user=root
password=root
[mysqld]
user=mysql
character-set-server=utf8mb4
collation-server=utf8mb4_general_ci
datadir=$MYSQL_HOME/data
log-bin=mysql-bin-master
expire_logs_days=5
server-id=2
log-slave-updates=1
plugin-load = "rpl_semi_sync_master=semisync_master.so;rpl_semi_sync_slave=semisync_slave.so"
rpl-semi-sync-master-enabled=1
rpl-semi-sync-slave-enabled=1
[mysqld_safe]
log-error=$MYSQL_HOME/logs/mysqld.log
pid-file=$MYSQL_HOME/pids/mysqld.pid
[client]
default-character-set=utf8mb4
EOF
# 启动
/etc/init.d/mysqld start
# 查看启动日志
tail -f $MYSQL_HOME/logs/mysqld.log
Master备机和主机如出一辙,主要区别的server-id字段不能一样,集群中的每台机器server-id都不能一样!
接下来就是创建同步用户,最好和Master主机创建成一样的,如果不一样,后面自动切换Slave的脚本就要针对Master主机和Master备机做调整,比较麻烦!
DROP USER 'repl'@'%' ;
CREATE USER 'repl'@'%' IDENTIFIED BY 'repl';
GRANT REPLICATION SLAVE ON *.* TO 'repl'@'%';
flush privileges;
看一下Master状态
SHOW MASTER STATUS;
+-------------------------+----------+--------------+------------------+-------------------+
| File | Position | Binlog_Do_DB | Binlog_Ignore_DB | Executed_Gtid_Set |
+-------------------------+----------+--------------+------------------+-------------------+
| mysql-bin-master.000001 | 154 | | | |
+-------------------------+----------+--------------+------------------+-------------------+
可以看到,两台干净的MySQL实例创建出来的Master File和Position都是一致的,两台Master角色都配置好了,接下来就是相互挂载成为对方的Slave
互为主从
其实很简单,互相指定下即可,将Master主挂载成为Master备机的从机
STOP SLAVE;
RESET SLAVE;
CHANGE MASTER TO
MASTER_HOST='172.17.0.10',
MASTER_USER='repl',
MASTER_PASSWORD='repl',
MASTER_LOG_FILE='mysql-bin-master.000001',
MASTER_LOG_POS=154;
START SLAVE;
SHOW SLAVE STATUS\G;
将Master备机挂载成为Master主机的从机,改下IP即可,其他的地方我们都配置成了一样的
STOP SLAVE;
RESET SLAVE;
CHANGE MASTER TO
MASTER_HOST='172.17.0.11',
MASTER_USER='repl',
MASTER_PASSWORD='repl',
MASTER_LOG_FILE='mysql-bin-master.000001',
MASTER_LOG_POS=154;
START SLAVE;
SHOW SLAVE STATUS\G;
分别查看下是否挂载成功
mysql> SHOW SLAVE STATUS\G;
*************************** 1. row ***************************
...
Slave_IO_Running: Yes
Slave_SQL_Running: Yes
...
再分别看下半同步是否开启
show status like 'Rpl_semi_sync_master_status';
+-----------------------------+-------+
| Variable_name | Value |
+-----------------------------+-------+
| Rpl_semi_sync_master_status | ON |
+-----------------------------+-------+
show status like 'Rpl_semi_sync_slave_status';
+----------------------------+-------+
| Variable_name | Value |
+----------------------------+-------+
| Rpl_semi_sync_slave_status | ON |
+----------------------------+-------+
最后在自行测试下载Master主机上面写数据Master备机能否同步,反过来再测试下,这里限于篇幅就不做测试了,实际上我做了很多次测试都是可以的。
在某一台宕机后另外一台第一次插入会等10s左右,说明半同步复制也是开启的,第二次插入就会退化了异步复制了,并且在宕机机器恢复后会自动同步缺失的数据。
配置Keepalived
编译安装
# 定义路径、定义版本
KEEPALIVED_HOME=/usr/local/keepalived
KEEPALIVED_VERSION=2.1.5
mkdir -p $KEEPALIVED_HOME
O_PATH=`pwd`
# 安装依赖
yum install -y wget make cmake gcc gcc-c++ openssl-devel net-tools
# 下载并解压
cd $KEEPALIVED_HOME && \
wget https://www.keepalived.org/software/keepalived-$KEEPALIVED_VERSION.tar.gz && \
tar -zxf keepalived-$KEEPALIVED_VERSION.tar.gz
# 编译安装
cd keepalived-$KEEPALIVED_VERSION
./configure --prefix=$KEEPALIVED_HOME
make && make install
# 软连
ln -s $KEEPALIVED_HOME/sbin/keepalived /sbin/
# 拷贝配置文件,默认路径为/etc/keepalived
mkdir /etc/keepalived
\cp -rf $KEEPALIVED_HOME/etc/keepalived/keepalived.conf /etc/keepalived/
# 添加环境变量
echo "export KEEPALIVED_HOME=$KEEPALIVED_HOME" >> ~/.bashrc
. ~/.bashrc
cd $O_PATH
编译安装过程如上,就不赘述了
Master主机配置
# 备份一下配置文件
cp /etc/keepalived/keepalived.conf /etc/keepalived/keepalived.conf.bak
# 重新编辑配置文件
echo '' > /etc/keepalived/keepalived.conf && vi /etc/keepalived/keepalived.conf
Master主机上的配置文件如下:
! Configuration File for keepalived
global_defs {
router_id HA-M1
}
vrrp_script chk_mysql {
script /etc/keepalived/chk_mysql.sh #健康监测脚本路径
interval 2
fall 3
rise 2
}
vrrp_instance VI_MYSQL {
state MASTER
interface eth0 # 监听网卡
virtual_router_id 100 # 虚拟路由编号,同一实例可以一致,但是其权重一定不能一致
priority 100 # 权重,两个节点不能一样
advert_int 1
mcast_src_ip 172.17.0.10
authentication {
auth_type PASS
auth_pass 1111
}
virtual_ipaddress {
172.17.0.101 dev eth0 label eth0:vip # 虚拟IP地址
# 240C::1234/64 # 支持IPv6
}
notify /etc/keepalived/kpad_notify.sh # keep状态传入脚本,通过该脚本可得知当前keep运行状态
track_script {
chk_mysql # 健康检查配置
}
}
router_id:标识,两台需不一样
state MASTER:表明这台是Master角色
priority:优先级,MASTER角色比BACKUP高!
virtual_router_id:虚拟路由编号,两台需要一致
172.17.0.101 dev eth0 label eth0:vip:这就是我们配置的VIP:172.17.0.101
script /etc/keepalived/chk_mysql.sh:MySQL的检测脚本,定时去检测MySQL进程是否挂掉,如果挂掉,在脚本里面重启之,定时通过interval配置,上面配置的是2s检测一次
notify /etc/keepalived/kpad_notify.sh:当Keepalived状态发生变化时,会调用该脚本,这也是我们实现Slave(m3)自动挂载Master的关键!!
当Keepalived从BACKUP提升为MASTER时,那么VIP就会漂移到自己身上,这个时候我们只需要将Slave挂载在自己身上即可
Slave的自动挂载Master,我们只需要实现上面这句话的逻辑即可!
下面是两个脚本文件,Master主机和Master备机都是一样的。分别创建脚本文件
MySQL状态监测脚本
在Master主机和Master备机上都要创建该脚本
touch /etc/keepalived/chk_mysql.sh
chmod +x /etc/keepalived/chk_mysql.sh
vi /etc/keepalived/chk_mysql.sh
# 创建个日志文件
mkdir -p /var/log/keepalived && touch /var/log/keepalived/chk_mysql.log
chk_mysql.sh内容如下,可以看到可以根据当前Keepalived的状态分别处理,但是此处的处理都是启动MySQL(如果MySQL进程已退出)
#!/bin/bash
STATE=`cat /tmp/keepalived-state`
log_file="/var/log/keepalived/chk_mysql.log"
service_name="mysqld"
service_cmd="/etc/init.d/mysqld"
get_pid=`pidof $service_name`
log() {
echo "$(date +"%Y-%m-%d %H:%M:%S.%4N") [$STATE] $1" >> $log_file
}
case $STATE in
"MASTER")
if [ "${get_pid}" == "" ]; then
log "$service_name service isn't exist."
log "Try to restart $service_name service."
$service_cmd start
if [ $? -eq 0 ]; then
log "restart $service_name service successfully."
else
log "restart $service_name service failed."
exit 1
fi
fi
exit 0
;;
"BACKUP")
if [ "${get_pid}" == "" ]; then
log "$service_name service isn't exist."
log "Try to restart $service_name service."
$service_cmd start
if [ $? -eq 0 ]; then
log "restart $service_name service successfully."
else
log "restart $service_name service failed."
exit 1
fi
fi
exit 0
;;
"FAULT")
exit 0
;;
*)
exit 1
;;
esac
Keepalived状态变更脚本
在Master主机和Master备机上都要创建该脚本
touch /etc/keepalived/kpad_notify.sh
chmod +x /etc/keepalived/kpad_notify.sh
vi /etc/keepalived/kpad_notify.sh
# 创建个日志文件
mkdir -p /var/log/keepalived && touch /var/log/keepalived/kpad_notify.log
kpad_notify.sh内容如下,
#!/bin/bash
TYPE=$1
NAME=$2
STATE=$3
log_file="/var/log/keepalived/kpad_notify.log"
log() {
echo "$(date +"%Y-%m-%d %H:%M:%S.%4N") [$STATE] $1" >> $log_file
}
case $STATE in
"MASTER")
echo 'MASTER' > /tmp/keepalived-state
# Slave自动切换Master
/etc/keepalived/chg_master.sh >> $log_file
log "changed master status $STATE"
exit 0
;;
"BACKUP")
echo 'BACKUP' > /tmp/keepalived-state
log "do nothing $STATE"
exit 0
;;
"FAULT")
echo 'FAULT' > /tmp/keepalived-state
log "keepalived status is fault."
exit 0
;;
*)
log "unknown keepalived status. $STATE"
exit 1
;;
esac
Slave自动切换Master脚本
当Keepalived从BACKUP提升为MASTER时,那么VIP就会漂移到自己身上,这个时候我们只需要将Slave挂载在自己身上即可
touch /etc/keepalived/chg_master.sh
vi /etc/keepalived/chg_master.sh
这句话的实现如下:chg_master.sh
#!/bin/bash
echo '====================================start master change==============================================='
# 目标IP,指的是Slave的IP,可以是多个
target_ip=(172.17.0.15)
# 本机IP
local_ip=`ifconfig eth0 | sed -n '2p' | awk '{ print $2 }'`
# 获取本机master信息,mysql免密登录
status=`mysql -uroot -e "show master status" | sed -n '2p'`
# binlog文件名字
binlog_file=`echo $status | awk '{ print $1 }'`
# binlog同步偏移量
binlog_pos=`echo $status | awk '{ print $2 }'`
echo "Got local ip is $local_ip and binary log is $binlog_file and binary log position is $binlog_pos"
# 循环执行转移Slave到自己身上
for t_ip in $target_ip
do
echo "start change slave for $t_ip to $local_ip"
# ssh远程到slave上执行切换master操作,需要ssh免密登录
ssh 172.17.0.15 <<SOF
mysql -uroot <<EOF
STOP SLAVE;
RESET SLAVE;
CHANGE MASTER TO
MASTER_HOST='$local_ip',
MASTER_USER='repl',
MASTER_PASSWORD='repl',
MASTER_LOG_FILE='$binlog_file',
MASTER_LOG_POS=$binlog_pos;
START SLAVE;
EOF
sleep 0.1
mysql -uroot -e "SHOW SLAVE STATUS\G";
SOF
done
echo '====================================end master change==============================================='
这个脚本用到了mysql -uroot免密登录和ssh免密登录,所以这里需要配置一下
mysql免密登录,在Master主机和Master备机上已经配置过了,这里无需配置,另外Slave(m3)上面也需要配置,这里讲到m3在说
ssh免密登录,上面已经安装过ssh服务端/客户端了,现在需要在Slave(m3)上配置ssh让Master主机/Master备机能登录上来操作
m3上执行下面两句话
# 启动ssh服务
systemctl start sshd
# 修改root密码,随便改个即可,后面要用到
passwd root
重点来了,需要在Master主机和Master备机上实现免密登录到m3,在Master主机和Master备机行执行下面操作
# 生成密钥
ssh-keygen -t rsa -f ~/.ssh/id_rsa
# 拷贝到目标主机,即m3
ssh-copy-id -p 22 172.17.0.15
# 试下下面的命令是否能直接登录,如果能说明免密登录成功
ssh 172.17.0.15
Master备机配置
# 备份一下配置文件
cp /etc/keepalived/keepalived.conf /etc/keepalived/keepalived.conf.bak
# 重新编辑配置文件
echo '' > /etc/keepalived/keepalived.conf && vi /etc/keepalived/keepalived.conf
Master备机上的配置文件如下:
! Configuration File for keepalived
global_defs {
router_id HA-M2
}
vrrp_script chk_mysql {
script /etc/keepalived/chk_mysql.sh #健康监测脚本路径
interval 2
fall 3
rise 2
}
vrrp_instance VI_MYSQL {
state BACKUP
interface eth0 # 监听网卡
virtual_router_id 100 # 虚拟路由编号,同一实例可以一致,但是其权重一定不能一致
nopreempt
priority 50 # 权重,两个节点不能一样
advert_int 1
mcast_src_ip 172.17.0.11
authentication {
auth_type PASS
auth_pass 1111
}
virtual_ipaddress {
172.17.0.101 dev eth0 label eth0:vip # 虚拟IP地址
# 240C::1234/64 # 支持IPv6
}
notify /etc/keepalived/kpad_notify.sh # keep状态传入脚本,通过该脚本可得知当前keep运行状态
track_script {
chk_mysql # 健康检查配置
}
}
注意router_id需要不一样,state BACKUP表明是备机,virtual_router_id需要一致,priority要比MASTER低!
启动Keepalived
最后在Master主机和Master备机上分别启动Keepalived
$KEEPALIVED_HOME/sbin/keepalived
不出意外的话在Master主机上执行ifconfig将会看到VIP,在Master备机上却看不到!
eth0:vip: flags=4163<UP,BROADCAST,RUNNING,MULTICAST> mtu 1500
inet 172.17.0.101 netmask 255.255.255.255 broadcast 0.0.0.0
ether 02:42:ac:11:00:0a txqueuelen 0 (Ethernet)
测试
测试的时候可以先把kpad_notify.sh中的
/etc/keepalived/chg_master.sh >> $log_file这句话注释掉再测试,因为现在我们还没有引入Slave(m3)
经过我的测试,有如下测试结论:
杀掉Master主机上的Keepalived和mysql:
VIP会漂移到Master备机上,ifconfig能看到VIP信息
通过/tmp/keepalived-state和日志kpad_notify.log观察Keepalived是从BACKUP状态转换为了MASTER状态
Master备机上做写操作,第一次写会发生半同步等待(10s左右),第二次写会退化成异步复制
启动Master上的Keepalived和mysql:
ifconfig看下VIP会漂移回来,并且Master备机上的VIP会消失
刚才在Master备机上写入的数据同步过来了
Keepalived的状态是从FAULT转变为BACKUP,最后变成MASTER
测试SQL
show databases;
create database tom;
use tom;
create table `user`(
`id` int comment 'id',
`name` varchar(50) comment '姓名'
);
insert into user values (1,'Tom'),(2,'Jerry');
select*from user;
insert into user values (3,'Tom3'),(4,'Jerry4');
配置Slave(m3)
接下来的配置都是一些常规的配置了,将m3配置成级联复制的中间节点,起到承上启下的作用,主要配置log-bin和log-slave-updates这两个参数!
# 先用-15优雅杀掉进程
ps aux | grep -v grep | grep mysql | awk '{ print $2 }' | xargs kill -15
sleep 1
# 修改配置
cat > /etc/my.cnf <<EOF
[client]
user=root
password=root
[mysqld]
user=mysql
character-set-server=utf8mb4
collation-server=utf8mb4_general_ci
datadir=$MYSQL_HOME/data
# 启用Binary Log
log-bin=mysql-bin
# server id唯一
server-id=3
# 重要!!!这个参数告知MySQL将本机收到的更新写入到自己的Binary Log中
log-slave-updates=1
[mysqld_safe]
log-error=$MYSQL_HOME/logs/mysqld.log
pid-file=$MYSQL_HOME/pids/mysqld.pid
[client]
default-character-set=utf8mb4
EOF
# 启动
/etc/init.d/mysqld start
# 查看启动日志
tail -f $MYSQL_HOME/logs/mysqld.log
异步复制配置
将Slave(m3)挂载到Master主机上(即m1上)
(注意,经过上面Master双主的测试后,可能binlog文件和偏移量已经变了,需要重置一下配置,请看开篇的重置配置)
STOP SLAVE;
RESET SLAVE;
CHANGE MASTER TO
MASTER_HOST='172.17.0.10',
MASTER_USER='repl',
MASTER_PASSWORD='repl',
MASTER_LOG_FILE='mysql-bin-master.000001',
MASTER_LOG_POS=154;
START SLAVE;
SHOW SLAVE STATUS\G;
创建同步用户
这里的同步用户是给m4和m5用的,没有什么特殊要求,简单起见Master主机/Master备机一致
DROP USER 'repl'@'%' ;
CREATE USER 'repl'@'%' IDENTIFIED BY 'repl';
GRANT REPLICATION SLAVE ON *.* TO 'repl'@'%';
flush privileges;
测试
这才是本篇中最重要的一个测试,一定要放开kpad_notify.sh中的
/etc/keepalived/chg_master.sh >> $log_file这句话,因为上面测试双主的时候关闭了
首先模拟Master主机宕机,杀掉Master主机上的Keepalived和mysql,查看Master备机上状态切换日志文件:/var/log/keepalived/kpad_notify.log,发现有如下输出
====================================start master change===============================================
Got local ip is 172.17.0.11 and binary log is mysql-bin-master.000001 and binary log position is 1598
start change slave for 172.17.0.15 to 172.17.0.11
*************************** 1. row ***************************
Master_Host: 172.17.0.11
Master_User: repl
Master_Port: 3306
Connect_Retry: 60
Master_Log_File: mysql-bin-master.000001
Read_Master_Log_Pos: 1598
Relay_Log_File: bbc1b4097f24-relay-bin.000002
Relay_Log_Pos: 327
Relay_Master_Log_File: mysql-bin-master.000001
Slave_IO_Running: Yes
Slave_SQL_Running: Yes
.....
====================================end master change===============================================
2021-04-10 06:15:12.2084 [MASTER] changed master status MASTER
这不就是我们Slave自动切换脚本输出的内容吗,说明自动切换生效了,通过Master_Host: 172.17.0.11就能看出来是Master备机的IP,在m3上执行SHOW SLAVE STATUS\G;看看是否和输出的一样,此时再在Master备机上进行写操作,Slave也是能同步过去的
然后模拟Master主机恢复,启动Master主机上的Keepalived和mysql,查看Master主机上状态切换日志文件:/var/log/keepalived/kpad_notify.log,发现有如下输出
====================================start master change===============================================
Got local ip is 172.17.0.10 and binary log is mysql-bin-master.000002 and binary log position is 412
start change slave for 172.17.0.15 to 172.17.0.10
*************************** 1. row ***************************
Slave_IO_State: Waiting for master to send event
Master_Host: 172.17.0.10
Master_User: repl
Master_Port: 3306
Connect_Retry: 60
Master_Log_File: mysql-bin-master.000002
Read_Master_Log_Pos: 412
Relay_Log_File: bbc1b4097f24-relay-bin.000002
Relay_Log_Pos: 327
Relay_Master_Log_File: mysql-bin-master.000002
Slave_IO_Running: Yes
Slave_SQL_Running: Yes
...
====================================end master change===============================================
2021-04-10 06:15:59.5336 [MASTER] changed master status MASTER
此时无需多言,聪明人一眼就能看出来Slave又自动切换回到了Master主机上,在m3上执行SHOW SLAVE STATUS\G;也可以看到和输出的一致
多个Slave?
多个Slave的情况也考虑到了,Slave自动切换Master脚本(chg_master.sh的目标IP:target_ip=(172.17.0.15)是个数组形式,多个Slave直接追加在后面,用空格隔开即可,当然Slave越多切换越慢,看实际情况而定!
级联复制
配置多个Slave
m4和m5作为m3的Slave,只需配置常规的异步复制即可,在m4和m5上分别执行,注意server-id集群中的实例必须唯一!server-id需自行更改
# 先用-15优雅杀掉进程
ps aux | grep -v grep | grep mysql | awk '{ print $2 }' | xargs kill -15
sleep 1
# 修改配置
cat > /etc/my.cnf <<EOF
[mysqld]
user=mysql
character-set-server=utf8mb4
collation-server=utf8mb4_general_ci
datadir=$MYSQL_HOME/data
server-id=4 # server id唯一
#server-id=5 # server id唯一
[mysqld_safe]
log-error=$MYSQL_HOME/logs/mysqld.log
pid-file=$MYSQL_HOME/pids/mysqld.pid
[client]
default-character-set=utf8mb4
EOF
# 启动
/etc/init.d/mysqld start
# 查看启动日志
tail -f $MYSQL_HOME/logs/mysqld.log
异步复制配置
注意:此时如果有数据的话需要将m1/m2/m3上的数据同步至m4和m5,总之要保证挂载之前m1到m5的数据是一致的!
先在m3上查看一下Master信息
show master status;
+------------------+----------+--------------+------------------+-------------------+
| File | Position | Binlog_Do_DB | Binlog_Ignore_DB | Executed_Gtid_Set |
+------------------+----------+--------------+------------------+-------------------+
| mysql-bin.000001 | 1185 | | | |
+------------------+----------+--------------+------------------+-------------------+
然后在m4和m5上执行下面命令分别挂载到m3上
STOP SLAVE;
RESET SLAVE;
CHANGE MASTER TO
MASTER_HOST='172.17.0.15',
MASTER_USER='repl',
MASTER_PASSWORD='repl',
MASTER_LOG_FILE='mysql-bin.000001',
MASTER_LOG_POS=1185;
START SLAVE;
SHOW SLAVE STATUS\G;
测试
这个测试基本没啥技术含量,在m1或者m2上任意进行写入操作,然后查看m1到m5的数据是否一致即可,当然经过上面的一番操作,数据肯定是一致的,但是级联复制链路越长,延迟就越大,这个需要结合实际情况。
总结
经过一顿操作,我们发现要实现MySQL的高可用整个过程还是比较复杂,特别是针对写节点,目前我们能实现的也是单点写,多点读,不过单点写有一台Master备机作为热备,那么有没有多点写呢?答案是有的,如果有时间,我们将来搭建一下PXC(Percona XtraDB Cluster)和MGR(MySQL Group Replication)
上面的架构中,Slave的切换需要通过脚本来实现,强烈依赖于keepalived的状态变化,那么我们能不能指定VIP作为m3的Master呢?答案是不能,因为指定Master除了指定IP还有binlog文件和偏移量,即使这三者都相同,也是不行的,每台MySQL都有一个server_uuid作为全局唯一标识,是不能简单的通过VIP欺骗MySQL Replication组件的(别问我怎么知道,因为我试过了不行最后才选择了脚本实现)
参考
https://www.cnblogs.com/kevingrace/p/6710136.html
http://ylong.net.cn/mysql-ha.html
https://blog.51cto.com/superleedo/1897966