一个庞大的分布式系统,各个组件间是如何协调工作的?组件是如何解耦的?线程运行如何更高效,减少阻塞带来的低效问题?本节将对 Yarn 的服务库和事件库进行介绍,看看 Yarn 是如何解决这些问题的。

一、服务库

一)简介

对于生命周期较长的对象,Yarn 采用基于服务的模型对其进行管理,有以下几个特点:

二)源码简析

源代码地址在 hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/serviceService 接口中。
其中定义了服务的四个状态,以及需要实现的状态转换、获取信息、注册等方法。

public interface Service extends Closeable {
  public enum STATE {
    NOTINITED(0, "NOTINITED"),
    INITED(1, "INITED"),
    STARTED(2, "STARTED"),
    STOPPED(3, "STOPPED");
  }
  void init(Configuration config);
  void start();
  void stop();
  void close() throws IOException;
  void registerServiceListener(ServiceStateChangeListener listener);
  // ......

抽象类 AbstractService 实现了 Service 接口,提供了基础的 Service 实现,非组合服务直接继承这个抽象类再开发即可。

public abstract class AbstractService implements Service {
  // 以 start 实现为例,执行后会触发其他的操作
  public void start() {
    if (isInState(STATE.STARTED)) {
      return;
    }
    //enter the started state
    synchronized (stateChangeLock) {
      if (stateModel.enterState(STATE.STARTED) != STATE.STARTED) {
        try {
          startTime = System.currentTimeMillis();
          serviceStart();
          if (isInState(STATE.STARTED)) {
            //if the service started (and isn't now in a later state), notify
            if (LOG.isDebugEnabled()) {
              LOG.debug("Service " + getName() + " is started");
            }
            notifyListeners();
          }
        } catch (Exception e) {
          noteFailure(e);
          ServiceOperations.stopQuietly(LOG, this);
          throw ServiceStateException.convert(e);
        }
      }
    }
  }
  // ......

对于组合类的服务如 ResourceManager、NodeManager 等,需要继承 CompositeService。其中会有对组合服务的逻辑处理。

  public List<Service> getServices() {
    synchronized (serviceList) {
      return new ArrayList<Service>(serviceList);
    }
  }
  protected void addService(Service service) {
    if (LOG.isDebugEnabled()) {
      LOG.debug("Adding service " + service.getName());
    }
    synchronized (serviceList) {
      serviceList.add(service);
    }
  }

二、事件库

传统函数式调用的问题:
整个执行过程是串行、同步进行的。调用另一个函数的时候,需要等待函数执行完毕,才会继续往下走。示意图如下:
【深入浅出 Yarn 架构与实现】2-3 Yarn 基础库 - 服务库与事件库

为了解决函数式调用的问题,可使用「事件驱动」的编程模型。

示意图如下:
【深入浅出 Yarn 架构与实现】2-3 Yarn 基础库 - 服务库与事件库

通过以上的方式,可以使程序有低耦合高内聚的特点,各个模块仅需完成各自的功能,同时提高了执行效率,把拆分的操作通过事件的方式发送出去即可。

三、服务库和事件库使用案例

本节将实现一个简化版的 MapReduce ApplicationMaster,帮助了解 service 和 event 的使用方法。
与 MR 类似,一个 job 将被分为多个 task 执行。因此涉及 job 和 task 两种对象的事件。并有一个 AsyncDispatcher 处理调度。
案例已上传至 github,有帮助可以点个 ⭐️
https://github.com/Simon-Ace/hadoop-yarn-study-demo/tree/master/service-event-demo

一)事件部分

参考 hadoop 源码中 Task 和 Job Event 的实现,进行一些简化。
1、task

public enum TaskEventType {
  //Producer:Client, Job
  T_KILL,
  //Producer:Job
  T_SCHEDULE
}
public class TaskEvent extends AbstractEvent<TaskEventType> {
  private String taskID;
  public TaskEvent(String taskID, TaskEventType type) {
    super(type);
    this.taskID = taskID;
  }
  public String getTaskID() {
    return taskID;
  }
}

2、job

public enum JobEventType {
  //Producer:Client
  JOB_KILL,
  //Producer:MRAppMaster
  JOB_INIT
}
public class JobEvent extends AbstractEvent<JobEventType> {
    private String jobID;
    public JobEvent(String jobID, JobEventType type) {
        super(type);
        this.jobID = jobID;
    }
    public String getJobId() {
        return jobID;
    }
}

二)事件调度器

import com.shuofxz.event.JobEvent;
import com.shuofxz.event.JobEventType;
import com.shuofxz.event.TaskEvent;
import com.shuofxz.event.TaskEventType;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.service.CompositeService;
import org.apache.hadoop.service.Service;
import org.apache.hadoop.yarn.event.AsyncDispatcher;
import org.apache.hadoop.yarn.event.Dispatcher;
import org.apache.hadoop.yarn.event.EventHandler;
@SuppressWarnings("unchecked")
public class MyMRAppMaster extends CompositeService {
    private Dispatcher dispatcher;  // AsyncDispatcher
    private String jobID;
    private int taskNumber;         // 一个 job 包含的 task 数
    private String[] taskIDs;
    public MyMRAppMaster(String name, String jobID, int taskNumber) {
        super(name);
        this.jobID = jobID;
        this.taskNumber = taskNumber;
        taskIDs = new String[taskNumber];
        for (int i = 0; i < taskNumber; i++) {
            taskIDs[i] = this.jobID + "_task_" + i;
        }
    }
    public void serviceInit(Configuration conf) throws Exception {
        dispatcher = new AsyncDispatcher();
        dispatcher.register(JobEventType.class, new JobEventDispatcher()); // register a job
        dispatcher.register(TaskEventType.class, new TaskEventDispatcher()); // register a task
        addService((Service) dispatcher);
        super.serviceInit(conf);
    }
    public void serviceStart() throws Exception {
        super.serviceStart();
    }
    public Dispatcher getDispatcher() {
        return dispatcher;
    }
    private class JobEventDispatcher implements EventHandler<JobEvent> {
        public void handle(JobEvent event) {
            if (event.getType() == JobEventType.JOB_KILL) {
                System.out.println("Receive JOB_KILL event, killing all the tasks");
                for (int i = 0; i < taskNumber; i++) {
                    dispatcher.getEventHandler().handle(new TaskEvent(taskIDs[i], TaskEventType.T_KILL));
                }
            } else if (event.getType() == JobEventType.JOB_INIT) {
                System.out.println("Receive JOB_INIT event, scheduling tasks");
                for (int i = 0; i < taskNumber; i++) {
                    dispatcher.getEventHandler().handle(new TaskEvent(taskIDs[i], TaskEventType.T_SCHEDULE));
                }
            }
        }
    }
    private class TaskEventDispatcher implements EventHandler<TaskEvent> {
        public void handle(TaskEvent event) {
            if (event.getType() == TaskEventType.T_KILL) {
                System.out.println("Receive T_KILL event of task id " + event.getTaskID());
            } else if (event.getType() == TaskEventType.T_SCHEDULE) {
                System.out.println("Receive T_SCHEDULE event of task id " + event.getTaskID());
            }
        }
    }
}

三)测试程序

public class MyMRAppMasterTest {
    public static void main(String[] args) {
        String jobID = "job_20221011_99";
        MyMRAppMaster appMaster = new MyMRAppMaster("My MRAppMaster Test", jobID, 10);
        YarnConfiguration conf = new YarnConfiguration(new Configuration());
        try {
            appMaster.serviceInit(conf);
            appMaster.serviceStart();
        } catch (Exception e) {
            e.printStackTrace();
        }
        appMaster.getDispatcher().getEventHandler().handle(new JobEvent(jobID, JobEventType.JOB_KILL));
        appMaster.getDispatcher().getEventHandler().handle(new JobEvent(jobID, JobEventType.JOB_INIT));
    }
}

输出结果:

Receive JOB_KILL event, killing all the tasks
Receive JOB_INIT event, scheduling tasks
Receive T_KILL event of task id job_20150723_11_task_0
Receive T_KILL event of task id job_20150723_11_task_1
Receive T_KILL event of task id job_20150723_11_task_2
Receive T_KILL event of task id job_20150723_11_task_3
Receive T_KILL event of task id job_20150723_11_task_4
Receive T_KILL event of task id job_20150723_11_task_5
Receive T_KILL event of task id job_20150723_11_task_6
Receive T_KILL event of task id job_20150723_11_task_7
Receive T_KILL event of task id job_20150723_11_task_8
Receive T_KILL event of task id job_20150723_11_task_9
Receive T_SCHEDULE event of task id job_20150723_11_task_0
Receive T_SCHEDULE event of task id job_20150723_11_task_1
Receive T_SCHEDULE event of task id job_20150723_11_task_2
Receive T_SCHEDULE event of task id job_20150723_11_task_3
Receive T_SCHEDULE event of task id job_20150723_11_task_4
Receive T_SCHEDULE event of task id job_20150723_11_task_5
Receive T_SCHEDULE event of task id job_20150723_11_task_6
Receive T_SCHEDULE event of task id job_20150723_11_task_7
Receive T_SCHEDULE event of task id job_20150723_11_task_8
Receive T_SCHEDULE event of task id job_20150723_11_task_9

四、总结

本节介绍了 Yarn 的服务和事件库。
服务库规范了生命周期较长的服务型对象,定义了服务的四种状态、启停注册等要实现的方法,给出了单一类型和组合类型服务的基本实现。
事件库的使用,解决了原始函数型调用的高耦合、阻塞低效等问题。可将一个大任务拆分成多个小任务,小任务变成不同的事件来触发处理。每一个事件处理器处理一种事件,并有一个中央异步调度器管理事件的收集和分发。
最后用一个简化的 MR ApplicationMaster 将事件库和服务库进行结合,更深体会如何在项目中将其结合使用。
学习过程中,写一个 demo 能更好的帮助你理解知识。

参考文章:
《Hadoop 技术内幕 - 深入解析 Yarn 结构设计与实现原理》3.4 节

发表回复