在 YARN 上运行任意命令

最后更新: 2026-08-04 06:22:21 +0800 +0800
tags: Spark | Yarn | Hadoop
原文以英文撰写。您正在阅读AI翻译版本。

向 YARN 提交应用的通用步骤 #

  1. 准备应用资源。 打包客户端和 ApplicationMaster 代码,然后确保 YARN 可以将其本地化。

    java
    FileSystem fs = FileSystem.get(conf);
    Path stagingDir = new Path(fs.getHomeDirectory(), ".yarn-binary-demo/" + appId);
    fs.mkdirs(stagingDir);
    
    Map<String, LocalResource> amLocalResources = new HashMap<>();
    Path appJarDst = new Path(stagingDir, YarnBinaryDemoSupport.APP_JAR_NAME);
    amLocalResources.put(
            YarnBinaryDemoSupport.APP_JAR_NAME,
            YarnBinaryDemoSupport.copyToHdfsAsLocalResource(fs, new Path(appJar), appJarDst));
  2. 创建 YARN 客户端并分配新的应用 ID。 客户端连接 ResourceManager,并请求一个新的应用句柄。

    java
    YarnConfiguration conf = new YarnConfiguration();
    YarnClient yarnClient = YarnClient.createYarnClient();
    yarnClient.init(conf);
    yarnClient.start();
    
    YarnClientApplication app = yarnClient.createApplication();
    ApplicationId appId = app.getNewApplicationResponse().getApplicationId();
  3. 构建 ApplicationMaster 启动命令。 这是 YARN 用于启动 AM 容器的确切 Java 命令。

    java
    StringBuilder amCommand = new StringBuilder();
    amCommand.append(ApplicationConstants.Environment.JAVA_HOME.$$())
            .append("/bin/java -Xmx256m ")
            .append("demo.yarn.YarnBinaryApplicationMaster ")
            .append("--command ")
            .append(YarnBinaryDemoSupport.shellQuote(command))
            .append(" ");
  4. 将 AM 命令封装在 ContainerLaunchContext 中。 这会为 AM 汇总命令、资源和环境。

    java
    Map<String, String> env = new HashMap<>();
    env.put("CLASSPATH", YarnBinaryDemoSupport.buildClasspath(conf));
    
    ContainerLaunchContext amContext = ContainerLaunchContext.newInstance(
            amLocalResources,
            env,
            Collections.singletonList(amCommand.toString()),
            null,
            null,
            null
    );
  5. 向 ResourceManager 提交应用。 提交内容包括 AM 规范、应用名称、队列和 AM 资源需求。

    java
    Resource amResource = Resource.newInstance(512, 1);
    app.getApplicationSubmissionContext().setApplicationName("yarn-binary-demo");
    app.getApplicationSubmissionContext().setAMContainerSpec(amContext);
    app.getApplicationSubmissionContext().setResource(amResource);
    app.getApplicationSubmissionContext().setQueue("default");
    
    yarnClient.submitApplication(app.getApplicationSubmissionContext());
  6. YARN 启动 ApplicationMaster 后注册它。 AM 在其容器内启动后,会向 ResourceManager 通告自身。

    java
    AMRMClient<AMRMClient.ContainerRequest> rmClient = AMRMClient.createAMRMClient();
    rmClient.init(conf);
    rmClient.start();
    
    rmClient.registerApplicationMaster("", 0, "");
  7. 向 ResourceManager 请求工作容器。 AM 为实际工作负载请求资源。

    java
    Resource capability = Resource.newInstance(256, 1);
    Priority priority = Priority.newInstance(0);
    AMRMClient.ContainerRequest request =
            new AMRMClient.ContainerRequest(capability, null, null, priority);
    rmClient.addContainerRequest(request);
  8. 等待容器分配。 AM 会持续轮询 RM,直到容器可用。

    java
    Container allocatedContainer = null;
    while (allocatedContainer == null) {
        AllocateResponse response = rmClient.allocate(0.1f);
        for (Container container : response.getAllocatedContainers()) {
            allocatedContainer = container;
            break;
        }
    }
  9. 请求所属的 NodeManager 启动实际负载。 此时 AM 会向工作容器发送确切的 shell 命令。

    java
    String containerCommand =
            "/bin/bash -lc " + YarnBinaryDemoSupport.shellQuote(payload)
                    + " 1>" + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout"
                    + " 2>" + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr";
    
    ContainerLaunchContext launchContext = ContainerLaunchContext.newInstance(
            localResources,
            new HashMap<String, String>(),
            Collections.singletonList(containerCommand),
            null,
            null,
            null
    );
    
    nmClient.startContainer(allocatedContainer, launchContext);
  10. 监控完成情况并结束应用。 AM 监视容器结果,报告最终状态,然后正常退出。

java
while (!finished) {
    AllocateResponse response = rmClient.allocate(0.5f);
    for (ContainerStatus status : response.getCompletedContainersStatuses()) {
        if (status.getContainerId().equals(allocatedContainer.getId())) {
            exitCode = status.getExitStatus();
            diagnostics = status.getDiagnostics();
            finished = true;
            break;
        }
    }
}

FinalApplicationStatus finalStatus =
        exitCode == 0 ? FinalApplicationStatus.SUCCEEDED : FinalApplicationStatus.FAILED;
rmClient.unregisterApplicationMaster(finalStatus, diagnostics, "");

评论

在 GitHub 上阅读或评论(发表评论需要登录)