此文已由作者岳猛授权网易云社区发布。
欢迎访问网易云社区,了解更多网易技术产品运营经验。
Flink on yarn的启动流程可以参见前面的文章 Flink on Yarn启动流程,下面主要是从源码角度看下这个实现,可能有的地方理解有误,请给予指正,多谢。
--> 1.命令行启动yarn session
bin/yarn-session.sh -n 3 -jm 1024 -nm 1024 -st
我们去看下启动脚本
$JAVA_RUN $JVM_ARGS -classpath "$CC_CLASSPATH:$HADOOP_CLASSPATH:$HADOOP_CONF_DIR:$YARN_CONF_DIR" $log_setting org.apache.flink.yarn.cli.FlinkYarnSessionCli -j "$FLINK_LIB_DIR"/flink-dist*.jar "$@"
主要是用java -cp的方式启动主类** *org.apache.flink.yarn.cli.FlinkYarnSessionCli * , $@ 就是我们传入的哪些参数 " -n 3 -jm 1024 -nm 1024 -st" **。
1. FlinkYarnSessionCli 的启动流程分析
首先看下Main函数
public static void main(String[] args) { FlinkYarnSessionCli cli = new FlinkYarnSessionCli("", ""); // no prefix for the YARN session System.exit(cli.run(args)); }
主要是构造FlinkYarnSessionCli,然后执行其run方法,这里主要介绍主要流程的代码。
public int run(String[] args)
-
1.解析命令行参数
cmd = parser.parse(options, args)
-
2.根据命令行参数决定执行那种模式。
# 第一种,判断命令是否包含 -q
** 如: **
# 第二种,判断是否有-id参数
这里我们看下交互模式是啥样的,共有两个可选项,help和stop,如果我们敲入stop,则应用对应的所有进程会退出。
# 第三种,为正常模式
** 这里主要为构造YarnClusterDescriptor,然后调用其deploy方法启动集群 ,接着将Jobmanager和web ui地址写入到out文件中去,如果采用分离模式,则等待集群启动之后yarn session自动退出,如果不是则进入交互模式,我们可以通过交互控制这个Applitcation **
接着看下是如何构造YarnClusterDescriptor的
----------------- **1 creat YarnClusterDescriptor ** ----------------------
直接new YarnClusterDescriptor对象,然后将依赖jar地址,配置参数如taskmanager个数,jar地址,配置文件地址,配置参数等设置到YarnClusterDescriptor对象中去,然后返回这个对象。
------------** 2 YarnClusterDescriptor deploy ** -------------------------
由于YarnClusterDescriptor没有重写depoy方法则直接调用其父类AbstractYarnClusterDescriptor的deploy方法,但是最终调用的是其deployInternal方法.
接着看下deployInternal方法,简单的描述下流程,后续代码分析下面的github地址
-
检查是否具备Deploy的条件,如配置文件,jar路径是否为空
-
获取yarn的client,用户和RM进行通信
-
增加动态的配置属性到配置conf对象中去,解析配置conf对象为kv对
-
获取HDFS FileSyetem,这里用于将本地jar及配置文件上传到HDFS,
-
判断JobManager和TaskManager申请的资源是否满足yarn分配单个container的最小分配,如果小于则将container最小分配用来初始化jobMananger和TaskMananer
-
通过yarn client创建Application,返回GetNewApplicationResponse对象用于跟RM进行RPC通信。
-
通过GetNewApplicationResponse对象获取RM能够为这个应用分配的最大资源,如果最大资源不能够满足jobManagerMemoryMb和taskManagerMemoryMb则报错,计算总的jobmanager和所有taskmanager总共需要的资源(jobManagerMemoryMb + taskManagerMemoryMb * taskManagerCount),计算RM中总共的空闲资源,判断空闲资源是否满足前面计算需要的需求,如果不满足,则可能先启动yarn session,task manager等到有资源再进行启动;先为jobManager分配一个nm,然后再在其他的nm上启动taskmanager
-
设置启动ApplicationMaster的 lanchcontext,这里主要是设置java home,主类,jvm参数数,log文件配置。ApplicationMaster的主类 YarnApplicationMasterRunner ** YarnApplicationMasterRunner **。
protected Class<?> getApplicationMasterClass() { return YarnApplicationMasterRunner.class; }``` - 设置ApplicationSubmissionContext,获取ApplicationId - 设置session需要的hdfs路径,然后将本地jar包及配置文件,配置文件上传到HDFS - 设置AM启动的token信息,设置AM启动的过程中需要从hdfs下载那些依赖的jar和配置文件,设置ApplicationMaster及Flink及其他进程的classpath,不多说 - 设置钩子函数在deploy的时候清理上传到hdfs的文件及本地下载的依赖文件 - *** 重点,提交Applicaiton到RM;设置这个Application的状态为NEW,然后监控这个应用,如果不是之前的NEW状态,则打印当前状态,如果Running状态则跳出这个循环,如果是其他状态,则抛出YarnDeploymentException异常,上层调用捕获处理吧,不然250ms判断一次 *** - depoly成功,钩子函数删除临时文件,如依赖的jar包和配置文件等,返回YarnClusterClient对象,包含了这YarnClusterDescriptor,ApplicationReport等重要的属性。 *** ***deploy 成功以后进入交互模式,在runInteractiveCli里面最重要的一步是构造ApplicationClient Actor用于和JobManager Actor进行通信,但是如果发送 RegisterInfoMessageListener、UnRegisterInfoMessageListener等消息,将会由jobmanager actor将forward方法路由到flink resource manager actor去处理,此时jobmanager作为flink resource manager的代理,此时收到这两个消息的时候,由于是forward的方法,sender仍然是application client actor,所以flink manager resource actor可以直接给application client返回消息*** *** > ------------ ** 3 代码展示主要流程**------ ![](https://img.haomeiwen.com/i3249301/d22456f0939a8365.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/9c80aa18467d4e10.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/85a2f462ff96e5fd.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/a3c81e3dc9b23db0.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/bf190e6a72366f0d.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/57eb01f090d38dd3.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/d548d544dbd1b713.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/2013feca33032c46.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/d0d8c8c1a56f28ff.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/be3a228edeffad9d.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)*** ---- ApplicationClient 和JobManager Actor通信代码 --*** ![](https://img.haomeiwen.com/i3249301/56371ec18930ba4f.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/ed28091d44dc3906.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/1a0df11e1a57941d.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/913d82bf6d5825b8.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/6309a49886d0cc4e.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/47f86bcfdee4f967.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/6a483d4af26931cc.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/4de478f435cb1356.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/27986c3659bd96cc.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/a27f89acbe3406de.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/f04a3da4f97a08dc.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)## 2. YarnApplicationMasterRunner 启动流程分析 *** RM首先分配一个NM的container去启动YarnApplicationMasterRunner ,接着下来我们看下是怎么做的*** 首先是进入main函数里面,构造一个YarnApplicationMasterRunner对象,直接调用其Run方法。 > run方法主要步骤 - 获取当前用户的UGI及远端UGI - 将当前用户ugi里面的token传递到远端的UGI中,用于数据和服务访问 - 在远端的UGI里面执行runApplicationMaster启动ApplicationMaster ![](https://img.haomeiwen.com/i3249301/7d3ac2af1bf091f6.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)> runApplicationMaster主要过程,这里注释很清楚,我只捡重要的提示下 - 1) load and parse / validate all configurations - 2) start the actor system,try to start the actor system, JobManager and JobManager actor system - 3) Generate the configuration for the TaskManagers,这里主要是JobManager的地址,taskManager注册的超时时间,slot个数,这里还有最重要的一步是构造TaskManager的ContainerLaunchContext,这个context里面包含了启动TaskManager的启动命令,***主类是YarnTaskManager***。 - start the actors and components in this order: 1) JobManager & Archive (in non-HA case, the leader service takes this),启动JobManagerActor,这里主类是***YarnJobManager*** 2) Web Monitor (we need its port to register) 启动WEB监控页面,创建LeaderRetrievalService对象,这个主要用于启动TaskManager的时候,告诉TaskManager JobManager得akka url,用于TaskManager向JobManager进行注册。 3) Resource Master for YARN 启动YarnFlinkResourceManager Actor,这里主要用于Flink container资源的管理包括申请与释放等。 4) Process reapers for the JobManager and Resource Master ***这里主要介绍YarnApplicationMasterRunner 是如何通过YarnFlinkResourceManager去完成container的申请与启动TaskManager的,这里相对来说,比较复杂,我跟到Yarn的代码里才算整明白*** ![YarnFlinkResourceManager的继承关系](https://img.haomeiwen.com/i3249301/cbc8215f8c356913.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)说明YarnFlinkResourceManager其实是一个actor,在runApplicationMaster方法中,通过下面的代码启动这个Actor
Props resourceMasterProps = YarnFlinkResourceManager.createActorProps(
getResourceManagerClass(),//YarnFlinkResourceManager
config,
yarnConfig,
leaderRetriever,
appMasterHostname,
webMonitorURL,
taskManagerParameters,
taskManagerContext,
numInitialTaskManagers,
LOG);
ActorRef resourceMaster = actorSystem.actorOf(resourceMasterProps);//启动YarnFlinkResourceManager actor
接着看下YarnFlinkResourceManager 的构造方法,这里主要有三个成员变量比较重要
//在yarn 的rm端会调用该回对象的回调函数进行container申请,resourceManagerCallbackHandler里面只有该actor的actor ref,所以回调的过程中能够与该actor进行通信
/** Callback handler for the asynchronous resourceManagerClient /
private YarnResourceManagerCallbackHandler resourceManagerCallbackHandler;
//AM与RM通信的client,resourceManagerClient对象持有resourceManagerCallbackHandler
/* Client to communicate with the Resource Manager (YARN's master) /
private AMRMClientAsync<AMRMClient.ContainerRequest> resourceManagerClient;
//AM与NM的通信client
/* Client to communicate with the Node manager and launch TaskManager processes */
private NMClient nodeManagerClient;
YarnFlinkResourceManager 启动的过程先执行preStart方法,自己没有实现则执行其父类FlinkResourceManager的preStart方法。接着调用YarnFlinkResourceManager 的initialize方法。 > ***在initialize方法里面*** *** resourceManagerClient.start() ----> AMRMClientAsyncImpl.serviceStart()---> CallbackHandlerThread.start()(守护线程)---> YarnResourceManagerCallbackHandler.onContainersAllocated(allocated)---> yarnFrameworkMaster.tell(new ContainersAllocated(containers),ActorRef.noSender())(yarnFrameworkMaster为YarnFlinkResourceManager ActorRef) --> YarnFlinkResourceManager .containersAllocated --> NMClient.startContainer(container, taskManagerLaunchContext) 至此通知各个NM启动container。*** ![](https://img.haomeiwen.com/i3249301/3a5b6377e93742b9.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/2c532578dc7a508d.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/6a2cd3067970fe39.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/e14cfee9289c57bf.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/2e3b73f252030ea8.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/a3239382a6bde777.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/0a52a463c263bbd7.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/52cecffb2670cc67.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)***至此,YarnApplicationMasterRunner 重要的流程已经说完,细节东西太多,就不再说了,有时间再看,接下来看YarnTaskManager的部分***## 3. YarnTaskManager启动流程分析***接上面nodeManagerClient.startContainer(container, taskManagerLaunchContext)将通知NM去启动container,NM根据taskManagerLaunchContext的启动信息,从HDFS下载YarnTaskManager启动过程依赖的jar和配置文件 (container_tokens default_container_executor_session.sh default_container_executor.sh flink-conf.yaml flink.jar launch_container.sh lib log4j.properties logback.xml),然后shell执行launch_container.sh,最终用java -cp启动YarnTaskManager进程,启动进程的时候首先执行YarnTaskManager run方法,TaskManager会拿到JobManager的akka地址,然后发送注册消息,JobManager收到注册消息以后,注册成功之后就发送ack确认注册信息给TaskManager,然后TaskManger根据配置以及JobManager返回过来的信息构建一些真正干活的成员变量。过程:*** > YarnTaskManagerRunner.runYarnTaskManager(args, classOf[YarnTaskManager])--> TaskManager.selectNetworkInterfaceAndRunTaskManager(configuration, resourceId, taskManager)--> TaskManager.runTaskManager --> TaskManager.startTaskManagerComponentsAndActor--> actorSystem.actorOf(tmProps, actorName)--> TaskManager.preStart--> StandaloneLeaderRetrievalService.start(TaskManager)--> TaskManger.notifyLeaderAddress--> TaskManager.handleJobManagerLeaderAddress--> TaskManager.triggerTaskManagerRegistration() TaskManager.handleRegistrationMessage--> instanceManager.registerTaskManager--> jobManager 发送消息AcknowledgeRegistration给TaskManager TaskManager.associateWithJobManager--> ![](https://img.haomeiwen.com/i3249301/8797abe4fead540e.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/6ada492f0e25718b.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/7d9e8a6893af056a.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/e4116544f6c7a677.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/03e551a6ebf785c1.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![Paste_Image.png](https://img.haomeiwen.com/i3249301/155d80ccd8b5bed4.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![Paste_Image.png](https://img.haomeiwen.com/i3249301/40bbe88a86f090c8.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![Paste_Image.png](https://img.haomeiwen.com/i3249301/2226948e829a5add.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![Paste_Image.png](https://img.haomeiwen.com/i3249301/b9c444d706737b6f.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/3ff9f4e9243c1264.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/08823728261241ec.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![associateWithJobManager](https://img.haomeiwen.com/i3249301/f3aa8048869975c8.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/1bb10a80436aa199.png?imageMogr2/auto-orient/strip%7CimageView2/2/w/1240)![](https://img.haomeiwen.com/i3249301/442effae29613f2c.png)###基本上Flink on yarn的流程就是这样,细节需要深入,有不正确的地方,希望给予指正。
链接:https://www.jianshu.com/p/8a3177095072
更多网易技术、产品、运营经验分享请点击。
相关文章:
【推荐】 分布式存储系统可靠性系列三:设计模式
【推荐】 种草社区缓存设计
网友评论