虽然spark master挂掉的几率很低，不过还是被我遇到了一次。以前在spark standalone的文章中也介绍过standalone的ha,现在详细说下部署流程，其实也比较简单。

一.机器

zookeeper集群

zk1:2181
zk2:2181
zk3:2181

spark master

spark-m1
spark-m2

spark worker

若干

二.步骤

1.进入spark-m1
修改conf/spark-env.sh

vi spark-env.sh
export SPARK_MASTER_IP=spark-m1
export SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=ZOOKEEPER -Dspark.deploy.zookeeper.url=zk1:2181,zk2:2181,zk3:2181 -Dspark.deploy.zookeeper.dir=/spark"   

启动master和slaves

./sbin/start-master.sh
./sbin/start-slaves.sh

2.进入spark-m2

修改conf/spark-env.sh

vi spark-env.sh
export SPARK_MASTER_IP=spark-m2
export SPARK_DAEMON_JAVA_OPTS="-Dspark.deploy.recoveryMode=ZOOKEEPER -Dspark.deploy.zookeeper.url=zk1:2181,zk2:2181,zk3:2181 -Dspark.deploy.zookeeper.dir=/spark"   

启动master和slaves

./sbin/start-master.sh
./sbin/start-slaves.sh

三.检测

在spark-m1的web ui中可以看到状态

spark-m2中可以看到处于STANDBY状态

application提交时,master改为

--master spark://spark-m1:7077,spark-m2:7077

spark shell 测试

在spark-m1中启动spark Shell

spark-shell --master spark://spark-m1:7077,spark-m2:7077

连接后关闭spark-m1 master

./bin/stop-master.sh

发现spark-shell不会断开而是转到spark-m2的master上继续执行（该过程持续大概1分钟,woker会重新注册到spark-m2上）,同时spark-m2变为alive状态。

可以在spark-m2的master日志中看到：

15/08/17 14:45:35 INFO ZooKeeperLeaderElectionAgent: We have gained leadership
15/08/17 14:45:36 INFO Master: I have been elected leader! New state: RECOVERING
15/08/17 14:45:36 INFO Master: Trying to recover worker:...
15/08/17 14:45:36 INFO Master: Trying to recover worker: ...
15/08/17 14:45:36 INFO Master: Trying to recover worker: ...
......
15/08/17 14:45:36 INFO Master: Worker has been re-registered: worker-...
15/08/17 14:45:36 INFO Master: Worker has been re-registered: worker-...
15/08/17 14:45:36 INFO Master: Worker has been re-registered: worker-...
...
15/08/17 14:45:36 INFO Master: Recovery complete - resuming operations!


部署结束