flink on yarn部分源码解析
转发请注明原创地址:https://www.cnblogs.com/dongxiao-yang/p/9403427.html
flink任务的deploy形式有很多种选择,常见的有standalone,on yarn , Meos , Kubernetes等方式,目前公司内部统一采用flink on yarn的 single job模式(每个flink job 单独在yarn上声明一个flink集群),本文分析的是flink1.5.1版本源码使用legacy 模式提交yarn single job到yarn集群的部分源码。
典型的flink提交single job命令格式如下: ./flink run -m yarn-cluster -d -yst -yqu flinkqu -yst -yn 4 -ys 2 -c flinkdemoclass flinkdemo.jar args1 args2 ...
flink脚本的入口类为org.apache.flink.client.cli.CliFrontend
在CliFrontend的main函数中首先通过loadCustomCommandLines方法加载了提交yarn任务初始化一个重要工具类
org.apache.flink.yarn.cli.FlinkYarnSessionCli
public static List<CustomCommandLine<?>> loadCustomCommandLines(Configuration configuration, String configurationDirectory) {
List<CustomCommandLine<?>> customCommandLines = new ArrayList<>(2); // Command line interface of the YARN session, with a special initialization here
// to prefix all options with y/yarn.
// Tips: DefaultCLI must be added at last, because getActiveCustomCommandLine(..) will get the
// active CustomCommandLine in order and DefaultCLI isActive always return true.
final String flinkYarnSessionCLI = "org.apache.flink.yarn.cli.FlinkYarnSessionCli";
try {
customCommandLines.add(
loadCustomCommandLine(flinkYarnSessionCLI,
configuration,
configurationDirectory,
"y",
"yarn"));
} catch (NoClassDefFoundError | Exception e) {
LOG.warn("Could not load CLI class {}.", flinkYarnSessionCLI, e);
} if (configuration.getString(CoreOptions.MODE).equalsIgnoreCase(CoreOptions.NEW_MODE)) {
customCommandLines.add(new DefaultCLI(configuration));
} else {
customCommandLines.add(new LegacyCLI(configuration));
} return customCommandLines;
}
根据启动参数,CliFrontend开始运行方法run()->runProgram(),runProgram内部与yarn相关的一个重点方法为
client = clusterDescriptor.deploySessionCluster(clusterSpecification);
上文中的clusterDescriptor就是前面的FlinkYarnSessionCli执行createClusterDescriptor()方法后产生的集群属性描述对象,在本模式中对应的具体类是org.apache.flink.yarn.LegacyYarnClusterDescriptor,父类为AbstractYarnClusterDescriptor
deploySessionCluster内部进一步调用deployInternal来向yarn集群提交一个flink集群。
protected ClusterClient<ApplicationId> deployInternal(
ClusterSpecification clusterSpecification,
String applicationName,
String yarnClusterEntrypoint,
@Nullable JobGraph jobGraph,
boolean detached) throws Exception { // ------------------ Check if configuration is valid --------------------
validateClusterSpecification(clusterSpecification); if (UserGroupInformation.isSecurityEnabled()) {
// note: UGI::hasKerberosCredentials inaccurately reports false
// for logins based on a keytab (fixed in Hadoop 2.6.1, see HADOOP-10786),
// so we check only in ticket cache scenario.
boolean useTicketCache = flinkConfiguration.getBoolean(SecurityOptions.KERBEROS_LOGIN_USETICKETCACHE); UserGroupInformation loginUser = UserGroupInformation.getCurrentUser();
if (loginUser.getAuthenticationMethod() == UserGroupInformation.AuthenticationMethod.KERBEROS
&& useTicketCache && !loginUser.hasKerberosCredentials()) {
LOG.error("Hadoop security with Kerberos is enabled but the login user does not have Kerberos credentials");
throw new RuntimeException("Hadoop security with Kerberos is enabled but the login user " +
"does not have Kerberos credentials");
}
} isReadyForDeployment(clusterSpecification); // ------------------ Check if the specified queue exists -------------------- checkYarnQueues(yarnClient); // ------------------ Add dynamic properties to local flinkConfiguraton ------
Map<String, String> dynProperties = getDynamicProperties(dynamicPropertiesEncoded);
for (Map.Entry<String, String> dynProperty : dynProperties.entrySet()) {
flinkConfiguration.setString(dynProperty.getKey(), dynProperty.getValue());
} // ------------------ Check if the YARN ClusterClient has the requested resources -------------- // Create application via yarnClient
final YarnClientApplication yarnApplication = yarnClient.createApplication();
final GetNewApplicationResponse appResponse = yarnApplication.getNewApplicationResponse(); Resource maxRes = appResponse.getMaximumResourceCapability(); final ClusterResourceDescription freeClusterMem;
try {
freeClusterMem = getCurrentFreeClusterResources(yarnClient);
} catch (YarnException | IOException e) {
failSessionDuringDeployment(yarnClient, yarnApplication);
throw new YarnDeploymentException("Could not retrieve information about free cluster resources.", e);
} final int yarnMinAllocationMB = yarnConfiguration.getInt(yarnConfiguration.RM_SCHEDULER_MINIMUM_ALLOCATION_MB, 0); final ClusterSpecification validClusterSpecification;
try {
validClusterSpecification = validateClusterResources(
clusterSpecification,
yarnMinAllocationMB,
maxRes,
freeClusterMem);
} catch (YarnDeploymentException yde) {
failSessionDuringDeployment(yarnClient, yarnApplication);
throw yde;
} LOG.info("Cluster specification: {}", validClusterSpecification); final ClusterEntrypoint.ExecutionMode executionMode = detached ?
ClusterEntrypoint.ExecutionMode.DETACHED
: ClusterEntrypoint.ExecutionMode.NORMAL; flinkConfiguration.setString(ClusterEntrypoint.EXECUTION_MODE, executionMode.toString()); ApplicationReport report = startAppMaster(
flinkConfiguration,
applicationName,
yarnClusterEntrypoint,
jobGraph,
yarnClient,
yarnApplication,
clusterSpecification); String host = report.getHost();
int port = report.getRpcPort(); // Correctly initialize the Flink config
flinkConfiguration.setString(JobManagerOptions.ADDRESS, host);
flinkConfiguration.setInteger(JobManagerOptions.PORT, port); flinkConfiguration.setString(RestOptions.ADDRESS, host);
flinkConfiguration.setInteger(RestOptions.PORT, port); // the Flink cluster is deployed in YARN. Represent cluster
return createYarnClusterClient(
this,
clusterSpecification.getNumberTaskManagers(),
clusterSpecification.getSlotsPerTaskManager(),
report,
flinkConfiguration,
true);
}
deployInternal方法开头对yarn集群的可用内存,queue等进行检查后申请了一个application,并调用startAppMaster声明了AM的启动类:YarnApplicationMasterRunner
public ApplicationReport startAppMaster(
Configuration configuration,
String applicationName,
String yarnClusterEntrypoint,
JobGraph jobGraph,
YarnClient yarnClient,
YarnClientApplication yarnApplication,
ClusterSpecification clusterSpecification) throws Exception {
..... setApplicationTags(appContext); // add a hook to clean up in case deployment fails
Thread deploymentFailureHook = new DeploymentFailureHook(yarnClient, yarnApplication, yarnFilesDir);
Runtime.getRuntime().addShutdownHook(deploymentFailureHook);
LOG.info("Submitting application master " + appId);
yarnClient.submitApplication(appContext); LOG.info("Waiting for the cluster to be allocated");
final long startTime = System.currentTimeMillis();
ApplicationReport report; }
YarnApplicationMasterRunner会在yarn集群上作为appmaster与resourcemanager通信申请对应的Taskmanagercontainer服务,启动jobmanager服务和webui服务等
protected int runApplicationMaster(Configuration config) {
......
......
webMonitor = BootstrapTools.startWebMonitorIfConfigured(
config,
highAvailabilityServices,
new AkkaJobManagerRetriever(actorSystem, webMonitorTimeout, 10, Time.milliseconds(50L)),
new AkkaQueryServiceRetriever(actorSystem, webMonitorTimeout),
webMonitorTimeout,
new ScheduledExecutorServiceAdapter(futureExecutor),
LOG); metricRegistry = new MetricRegistryImpl(
MetricRegistryConfiguration.fromConfiguration(config)); metricRegistry.startQueryService(actorSystem, null); // 2: the JobManager
LOG.debug("Starting JobManager actor"); // we start the JobManager with its standard name
ActorRef jobManager = JobManager.startJobManagerActors(
config,
actorSystem,
futureExecutor,
ioExecutor,
highAvailabilityServices,
metricRegistry,
webMonitor == null ? Option.empty() : Option.apply(webMonitor.getRestAddress()),
new Some<>(JobMaster.JOB_MANAGER_NAME),
Option.<String>empty(),
getJobManagerClass(),
getArchivistClass())._1(); final String webMonitorURL = webMonitor == null ? null : webMonitor.getRestAddress(); // 3: Flink's Yarn ResourceManager
LOG.debug("Starting YARN Flink Resource Manager"); Props resourceMasterProps = YarnFlinkResourceManager.createActorProps(
getResourceManagerClass(),
config,
yarnConfig,
highAvailabilityServices.getJobManagerLeaderRetriever(HighAvailabilityServices.DEFAULT_JOB_ID),
appMasterHostname,
webMonitorURL,
taskManagerParameters,
taskManagerContext,
numInitialTaskManagers,
LOG); ActorRef resourceMaster = actorSystem.actorOf(resourceMasterProps);
另一方面,flink客户端在提交完集群后从runprogram()方法进入executeProgram();
protected void executeProgram(PackagedProgram program, ClusterClient<?> client, int parallelism) throws ProgramMissingJobException, ProgramInvocationException {
logAndSysout("Starting execution of program"); final JobSubmissionResult result = client.run(program, parallelism); if (null == result) {
throw new ProgramMissingJobException("No JobSubmissionResult returned, please make sure you called " +
"ExecutionEnvironment.execute()");
} if (result.isJobExecutionResult()) {
logAndSysout("Program execution finished");
JobExecutionResult execResult = result.getJobExecutionResult();
System.out.println("Job with JobID " + execResult.getJobID() + " has finished.");
System.out.println("Job Runtime: " + execResult.getNetRuntime() + " ms");
Map<String, Object> accumulatorsResult = execResult.getAllAccumulatorResults();
if (accumulatorsResult.size() > 0) {
System.out.println("Accumulator Results: ");
System.out.println(AccumulatorHelper.getResultsFormatted(accumulatorsResult));
}
} else {
logAndSysout("Job has been submitted with JobID " + result.getJobID());
}
}
代码从ClusterClient.run()->prog.invokeInteractiveModeForExecution()开始真正进入用户flink job的main方法。
main方法中,代码最后的env.execute() 会把生成job的执行plan并返回对应的DetachedEnvironment对象。
方法调用链路为DetachedEnvironment.finalizeExecute()->ClusterClient.run()->YarnClusterClient.submitJob->ClusterClient.runDetached();
/**
* Submits a JobGraph detached.
* @param jobGraph The JobGraph
* @param classLoader User code class loader to deserialize the results and errors (may contain custom classes).
* @return JobSubmissionResult
* @throws ProgramInvocationException
*/
public JobSubmissionResult runDetached(JobGraph jobGraph, ClassLoader classLoader) throws ProgramInvocationException { waitForClusterToBeReady(); final ActorGateway jobManagerGateway;
try {
jobManagerGateway = getJobManagerGateway();
} catch (Exception e) {
throw new ProgramInvocationException("Failed to retrieve the JobManager gateway.", e);
} try {
logAndSysout("Submitting Job with JobID: " + jobGraph.getJobID() + ". Returning after job submission.");
JobClient.submitJobDetached(
new AkkaJobManagerGateway(jobManagerGateway),
flinkConfig,
jobGraph,
Time.milliseconds(timeout.toMillis()),
classLoader);
return new JobSubmissionResult(jobGraph.getJobID());
} catch (JobExecutionException e) {
throw new ProgramInvocationException("The program execution failed: " + e.getMessage(), e);
}
}
@Override
public JobSubmissionResult submitJob(JobGraph jobGraph, ClassLoader classLoader) throws ProgramInvocationException {
if (isDetached()) {
if (newlyCreatedCluster) {
stopAfterJob(jobGraph.getJobID());
}
LOG.info("super.runDetached");
return super.runDetached(jobGraph, classLoader);
} else {
LOG.info("super.run");
return super.run(jobGraph, classLoader);
}
}
最后,客户端连接到前文对应的jobmanager服务并把flink job grafaph提交给yarn上已经申请好的flink集群。
结论:flink on yarn的single job模式提交作业的逻辑为flink客户端首先申请一个yarn集群的application,等待集群成功部署后再联系jobmanager并把job提交到集群上面。这个模式的优点是每个
flink job有一个独立的集群便于资源规划和管理,缺点是经过验证在am挂掉后yarn只能把原来的集群重启回来但是无法恢复flink jobgraph的行为,所以需要额外配置ha信息。
flink on yarn部分源码解析的更多相关文章
- flink on yarn部分源码解析 (FLIP-6 new mode)
我们在https://www.cnblogs.com/dongxiao-yang/p/9403427.html文章里分析了flink提交single job到yarn集群上的代码,flink在1.5版 ...
- [源码解析] 从TimeoutException看Flink的心跳机制
[源码解析] 从TimeoutException看Flink的心跳机制 目录 [源码解析] 从TimeoutException看Flink的心跳机制 0x00 摘要 0x01 缘由 0x02 背景概念 ...
- [源码解析] 当 Java Stream 遇见 Flink
[源码解析] 当 Java Stream 遇见 Flink 目录 [源码解析] 当 Java Stream 遇见 Flink 0x00 摘要 0x01 领域 1.1 Flink 1.2 Java St ...
- [源码解析] Flink的Slot究竟是什么?(1)
[源码解析] Flink的Slot究竟是什么?(1) 目录 [源码解析] Flink的Slot究竟是什么?(1) 0x00 摘要 0x01 概述 & 问题 1.1 Fllink工作原理 1.2 ...
- [源码解析] Flink的Slot究竟是什么?(2)
[源码解析] Flink 的slot究竟是什么?(2) 目录 [源码解析] Flink 的slot究竟是什么?(2) 0x00 摘要 0x01 前文回顾 0x02 注册/更新Slot 2.1 Task ...
- 《Flink 源码解析》—— 源码编译运行
更新一篇知识星球里面的源码分析文章,去年写的,周末自己录了个视频,大家看下效果好吗?如果好的话,后面补录发在知识星球里面的其他源码解析文章. 前言 之前自己本地 clone 了 Flink 的源码,编 ...
- Flink 源码解析 —— 源码编译运行
更新一篇知识星球里面的源码分析文章,去年写的,周末自己录了个视频,大家看下效果好吗?如果好的话,后面补录发在知识星球里面的其他源码解析文章. 前言 之前自己本地 clone 了 Flink 的源码,编 ...
- Flink 源码解析 —— 如何获取 ExecutionGraph ?
https://t.zsxq.com/UnA2jIi 博客 1.Flink 从0到1学习 -- Apache Flink 介绍 2.Flink 从0到1学习 -- Mac 上搭建 Flink 1.6. ...
- Flink 源码解析 —— 深度解析 Flink 是如何管理好内存的?
前言 如今,许多用于分析大型数据集的开源系统都是用 Java 或者是基于 JVM 的编程语言实现的.最着名的例子是 Apache Hadoop,还有较新的框架,如 Apache Spark.Apach ...
随机推荐
- C语言基本数据类型简介
1.概述 C 语言包含的数据类型如下图所示: 2.各种数据类型介绍 2.1整型 整形包括短整型.整形和长整形. 2.1.1短整形 short a=1; 2.1.2整形 一般占4个字节(32位),最高位 ...
- React个人学习笔记
元素渲染 通过 ReactDOM.render() 方法渲染页面, 可以使用 ES6 class 来定义一个组件: 如何解析HTMl里面的空格: 1. 使用空格的 unicod 编码 : \u0020 ...
- Flask参数解析、请求钩子
转载请注明出处 https://www.cnblogs.com/chenxianpao/p/9949279.html 参数解析 Flask的参数解析主要用Request完成(from flask i ...
- 【笔记】关于require.js 的用法
最近忙于学校的一个新网站建设,对于以前的前端程序编写方式的不正规特意上网学习了require.js 的用法,使此次的工程更加有条理同时符合当前前端的开发模式——前端模块化. 网上有不少很好的学习文章这 ...
- (转)MySQL的JDBC驱动源码解析
一.背景 MySQL是一个中小型关系型数据库管理系统,目前我们淘宝也使用的也非常广泛.为了对开发中间DAO持久层的问题能有更深的理解以及最近在使用的phoenix on Hbase的SQL也是实现的J ...
- 2017.12.27 sqlSessionFactory和sqlSession(to be continued)
参考来自:<深入浅出MyBatis技术原理与实践-第6章 > 1.SqlSessionFactory SqlSessionFactory是一个接口,最重要的功能是提供SqlSession. ...
- BST数据结构题
给定BST.改动BST,使得每一个点都是大于他的结点的值之和 关键是这题递归參数怎么设计,每一个点比他大的有两快.一个是右子书(假设有的话),还有一个是祖先里面比他大的,假设直接用这两个的话,找不到递 ...
- STL经常使用遍历算法for_each和transform的比較
for_each()和transform()算法比較 1)STL 算法 – 改动性算法 for_each() copy() copy_backward() transform() merge ...
- ORA-12519, TNS:no appropriate service handler found(数据库上当前的连接数目已经超过最大值)
报错: ORA-12519, TNS:no appropriate service handler foundThe Connection descriptor used by the client ...
- 微信小程序之下拉刷新,上拉更多列表实现
代码地址如下:http://www.demodashi.com/demo/11110.html 一.准备工作 首先需要下载小程序开发工具 官方下载地址: https://mp.weixin.qq.co ...