Spark on YARN两种运行模式介绍

问题导读

1.Spark在YARN中有几种模式？
2.Yarn Cluster模式，Driver程序在YARN中运行，应用的运行结果在什么地方可以查看？
3.由client向ResourceManager提交请求，并上传jar到HDFS上包含哪些步骤？
4.传递给app的参数应该通过什么来指定？
5.什么模式下最后将结果输出到terminal中？

Spark在YARN中有yarn-cluster和yarn-client两种运行模式:

1.Yarn Cluster

Spark Driver首选作为一个ApplicationMaster在Yarn集群中启动，客户端提交给ResourceManager的每一个job都会在集群的worker节点上分配一个唯一的ApplicationMaster,

由该ApplicationMaster管理全生命周期的应用。因为Driver程序在YARN中运行，所以事先不用启动Spark Master/Client，应用的运行结果不能再客户端显示(可以在history server中查看)

，所以最好将结果保存在HDFS而非stdout输出，客户端的终端显示的是作为YARN的job的简单运行状况。

by @Sandy Ryza

by 明风@taobao

从terminal的output中看到任务初始化更详细的四个步骤:

14/09/28 11:24:52 INFO RMProxy: Connecting to ResourceManager at hdp01/172.19.1.231:8032
14/09/28 11:24:52 INFO Client: Got Cluster metric info from ApplicationsManager (ASM), number of NodeManagers: 4
14/09/28 11:24:52 INFO Client: Queue info ... queueName: root.default, queueCurrentCapacity: 0.0, queueMaxCapacity: -1.0,
queueApplicationCount = 0, queueChildQueueCount = 0
14/09/28 11:24:52 INFO Client: Max mem capabililty of a single resource in this cluster 8192
14/09/28 11:24:53 INFO Client: Uploading file:/usr/lib/spark/examples/lib/spark-examples_2.10-1.0.0-cdh5.1.0.jar to hdfs://hdp01:8020/user/spark/.sparkStaging/application_1411874193696_0003/spark-examples_2.10-1.0.0-cdh5.1.0.jar
14/09/28 11:24:54 INFO Client: Uploading file:/usr/lib/spark/assembly/lib/spark-assembly-1.0.0-cdh5.1.0-hadoop2.3.0-cdh5.1.0.jar to hdfs://hdp01:8020/user/spark/.sparkStaging/application_1411874193696_0003/spark-assembly-1.0.0-cdh5.1.0-hadoop2.3.0-cdh5.1.0.jar
14/09/28 11:24:55 INFO Client: Setting up the launch environment
14/09/28 11:24:55 INFO Client: Setting up container launch context
14/09/28 11:24:55 INFO Client: Command for starting the Spark ApplicationMaster: List($JAVA_HOME/bin/java, -server, -Xmx512m, -Djava.io.tmpdir=$PWD/tmp, -Dspark.master="spark://hdp01:7077", -Dspark.app.name="org.apache.spark.examples.SparkPi", -Dspark.eventLog.enabled="true", -Dspark.eventLog.dir="/user/spark/applicationHistory", -Dlog4j.configuration=log4j-spark-container.properties, org.apache.spark.deploy.yarn.ApplicationMaster, --class, org.apache.spark.examples.SparkPi, --jar , file:/usr/lib/spark/examples/lib/spark-examples_2.10-1.0.0-cdh5.1.0.jar, , --executor-memory, 1024, --executor-cores, 1, --num-executors , 2, 1>, <LOG_DIR>/stdout, 2>, <LOG_DIR>/stderr)
14/09/28 11:24:55 INFO Client: Submitting application to ASM
14/09/28 11:24:55 INFO YarnClientImpl: Submitted application application_1411874193696_0003
14/09/28 11:24:56 INFO Client: Application report from ASM:
application identifier: application_1411874193696_0003
appId: 3
clientToAMToken: null
appDiagnostics:
appMasterHost: N/A
appQueue: root.spark
appMasterRpcPort: -1
appStartTime: 1411874695327
yarnAppState: ACCEPTED
distributedFinalState: UNDEFINED
appTrackingUrl: http://hdp01:8088/proxy/application_1411874193696_0003/
appUser: spark

1.由client向ResourceManager提交请求，并上传Jar到HDFS上

这期间包括四个步骤:

a).连接到RM

b).从RM ASM(applicationsManager)中获得metric，queue和resource等信息。

c).upload app jar and spark-assembly jar

d).设置运行环境和container上下文

2.ResourceManager向NodeManager申请资源，创建Spark ApplicationMaster（每个SparkContext都有一个ApplicationManager）

3.NodeManager启动Spark App Master，并向ResourceManager ASM注册

4.Spark ApplicationMaster从HDFS中找到jar文件，启动DAGScheduler和YARN Cluster Scheduler

5.ResourceManager向ResourceManager ASM注册申请container资源(INFO YarnClientImpl: Submitted application)

6.ResourceManager通知NodeManager分配Container，这是可以收到来自ASM关于container的报告。(每个container的对应一个executor)

7.Spark ApplicationMaster直接和container(executor)进行交互，完成这个分布式任务。

需要注意的是：
a). Spark中的localdir会被yarn.nodemanager.local-dirs替换
b). 允许失败的节点数(spark.yarn.max.worker.failures)为executor数量的两倍数量，最小为3.
c). SPARK_YARN_USER_ENV传递给spark进程的环境变量
d). 传递给app的参数应该通过–args指定

II. yarn-client

(YarnClientClusterScheduler)查看对应类的文件

在Yarn-client模式下，Driver运行在Client上，通过ApplicationMaster向RM获取资源。本地Driver负责与所有的executor container进行交互，并将最后的结果汇总。结束掉终端，相当于kill掉这个spark应用。一般来说，如果运行的结果仅仅返回到terminal上时需要配置这个。

客户端的Driver将应用提交给Yarn后，Yarn会先后启动ApplicationMaster和excutor,另外ApplicationMaster和executor都装在在container里运行，container默认的内存是1g，ApplicationMaster分配的内存是driver-memory,executor分配的内存是executor-memory.同时，因为Driver在客户端，所以程序的运行结果可以在客户端显示，Driver以进程名为SparkSubmit的形式存在。

配置Yarn-client模式统一需要HADOOP_CONF_DIR/YARN_CONF_DIR和SPARK_JAR变量

提交任务测试:

spark-submit --class org.apache.spark.examples.SparkPi --deploy-mode client /usr/lib/spark/examples/lib/spark-examples_2.10-1.0.0-cdh5.1.0.jar
terminal output：

14/09/28 11:18:34 INFO Client: Command for starting the Spark ApplicationMaster: List($JAVA_HOME/bin/java, -server, -Xmx512m, -Djava.io.tmpdir=$PWD/tmp, -Dspark.tachyonStore.folderName="spark-9287f0f2-2e72-4617-a418-e0198626829b", -Dspark.eventLog.enabled="true", -Dspark.yarn.secondary.jars="", -Dspark.driver.host="hdp01", -Dspark.driver.appUIHistoryAddress="", -Dspark.app.name="Spark Pi", -Dspark.jars="file:/usr/lib/spark/examples/lib/spark-examples_2.10-1.0.0-cdh5.1.0.jar", -Dspark.fileserver.uri="http://172.19.17.231:53558", -Dspark.eventLog.dir="/user/spark/applicationHistory", -Dspark.master="yarn-client", -Dspark.driver.port="35938", -Dspark.httpBroadcast.uri="http://172.19.17.231:43804", -Dlog4j.configuration=log4j-spark-container.properties, org.apache.spark.deploy.yarn.ExecutorLauncher, --class, notused, --jar , null, --args ‘hdp01:35938‘ , --executor-memory, 1024, --executor-cores, 1, --num-executors , 2, 1>, <LOG_DIR>/stdout, 2>, <LOG_DIR>/stderr)
14/09/28 11:18:34 INFO Client: Submitting application to ASM
14/09/28 11:18:34 INFO YarnClientSchedulerBackend: Application report from ASM:
appMasterRpcPort: -1
appStartTime: 1411874314198
yarnAppState: ACCEPTED
......

最后将结果输出到terminal中