Netflix Conductor源码分析--Client层源码分析

96
小程故事多
0.2 2018.12.10 14:09 字数 925

一、Client层总体介绍
在正式介绍Client层源码前,我们先来看一下如何在client端与server端通信,demo代码如下:

TaskClient taskClient = new TaskClient();
taskClient.setRootURI("http://localhost:8080/api/");        //Point this to the server API
int threadCount = 2;            //number of threads used to execute workers.  To avoid starvation, should be same or more than number of workers
Worker worker1 = new OrderWorker("order");
Worker worker2 = new PaymentWorker("payment");
//Create WorkflowTaskCoordinator
WorkflowTaskCoordinator.Builder builder = new WorkflowTaskCoordinator.Builder();
WorkflowTaskCoordinator coordinator = builder.withWorkers(worker1, worker2).withThreadCount(threadCount).withTaskClient(taskClient).build();
//Start for polling and execution of the tasks
coordinator.init();

代码说明:

1、第一步需要创建TaskClient类并设置server端的API URL路径以便客户端能够与服务端通信。

2、创建任务工作者Worker对象,具体的任务是由Worker来执行。

3、将Worker对象传入WorkerflowTaskCoordinator对象中,WorkerflowTaskCoordinator负责启动线程池来执行Worker任务,同时维护与server端的心跳以及最新任务数据的拉取操作。

通过阅读上述代码引出了几个类名称的解释:

  • WorkerflowTaskCoordinator:工作流的协调者,负责管理Task Worker的线程池以及和服务端的通信。
  • TaskClient:conductor的任务管理客户端类,负责从server端轮询任务以及更新任务状态等。
  • Builder:用于创建WorkerflowTaskCoordinator实例的建造类。

这三个类的类图如图1-1所示,从图中可以看到类的依赖、组合等关系。

图1-1

图1-1展示是Client层最核心的三个类的依赖关系,我们接下来的源码解析就是围绕这三个类来展开。

整个Client模块的包结构和关键类如图1-2所示:


图1-2

其中:

  • config包是关于Client的一些配置类
  • exceptions包是自定义的client异常类
  • http包是与服务端通信的基础类,包括基础基类ClientBase,还有元数据、负载、客户端任务,工作流等通信类
  • task包主要包括工作流协调者和工作流任务统计类
  • worker包主要包括Worker工作者接口类

二、Client层源码执行的全流程解析

我们拿文章 深入浅出Netflix Conductor使用 中介绍的案例来讲解源码流程(文章中包括了任务、工作流的DSL定义以及如何使用),流程图形表示如图1-3所示:

图1-3

这张图的含义非常简单,用户走下单流程到order模块,如果下单成功则走payment支付模块进行支付,如果下单失败则走失败模块进行重试等操作。

在Swagger界面上输入如下参数启动工作流,如图1-4


图1-4

启动的过程实际上是通过Swagger API接口调用server端的相关类,而client端则是通过拉取的方式来得到需要自己执行任务的通知和输入参数。

启动完工作流之后Client端的代码进入WorkerflowTaskCoordinator中的init方法,代码如下所示:

public synchronized void init() {
    if(threadCount == -1) {
        threadCount = workers.size();
    }
    logger.info("Initialized the worker with {} threads", threadCount);
    this.workerQueue = new LinkedBlockingQueue<Runnable>(workerQueueSize);
    AtomicInteger count = new AtomicInteger(0);
    this.executorService = new ThreadPoolExecutor(threadCount, threadCount,
            0L, TimeUnit.MILLISECONDS,
            workerQueue,
            (runnable) -> {
                Thread thread = new Thread(runnable);
                thread.setName(workerNamePrefix + count.getAndIncrement());
                return thread;
            });
    this.scheduledExecutorService = Executors.newScheduledThreadPool(workers.size());
  
   //定时轮询server状态策略,默认每隔1秒进行轮询,根据任务名获取当前任务信息
    workers.forEach(worker -> {
        scheduledExecutorService.scheduleWithFixedDelay(()->pollForTask(worker), worker.getPollingInterval(), worker.getPollingInterval(), TimeUnit.MILLISECONDS);
    });
}

代码说明:

这段代码通过JDK中的scheduledExecutorService.scheduleWithFixedDelay方法每隔一秒对server端进行轮询,轮询任务的方法是pollForTask,代码如下:

private void pollForTask(Worker worker) {
    if(eurekaClient != null && !eurekaClient.getInstanceRemoteStatus().equals(InstanceStatus.UP)) {
        logger.debug("Instance is NOT UP in discovery - will not poll");
        return;
    }
    if(worker.paused()) {
        WorkflowTaskMetrics.incrementTaskPausedCount(worker.getTaskDefName());
        logger.debug("Worker {} has been paused. Not polling anymore!", worker.getClass());
        return;
    }
    String domain = Optional.ofNullable(PropertyFactory.getString(worker.getTaskDefName(), DOMAIN, null))
            .orElse(PropertyFactory.getString(ALL_WORKERS, DOMAIN, null));
    logger.debug("Polling {}, domain={}, count = {} timeout = {} ms", worker.getTaskDefName(), domain, worker.getPollCount(), worker.getLongPollTimeoutInMS());
    List<Task> tasks = Collections.emptyList();
    try{
        // get the remaining capacity of worker queue to prevent queue full exception
        int realPollCount = Math.min(workerQueue.remainingCapacity(), worker.getPollCount());
        if (realPollCount <= 0) {
            logger.warn("All workers are busy, not polling. queue size = {}, max = {}", workerQueue.size(), workerQueueSize);
            return;
        }
        //获取当前客户端的任务名称
        String taskType = worker.getTaskDefName();
        //根据当前客户端的任务名称从server端的状态机获取是否有自己要执行的任务,如果有任务则获取执行,只能获取一次。
        tasks = getPollTimer(taskType)
                .record(() -> taskClient.batchPollTasksInDomain(taskType, domain, worker.getIdentity(), realPollCount, worker.getLongPollTimeoutInMS()));
        incrementTaskPollCount(taskType, tasks.size());
        logger.debug("Polled {}, domain {}, received {} tasks in worker - {}", worker.getTaskDefName(), domain, tasks.size(), worker.getIdentity());
    } catch (Exception e) {
        WorkflowTaskMetrics.incrementTaskPollErrorCount(worker.getTaskDefName(), e);
        logger.error("Error when polling for tasks", e);
    }
    //根据获取的任务列表,以线程的方式启动执行任务
    for (Task task : tasks) {
        try {
            executorService.submit(() -> {
                try {
                    logger.debug("Executing task {}, taskId - {} in worker - {}", task.getTaskDefName(), task.getTaskId(), worker.getIdentity());
                    //这步就是执行用户自定义的任务逻辑
                    execute(worker, task);
                } catch (Throwable t) {
                    //执行失败,置任务状态为失败,并将失败结果返回到server端
                    task.setStatus(Task.Status.FAILED);
                    TaskResult result = new TaskResult(task);
                    handleException(t, result, worker, task);
                }
            });
        } catch (RejectedExecutionException e) {
            WorkflowTaskMetrics.incrementTaskExecutionQueueFullCount(worker.getTaskDefName());
            logger.error("Execution queue is full, returning task: {}", task.getTaskId(), e);
            returnTask(worker, task);
        }
    }
}

代码说明:

每隔一秒从服务端的(tasks/poll/batch/{taskType})获取当前需要执行的任务列表,任务只能获取一次不能重新获取。然后将任务通过异步线程的方式启动执行,每一个任务都是由用户自定义的逻辑实现,任务的返回值被封装到了TaskResult类中,execute方法的内容如下所示:

private void execute(Worker worker, Task task) {
    String taskType = task.getTaskDefName();
    try {
        if(!worker.preAck(task)) {
            logger.debug("Worker decided not to ack the task {}, taskId = {}", taskType, task.getTaskId());
            return;
        }
        if (!taskClient.ack(task.getTaskId(), worker.getIdentity())) {
            WorkflowTaskMetrics.incrementTaskAckFailedCount(worker.getTaskDefName());
            logger.error("Ack failed for {}, taskId = {}", taskType, task.getTaskId());
            returnTask(worker, task);
            return;
        }
    } catch (Exception e) {
        logger.error(String.format("ack exception for task %s, taskId = %s in worker - %s", task.getTaskDefName(), task.getTaskId(), worker.getIdentity()), e);
        WorkflowTaskMetrics.incrementTaskAckErrorCount(worker.getTaskDefName(), e);
        returnTask(worker, task);
        return;
    }
    com.google.common.base.Stopwatch stopwatch = com.google.common.base.Stopwatch.createStarted();
    TaskResult result = null;
    try {
        //前面大部分都是做监控和统计功能的,在这里不细说
        //这段代码是真正执行用户Task任务的代码,执行完后返回值被封装为TaskResult对象
        logger.debug("Executing task {} in worker {} at {}", task, worker.getClass().getSimpleName(), worker.getIdentity());
        result = worker.execute(task);
        result.setWorkflowInstanceId(task.getWorkflowInstanceId());
        result.setTaskId(task.getTaskId());
        result.setWorkerId(worker.getIdentity());
    } catch (Exception e) {
        logger.error("Unable to execute task {}", task, e);
        if (result == null) {
            task.setStatus(Task.Status.FAILED);
            result = new TaskResult(task);
        }
        handleException(e, result, worker, task);
    } finally {
        stopwatch.stop();
        WorkflowTaskMetrics.getExecutionTimer(worker.getTaskDefName())
                .record(stopwatch.elapsed(TimeUnit.MILLISECONDS), TimeUnit.MILLISECONDS);
    }
    logger.debug("Task {} executed by worker {} at {} with status {}", task.getTaskId(), worker.getClass().getSimpleName(), worker.getIdentity(), task.getStatus());
    //更新任务状态,成功或者失败
    updateWithRetry(updateRetryCount, task, result, worker);
}

代码说明:

通过worker.execute方法执行用户定义的任务逻辑,不管是否成功都执行updatewithRetry方法更新server端的任务状态和任务执行返回结果。

访问的URL是/api/tasks。

三、完整流程时序图


image.png
日记本
Web note ad 1