Drill-on-YARN之原始碼解析

1. 概要

前面介紹瞭如何把Drill部署在YARN上，然後通過Drill-on-YARN客戶端，你可以啟動、停止、調整、清零命令操作Drill。但是在這麼命令背後，到底是如何執行的呢，下面會對Drill-on-YARN的原始碼進行詳細的解析，重點解析啟動過程，其他命令簡單介紹。

說明：下面涉及到的程式碼，以drill 1.14.0為準，並且為了減少篇幅，進行了刪減。

2. Drill-on-YARN start

2.1 drill-on-yarn.sh

通過檢視drill-on-yarn.sh指令碼，很容易發現最終執行的java類是CLIENT_CMD="$JAVA $VM_OPTS -cp $CP org.apache.drill.yarn.client.DrillOnYarn ${args[@]}"。 org.apache.drill.yarn.client.DrillOnYarn便是啟動Drill-on-YARN的入口。我們可以總覽一下這個類：

public class DrillOnYarn {
  public static void main(String argv[]) {
    BasicConfigurator.configure();
    ClientContext.init();
    run(argv);
  }
  public static void run(String argv[]) {
    ClientContext context = ClientContext.instance();
    CommandLineOptions opts = new CommandLineOptions();
    if (!opts.parse(argv)) {
      opts.usage();
      context.exit(-1);
    }
    if (opts.getCommand() == null) {
      opts.usage();
      context.exit(-1);
    }
    try {
      DrillOnYarnConfig.load().setClientPaths();
    } catch (DoyConfigException e) {
      ClientContext.err.println(e.getMessage());
      context.exit(-1);
    }
    ClientCommand cmd;
    switch (opts.getCommand()) {
    case UPLOAD:
      cmd = new StartCommand(true, false);
      break;
    case START:
      cmd = new StartCommand(true, true);
      break;
    case DESCRIBE:
      cmd = new PrintConfigCommand();
      break;
    case STATUS:
      cmd = new StatusCommand();
      break;
    case STOP:
      cmd = new StopCommand();
      break;
    case CLEAN:
      cmd = new CleanCommand();
      break;
    case RESIZE:
      cmd = new ResizeCommand();
      break;
    default:
      cmd = new HelpCommand();
    }
    cmd.setOpts(opts);
    try {
      cmd.run();
    } catch (ClientException e) {
      displayError(opts, e);
      context.exit(1);
    }
  }
}

可以看到入口main方法，其中最關鍵的便是run方法，包含了很多的命令，我們重點看start命令，程式碼如下：

public void run() throws ClientException {
    checkExistingApp();

    dryRun = opts.dryRun;
    config = DrillOnYarnConfig.config();
    FileUploader uploader = upload();
    if (launch) {
      launch(uploader);
    }
}

概括的來說，它主要包含以下流程：

1. 檢查application是否已經存在，如果已經存在，便不允許啟動，否則執行啟動操作（此處檢查的application是YARN的application，啟動成功會將YARN的applicationId寫入本地磁碟的一個檔案，通過此檔案來檢查）。

2. 上傳Drill二方包和site目錄下的內容至DFS上，其中site目錄下的內容會被打包為site.tar.gz

   public void run() throws ClientException {
     setup();
     uploadDrillArchive();
     if (hasSiteDir()) {
       uploadSite();
     }
   }

3. 啟動ApplicationMaster，主要流程為：

   • 建立YARN客戶端，並啟動

     // AMRunner#connectToYarn
     private void connectToYarn() {
         System.out.print("Loading YARN Config...");
         client = new YarnRMClient();
         System.out.println(" Loaded.");
     }

   • 建立ApplicationMaster

     // AMRunner#createApp
     private void createApp() throws ClientException {
         try {
           appResponse = client.createAppMaster();
         } catch (YarnClientException e) {
           throw new ClientException("Failed to allocate Drill application master",
               e);
         }
         appId = appResponse.getApplicationId();
         System.out.println("Application ID: " + appId.toString());
     }

   • 設定ApplicationMaster上下文，包括：Heap memory、Class Path、啟動的命令（dirll-am.sh）、啟動am容器使用的資源（memory、vCores、disks）
   • 校驗資源，主要是ApplicationMaster使用資源是否超過了YARN的設定

   • 提交ApplicationMaster

     private void launchApp(AppSpec master) throws ClientException {
         try {
           client.submitAppMaster(master);
         } catch (YarnClientException e) {
           throw new ClientException("Failed to start Drill application master", e);
         }
     }

   • 等待啟動，並列印啟動日誌
   • 將ApplicationMaster的appid寫入檔案（在第1步，檢測Application是否存在，就是使用這個檔案）

ApplicationMaster啟動後，會向RM申請資源，啟動Drillbits，下面詳細介紹ApplicationMaster啟動後的操作

2.2 drill-am.sh

通過檢視drill-am.sh指令碼，很容易發現最終執行的java類是AMCMD="$JAVA $AM_JAVA_OPTS ${args[@]} -cp $CP org.apache.drill.yarn.appMaster.DrillApplicationMaster"。org.apache.drill.yarn.appMaster.DrillApplicationMaste表示ApplicationMaster執行的入口，下面總覽一下這個類：

public class DrillApplicationMaster {
  public static void main(String[] args) {
    LOG.trace("Drill Application Master starting.");
    try {
      DrillOnYarnConfig.load().setAmDrillHome();
    } catch (DoyConfigException e) {
      System.err.println(e.getMessage());
      System.exit(-1);
    }
    Dispatcher dispatcher;
    try {
      dispatcher = (new DrillControllerFactory()).build();
    } catch (ControllerFactoryException e) {
      LOG.error("Setup failed, exiting: " + e.getMessage(), e);
      System.exit(-1);
      return;
    }
    try {
      if (!dispatcher.start()) {
        return;
      }
    } catch (Throwable e) {
      LOG.error("Fatal error, exiting: " + e.getMessage(), e);
      System.exit(-1);
    }
    WebServer webServer = new WebServer(dispatcher);
    try {
      webServer.start();
    } catch (Exception e) {
      LOG.error("Web server setup failed, exiting: " + e.getMessage(), e);
      System.exit(-1);
    }
    try {
      dispatcher.run();
    } catch (Throwable e) {
      LOG.error("Fatal error, exiting: " + e.getMessage(), e);
      System.exit(-1);
    } finally {
      try {
        webServer.close();
      } catch (Exception e) {
      }
    }
  }
}

概況的來說，它主要包含以下流程：

1. 載入Drill-on-YARN的配置，並設定AM的DirllHome，比如/home/admin/tmp2/hadoop/nm-local-dir/usercache/admin/appcache/application_1534698866098_0022/container_1534698866098_0022_01_000001/drill/apache-drill-1.14.0

2. 構造Dispatcher，Dispatcher用於分配YARN、timer、ZooKeeper事件給給叢集控制器，它是輕量級多執行緒的，用於響應RM、NM、timer執行緒的事件，對於某一個事件，它是連續的，所以需要同步，但是不同型別的事件不需要同步。整個的構造流程如下：

   • 準備資源，包括：drill二方包、site壓縮包的目錄

     private Map<String, LocalResource> prepareResources() {
         ...
         drillArchivePath = drillConfig.getDrillArchiveDfsPath();
         siteArchivePath = drillConfig.getSiteArchiveDfsPath();
         ...
     }

   • 定義任務啟動的規格（TaskSpec），包括：執行時環境、YARN container的規格、dirllbit的規格

     private TaskSpec buildDrillTaskSpec(Map<String, LocalResource> resources) throws DoyConfigException {
         ...
         ContainerRequestSpec containerSpec = new ContainerRequestSpec();
         containerSpec.memoryMb = config.getInt(DrillOnYarnConfig.DRILLBIT_MEMORY);
         ...
         LaunchSpec drillbitSpec = new LaunchSpec();
         ...
         TaskSpec taskSpec = new TaskSpec();
         taskSpec.name = "Drillbit";
         taskSpec.containerSpec = containerSpec;
         taskSpec.launchSpec = drillbitSpec;
     }

   • 設定Dispatcher的控制器：實現類為ClusterControllerImpl，它主要通過狀態來控制Drill叢集、調整整個叢集的任務（Drill啟動、停止等任務）、處理container的回撥

     public void setYarn(AMYarnFacade yarn) throws YarnFacadeException {
             this.yarn = yarn;
             controller = new ClusterControllerImpl(yarn);
         }

   • 為控制器註冊Scheduler，比如DrillbitScheduler，此外Scheduler配置來源於之前drill-on-yarn.conf

     cluster: [
         {
           name: "drill-group1"
           type: "basic"
           count: 1
         }
     ]

     ...
     ClusterDef.ClusterGroup pool = ClusterDef.getCluster(config, 0);
     Scheduler testGroup = new DrillbitScheduler(pool.getName(), taskSpec,
     pool.getCount(), requestTimeoutSecs, maxExtraNodes);
     dispatcher.getController().registerScheduler(testGroup);
     ...

   • 建立ZooKeeper叢集協調器

     String zkConnect = config.getString(DrillOnYarnConfig.ZK_CONNECT);
     String zkRoot = config.getString(DrillOnYarnConfig.ZK_ROOT);
     String clusterId = config.getString(DrillOnYarnConfig.CLUSTER_ID);

3. 啟動Dispatcher，主要啟動AMRMClientAsync、NMClientAsync、YarnClient

   ...
   yarn.start(new ResourceCallback(), new NodeCallback());
   String url = trackingUrl.replace("<port>", Integer.toString(httpPort));
   if (DrillOnYarnConfig.config().getBoolean(DrillOnYarnConfig.HTTP_ENABLE_SSL)) {
     url = url.replace("http:", "https:");
   }
   yarn.register(url);
   controller.started();
   ...

   ...
   resourceMgr = AMRMClientAsync.createAMRMClientAsync(pollPeriodMs, resourceCallback);
   resourceMgr.init(conf);
   resourceMgr.start();
   ...
   nodeMgr = NMClientAsync.createNMClientAsync(nodeCallback);
   nodeMgr.init(conf);
   nodeMgr.start();
   ...
   client = YarnClient.createYarnClient();
   client.init(conf);
   client.start();
   ...

4. 啟動dirll運維介面

   WebServer webServer = new WebServer(dispatcher);
   webServer.start();

5. 執行Dispatcher，主要是啟動一個執行緒，此執行緒會不斷的輪詢當前的任務佇列中的任務情況，比如啟動、停止、resize等型別的任務，然後執行相應的動作，拿啟動來說

   • 新增一個啟動任務，然後放入pendingTask佇列中

     if (state == State.LIVE) {
       adjustTasks(curTime);
       requestContainers();
     }

   • 向RM請求container：建立一個ContainerRequest

     ContainerRequest request = containerSpec.makeRequest();
     resourceMgr.addContainerRequest(containerSpec.makeRequest());
     return request;

   • ResourceCallback監聽container分配，然後啟動container

     private class ResourceCallback implements AMRMClientAsync.CallbackHandler {
         @Override
         public void onContainersAllocated(List<Container> containers) {
           controller.containersAllocated(containers);
         }
     }

     public void containerAllocated(EventContext context, Container container) {
       Task task = context.task;
       LOG.info(task.getLabel() + " - Received container: "
           + DoYUtil.describeContainer(container));
       context.group.dequeueAllocatingTask(task);
     
       // No matter what happens below, we don`t want to ask for this
       // container again. The RM async API is a bit bizarre in this
       // regard: it will keep asking for container over and over until
       // we tell it to stop.
     
       context.yarn.removeContainerRequest(task.containerRequest);
     
       // The container is need both in the normal and in the cancellation
       // path, so set it here.
     
       task.container = container;
       if (task.cancelled) {
         context.yarn.releaseContainer(container);
         taskStartFailed(context, Disposition.CANCELLED);
         return;
       }
       task.error = null;
       task.completionStatus = null;
       transition(context, LAUNCHING);
     
       // The pool that manages this task wants to know that we have
       // a container. The task manager may want to do some task-
       // specific setup.
     
       context.group.containerAllocated(context.task);
       context.getTaskManager().allocated(context);
     
       // Go ahead and launch a task in the container using the launch
       // specification provided by the task group (pool).
     
       try {
         context.yarn.launchContainer(container, task.getLaunchSpec());
         task.launchTime = System.currentTimeMillis();
       } catch (YarnFacadeException e) {
         LOG.error("Container launch failed: " + task.getContainerId(), e);
     
         // This may not be the right response. RM may still think
         // we have the container if the above is a local failure.
     
         task.error = e;
         context.group.containerReleased(task);
         task.container = null;
         taskStartFailed(context, Disposition.LAUNCH_FAILED);
       }
     }

   • NodeCallback監聽container啟動

     public class NodeCallback implements NMClientAsync.CallbackHandler {
         @Override
         public void onStartContainerError(ContainerId containerId, Throwable t) {
           controller.taskStartFailed(containerId, t);
         }
     
         @Override
         public void onContainerStarted(ContainerId containerId, Map<String, ByteBuffer> allServiceResponse) {
           controller.containerStarted(containerId);
         }
     
         @Override
         public void onContainerStatusReceived(ContainerId containerId, ContainerStatus containerStatus) {
         }
     
         @Override
         public void onGetContainerStatusError(ContainerId containerId, Throwable t) {
         }
     
         @Override
         public void onStopContainerError(ContainerId containerId, Throwable t) {
           controller.stopTaskFailed(containerId, t);
         }
     
         @Override
         public void onContainerStopped(ContainerId containerId) {
           controller.containerStopped(containerId);
         }
     }

2.3 fail over

Drill-on-YARN除了提供start、stop、resize功能外，還提供了fail over功能，當前某個drillbit掛掉後，Drill-on-YARN會嘗試再次啟動drillbit，目前重試的次數為2。此外，如果一個drillbit所在的節點頻繁掛掉，會被列入黑名單。

我們可以通過手動kill drillbit來模擬drillbit掛掉的情況，然後等待一會兒，可以看到，drillbit程式重新啟動了。下面我們看看，程式碼的執行流程

1. drillbit掛掉，container結束

private class ResourceCallback implements AMRMClientAsync.CallbackHandler {
    @Override
    public void onContainersCompleted(List<ContainerStatus> statuses) {
      controller.containersCompleted(statuses);
    }
}

1. retry task：重新將這個task加入pendingTasks，然後輪詢的執行緒檢測到pendingTasks不為空，執行啟動操作

protected void taskTerminated(EventContext context) {
    Task task = context.task;
    context.getTaskManager().completed(context);
    context.group.containerReleased(task);
    assert task.completionStatus != null;
    // container結束的狀態不是0，說明不是正常結束
    if (task.completionStatus.getExitStatus() == 0) {
      taskEnded(context, Disposition.COMPLETED);
      context.group.taskEnded(context.task);
    } else {
      taskEnded(context, Disposition.RUN_FAILED);
      retryTask(context);
    }
}

private void retryTask(EventContext context) {
    Task task = context.task;
    assert task.state == END;
    if (!context.controller.isLive() || !task.retryable()) {
      context.group.taskEnded(task);
      return;
    }
    if (task.tryCount > task.taskGroup.getMaxRetries()) {
      LOG.error(task.getLabel() + " - Too many retries: " + task.tryCount);
      task.disposition = Disposition.TOO_MANY_RETRIES;
      context.group.taskEnded(task);
      return;
    }
    LOG.info(task.getLabel() + " - Retrying task, try " + task.tryCount);
    context.group.taskRetried(task);
    task.reset();
    transition(context, START);
    context.group.enqueuePendingRequest(task);
}

3. 停止

除了前面詳情介紹的start命令外，Drill-on-YARN也提供了stop命令，其中stop分兩種：

1. 強制停止：直接呼叫yarn客戶端的killApplication api yarnClient.killApplication(appId);
2. 優雅停止：先清理所有的任務，包括pending、running的，然後呼叫yarn的api殺死容器，關閉controller，然後通知am執行結束

...
for (Task task : getStartingTasks()) {
  context.setTask(task);
  context.getState().cancel(context);
}
for (Task task : getActiveTasks()) {
  context.setTask(task);
  context.getState().cancel(context);
}
...

...
context.yarn.killContainer(task.container);
...

public void run() throws YarnFacadeException {
    ...
    boolean success = controller.waitForCompletion();
    ...
    ...
    finish(success, null);
    ...
  }

public boolean waitForCompletion() {
    start();
    synchronized (completionMutex) {
      try {
        completionMutex.wait();
      } catch (InterruptedException e) {
        
      }
    }
    return succeeded();
}

public void finish(boolean succeeded, String msg) throws YarnFacadeException {
    nodeMgr.stop();
    String appMsg = "Drill Cluster Shut-Down";
    FinalApplicationStatus status = FinalApplicationStatus.SUCCEEDED;
    if (!succeeded) {
      appMsg = "Drill Cluster Fatal Error - check logs";
      status = FinalApplicationStatus.FAILED;
    }
    if (msg != null) {
      appMsg = msg;
    }
    try {
      resourceMgr.unregisterApplicationMaster(status, appMsg, "");
    } catch (YarnException | IOException e) {
      throw new YarnFacadeException("Deregister AM failed", e);
    }
    resourceMgr.stop();
  }

4. resize

resize流程為：調整quantity（保留多少個container），之後輪詢執行緒會根據quantity，調整任務，執行resize操作

public int resize(int level) {
    int limit = quantity + state.getController().getFreeNodeCount() +maxExtraNodes;
    return super.resize( Math.min( limit, level ) );
}

5. 總結

總的來說，Drill-on-YARN分為兩大模組，drill-on-yarn.sh和drill-am.sh。drill-on-yarn.sh用於啟動ApplicationMaster，drill-am.sh用於向ResourceManager申請資源並啟動Drill叢集。其中Drill的啟動、停止、縮容、擴容，都被封裝為一個任務，在執行這些命令時，會構建一個任務，放入任務佇列中。有一個執行緒會一直輪詢此佇列，根據佇列中的任務執行不同的操作，從而達到啟動、停止、縮容、擴容Drill叢集的功能。此外，相比獨立部署，Drill-on-YARN提供的failover功能強化了Drill的穩定性。