redis集群实现(六) 容灾与宕机恢复

实现集群，一个重要的保证就是高可用性，要在各种软件和硬件的故障情况下仍然能够提供服务。一般来说有两种解决思路，一种是每一个节点互相之间都会进行数据交互以及监控，出现故障的时候，各个节点都可以做协调任务。另一种就是增加一个协调组件来对集群进行实时监控以及故障处理。现在使用比较广泛的是第二种方案，各个模块之间低耦合，工程师先也比较简单(相对第一种而言)。上一节已经介绍过了raft协议，有了raft协议的基础，相信大家理解sentinel也会比较轻松了。redis内的sentinel会实时扫描节点，如果发现了宕机的节点就会执行故障转移，选主等操作，我们来看一下具体的过程。
首先我们启动一个具有三个节点的sentinel集群，首先需要修改sentinel的配置文件，sentinel里有以下几个配置项需要修改的：
port：我们需要修改，因为要启动三个节点，端口必须是不一样的。
dir：sentinel的运行时目录。
sentinel monitor :监视一个名叫 的master，我们不需要监视slave，监视了master的话，slave会自动加入到sentinel里边。后边的quorum表示达成一致的最小数目，至少quorum台机器达成一致，才能保证一致性。
sentinel down-after-milliseconds 表示监视的节点在后没有回复就会被认为主观下线，当quorum个节点都认为此节点下线了以后就会被认为客观下线。
sentinel parallel-syncs 表示在故障转移的时候最多有numslaves在同步更新新的master。
我们修改过的三个sentinel.conf是sentinel1.conf，sentinel2.conf，sentinel3.conf，具体内容如下：
sentinel1.conf:

# Example sentinel.conf

# port 
# The port that this sentinel instance will run on
port 27000

# dir 
# Every long running process should have a well-defined working directory.
# For Redis Sentinel to chdir to /tmp at startup is the simplest thing
# for the process to don't interfere with administrative tasks such as
# unmounting filesystems.
dir /tmp

# sentinel monitor    
#
# Tells Sentinel to monitor this master, and to consider it in O_DOWN
# (Objectively Down) state only if at least  sentinels agree.
#
# Note that whatever is the ODOWN quorum, a Sentinel will require to
# be elected by the majority of the known Sentinels in order to
# start a failover, so no failover can be performed in minority.
#
# Slaves are auto-discovered, so you don't need to specify slaves in
# any way. Sentinel itself will rewrite this configuration file adding
# the slaves using additional configuration options.
# Also note that the configuration file is rewritten when a
# slave is promoted to master.
#
# Note: master name should not include special characters or spaces.
# The valid charset is A-z 0-9 and the three characters ".-_".
sentinel monitor master1 127.0.0.1 7000 2
sentinel monitor master2 127.0.0.1 7004 2
sentinel monitor master3 127.0.0.1 7005 2

# sentinel down-after-milliseconds  
#
# Number of milliseconds the master (or any attached slave or sentinel) should
# be unreachable (as in, not acceptable reply to PING, continuously, for the
# specified period) in order to consider it in S_DOWN state (Subjectively
# Down).
#
# Default is 30 seconds.
sentinel down-after-milliseconds master1 30000
sentinel down-after-milliseconds master2 30000
sentinel down-after-milliseconds master3 30000

# sentinel parallel-syncs  
#
# How many slaves we can reconfigure to point to the new slave simultaneously
# during the failover. Use a low number if you use the slaves to serve query
# to avoid that all the slaves will be unreachable at about the same
# time while performing the synchronization with the master.
sentinel parallel-syncs master1 1
sentinel parallel-syncs master2 1
sentinel parallel-syncs master3 1

# sentinel failover-timeout  
#
# Specifies the failover timeout in milliseconds. It is used in many ways:
#
# - The time needed to re-start a failover after a previous failover was
#   already tried against the same master by a given Sentinel, is two
#   times the failover timeout.
#
# - The time needed for a slave replicating to a wrong master according
#   to a Sentinel current configuration, to be forced to replicate
#   with the right master, is exactly the failover timeout (counting since
#   the moment a Sentinel detected the misconfiguration).
#
# - The time needed to cancel a failover that is already in progress but
#   did not produced any configuration change (SLAVEOF NO ONE yet not
#   acknowledged by the promoted slave).
#
# - The maximum time a failover in progress waits for all the slaves to be
#   reconfigured as slaves of the new master. However even after this time
#   the slaves will be reconfigured by the Sentinels anyway, but not with
#   the exact parallel-syncs progression as specified.
#
# Default is 3 minutes.
sentinel failover-timeout master1 180000
sentinel failover-timeout master2 180000
sentinel failover-timeout master3 180000

sentinel2.conf

# Example sentinel.conf

# port 
# The port that this sentinel instance will run on
port 27001

# dir 
# Every long running process should have a well-defined working directory.
# For Redis Sentinel to chdir to /tmp at startup is the simplest thing
# for the process to don't interfere with administrative tasks such as
# unmounting filesystems.
dir /tmp

# sentinel monitor    
#
# Tells Sentinel to monitor this master, and to consider it in O_DOWN
# (Objectively Down) state only if at least  sentinels agree.
#
# Note that whatever is the ODOWN quorum, a Sentinel will require to
# be elected by the majority of the known Sentinels in order to
# start a failover, so no failover can be performed in minority.
#
# Slaves are auto-discovered, so you don't need to specify slaves in
# any way. Sentinel itself will rewrite this configuration file adding
# the slaves using additional configuration options.
# Also note that the configuration file is rewritten when a
# slave is promoted to master.
#
# Note: master name should not include special characters or spaces.
# The valid charset is A-z 0-9 and the three characters ".-_".
sentinel monitor master1 127.0.0.1 7000 2
sentinel monitor master2 127.0.0.1 7004 2
sentinel monitor master3 127.0.0.1 7005 2

# sentinel down-after-milliseconds  
#
# Number of milliseconds the master (or any attached slave or sentinel) should
# be unreachable (as in, not acceptable reply to PING, continuously, for the
# specified period) in order to consider it in S_DOWN state (Subjectively
# Down).
#
# Default is 30 seconds.
sentinel down-after-milliseconds master1 30000
sentinel down-after-milliseconds master2 30000
sentinel down-after-milliseconds master3 30000

# sentinel parallel-syncs  
#
# How many slaves we can reconfigure to point to the new slave simultaneously
# during the failover. Use a low number if you use the slaves to serve query
# to avoid that all the slaves will be unreachable at about the same
# time while performing the synchronization with the master.
sentinel parallel-syncs master1 1
sentinel parallel-syncs master2 1
sentinel parallel-syncs master3 1

# sentinel failover-timeout  
#
# Specifies the failover timeout in milliseconds. It is used in many ways:
#
# - The time needed to re-start a failover after a previous failover was
#   already tried against the same master by a given Sentinel, is two
#   times the failover timeout.
#
# - The time needed for a slave replicating to a wrong master according
#   to a Sentinel current configuration, to be forced to replicate
#   with the right master, is exactly the failover timeout (counting since
#   the moment a Sentinel detected the misconfiguration).
#
# - The time needed to cancel a failover that is already in progress but
#   did not produced any configuration change (SLAVEOF NO ONE yet not
#   acknowledged by the promoted slave).
#
# - The maximum time a failover in progress waits for all the slaves to be
#   reconfigured as slaves of the new master. However even after this time
#   the slaves will be reconfigured by the Sentinels anyway, but not with
#   the exact parallel-syncs progression as specified.
#
# Default is 3 minutes.
sentinel failover-timeout master1 180000
sentinel failover-timeout master2 180000
sentinel failover-timeout master3 180000

sentinel3.conf

# Example sentinel.conf

# port 
# The port that this sentinel instance will run on
port 27002

# dir 
# Every long running process should have a well-defined working directory.
# For Redis Sentinel to chdir to /tmp at startup is the simplest thing
# for the process to don't interfere with administrative tasks such as
# unmounting filesystems.
dir /tmp

# sentinel monitor    
#
# Tells Sentinel to monitor this master, and to consider it in O_DOWN
# (Objectively Down) state only if at least  sentinels agree.
#
# Note that whatever is the ODOWN quorum, a Sentinel will require to
# be elected by the majority of the known Sentinels in order to
# start a failover, so no failover can be performed in minority.
#
# Slaves are auto-discovered, so you don't need to specify slaves in
# any way. Sentinel itself will rewrite this configuration file adding
# the slaves using additional configuration options.
# Also note that the configuration file is rewritten when a
# slave is promoted to master.
#
# Note: master name should not include special characters or spaces.
# The valid charset is A-z 0-9 and the three characters ".-_".
sentinel monitor master1 127.0.0.1 7000 2
sentinel monitor master2 127.0.0.1 7004 2
sentinel monitor master3 127.0.0.1 7005 2

# sentinel down-after-milliseconds  
#
# Number of milliseconds the master (or any attached slave or sentinel) should
# be unreachable (as in, not acceptable reply to PING, continuously, for the
# specified period) in order to consider it in S_DOWN state (Subjectively
# Down).
#
# Default is 30 seconds.
sentinel down-after-milliseconds master1 30000
sentinel down-after-milliseconds master2 30000
sentinel down-after-milliseconds master3 30000

# sentinel parallel-syncs  
#
# How many slaves we can reconfigure to point to the new slave simultaneously
# during the failover. Use a low number if you use the slaves to serve query
# to avoid that all the slaves will be unreachable at about the same
# time while performing the synchronization with the master.
sentinel parallel-syncs master1 1
sentinel parallel-syncs master2 1
sentinel parallel-syncs master3 1

# sentinel failover-timeout  
#
# Specifies the failover timeout in milliseconds. It is used in many ways:
#
# - The time needed to re-start a failover after a previous failover was
#   already tried against the same master by a given Sentinel, is two
#   times the failover timeout.
#
# - The time needed for a slave replicating to a wrong master according
#   to a Sentinel current configuration, to be forced to replicate
#   with the right master, is exactly the failover timeout (counting since
#   the moment a Sentinel detected the misconfiguration).
#
# - The time needed to cancel a failover that is already in progress but
#   did not produced any configuration change (SLAVEOF NO ONE yet not
#   acknowledged by the promoted slave).
#
# - The maximum time a failover in progress waits for all the slaves to be
#   reconfigured as slaves of the new master. However even after this time
#   the slaves will be reconfigured by the Sentinels anyway, but not with
#   the exact parallel-syncs progression as specified.
#
# Default is 3 minutes.
sentinel failover-timeout master1 180000
sentinel failover-timeout master2 180000
sentinel failover-timeout master3 180000

然后我们输入

	redis-sentinel sentinel1.conf
	redis-sentinel sentinel2.conf
	redis-sentinel sentinel3.conf

就可以建立好三个sentinel伪集群，我们会看到如下打印，说明三个master和sentinel都被识别了。

56161:X 04 Dec 09:23:09.855 # Sentinel runid is 4dd7b82766f7faac95c251235682e42079e0a701
56161:X 04 Dec 09:23:09.855 # +monitor master master0 192.168.39.153 7000 quorum 2
56161:X 04 Dec 09:23:09.855 # +monitor master master2 192.168.39.153 7005 quorum 2
56161:X 04 Dec 09:23:09.856 # +monitor master master1 192.168.39.153 7004 quorum 2
56161:X 04 Dec 09:23:10.842 * +slave slave 192.168.39.153:7003 192.168.39.153 7003 @ master0 192.168.39.153 7000
56161:X 04 Dec 09:23:10.842 * +slave slave 192.168.39.153:7002 192.168.39.153 7002 @ master2 192.168.39.153 7005
56161:X 04 Dec 09:23:10.843 * +slave slave 192.168.39.153:7001 192.168.39.153 7001 @ master1 192.168.39.153 7004
56161:X 04 Dec 09:23:19.505 * +sentinel sentinel 192.168.39.153:27001 192.168.39.153 27001 @ master0 192.168.39.153 7000
56161:X 04 Dec 09:23:19.506 * +sentinel sentinel 192.168.39.153:27001 192.168.39.153 27001 @ master2 192.168.39.153 7005
56161:X 04 Dec 09:23:19.508 * +sentinel sentinel 192.168.39.153:27001 192.168.39.153 27001 @ master1 192.168.39.153 7004
56161:X 04 Dec 09:23:25.240 * +sentinel sentinel 192.168.39.153:27002 192.168.39.153 27002 @ master1 192.168.39.153 7004
56161:X 04 Dec 09:23:25.241 * +sentinel sentinel 192.168.39.153:27002 192.168.39.153 27002 @ master2 192.168.39.153 7005
56161:X 04 Dec 09:23:25.242 * +sentinel sentinel 192.168.39.153:27002 192.168.39.153 27002 @ master0 192.168.39.153 7000

一般来说，我们把一个master下线了以后，集群就会变成不可用状态，但是现在有了sentinel了，一旦master下线就会立刻执行故障转移，就能够在很短的时间内恢复可用。
开始的时候有六个节点，三个master，三个slave，状态如下：

127.0.0.1:7000> cluster info
cluster_state:ok
cluster_slots_assigned:16384
cluster_slots_ok:16384
cluster_slots_pfail:0
cluster_slots_fail:0
cluster_known_nodes:6
cluster_size:3
cluster_current_epoch:12
cluster_my_epoch:10
cluster_stats_messages_sent:2424257
cluster_stats_messages_received:2423717
127.0.0.1:7000> cluster nodes
930daea84150b5fabd32a95592781b27ceab1b71 192.168.39.153:7001 slave 81c884ebfc919ad293f02d797aff1033025ac27e 0 1480817793875 9 connected
8a6707d5b9269b6260315b47f300c1ab599733b7 192.168.39.153:7005 master - 0 1480817794879 11 connected 10923-16383
bdb62bb6ffce71588961f513c74b0d5a1a7145ea 192.168.39.153:7002 slave 8a6707d5b9269b6260315b47f300c1ab599733b7 0 1480817793372 11 connected
81c884ebfc919ad293f02d797aff1033025ac27e 192.168.39.153:7004 master - 0 1480817794378 9 connected 5461-10922
099cfc6fbb785449a8bf5369a53d21a9e127fa42 192.168.39.153:7000 myself,master - 0 0 10 connected 0-5460
a8081e97862d9cf76c72d364f9a173187376f215 192.168.39.153:7003 slave 099cfc6fbb785449a8bf5369a53d21a9e127fa42 0 1480817792868 10 connected

我们手动发送int信号终止这个进程，发现redis-server:7004进程已经被我们杀死了。

ubuntu@ubuntu-virtual-machine:~/redis-3.0.0/src$ ps aux | grep redis
ubuntu     6067  0.0  0.4  33148  4080 ?        Ss   11月27   0:00 SCREEN -S redis
ubuntu     7192  0.0  0.8  42300  8392 ?        Ssl  11月27   7:22 redis-server *:7000 [cluster]          
ubuntu     7196  0.0  1.0  42300 10632 ?        Ssl  11月27   7:19 redis-server *:7001 [cluster]          
ubuntu     7200  0.0  1.0  42300 10504 ?        Ssl  11月27   7:21 redis-server *:7002 [cluster]          
ubuntu     7205  0.0  1.0  42300 10524 ?        Ssl  11月27   7:21 redis-server *:7003 [cluster]          
ubuntu     7218  0.0  0.8  42300  8556 ?        Ssl  11月27   7:21 redis-server *:7005 [cluster]          
ubuntu    56036  0.0  0.3  31128  3232 pts/6    S+   09:15   0:00 screen -r redis
ubuntu    56161  0.2  0.7  42304  7532 pts/25   Sl+  09:23   0:10 redis-sentinel *:27000 [sentinel]
ubuntu    56176  0.2  0.7  42304  7444 pts/26   Sl+  09:23   0:10 redis-sentinel *:27001 [sentinel]
ubuntu    56192  0.2  0.9  42304  9424 pts/27   Sl+  09:23   0:10 redis-sentinel *:27002 [sentinel]
ubuntu    56536  0.0  0.2  15944  2396 pts/12   R+   10:29   0:00 grep --color=auto redis
ubuntu@ubuntu-virtual-machine:~/redis-3.0.0/src$ redis-cli -p 7000 
127.0.0.1:7000> cluster info
cluster_state:ok
cluster_slots_assigned:16384
cluster_slots_ok:16384
cluster_slots_pfail:0
cluster_slots_fail:0
cluster_known_nodes:6
cluster_size:3
cluster_current_epoch:13
cluster_my_epoch:10
cluster_stats_messages_sent:2433366
cluster_stats_messages_received:2433005
127.0.0.1:7000> cluster nodes
930daea84150b5fabd32a95592781b27ceab1b71 192.168.39.153:7001 master - 0 1480818606296 13 connected 5461-10922
8a6707d5b9269b6260315b47f300c1ab599733b7 192.168.39.153:7005 master - 0 1480818606797 11 connected 10923-16383
bdb62bb6ffce71588961f513c74b0d5a1a7145ea 192.168.39.153:7002 slave 8a6707d5b9269b6260315b47f300c1ab599733b7 0 1480818608306 11 connected
81c884ebfc919ad293f02d797aff1033025ac27e 192.168.39.153:7004 master,fail - 1480818583889 1480818583084 9 disconnected
099cfc6fbb785449a8bf5369a53d21a9e127fa42 192.168.39.153:7000 myself,master - 0 0 10 connected 0-5460
a8081e97862d9cf76c72d364f9a173187376f215 192.168.39.153:7003 slave 099cfc6fbb785449a8bf5369a53d21a9e127fa42 0 1480818607301 10 connected

我们发现，在停止了一个master节点以后，集群在很短的时间内处理了故障转移，然后集群立刻恢复可用，原来的slave变成了master。可以看出来，sentinel成功发挥了故障处理的作用，在分布式的集群中，保证高可用性是很重要的一点，下节我们从源代码层次看看sentinel如何实现的故障转移。