DolphinScheduler容错源码分析之Worker

@PostConstruct
public void run() {
	this.workerrpcServer.start();
	this.workerRpcClient.start();
	this.taskPluginManager.loadPlugin();
	this.workerReGIStryClient.setRegistryStoppable(this);
	this.workerRegistryClient.start();
	this.workerManagerThread.start();
	this.messageRetryRunner.start();
	
	Runtime.getRuntime().addShutdownHook(new Thread(() -> {
		if (!ServerLifeCycleManager.isStopped()) {
			close("WorkerServer shutdown hook");
		}
	}));
}

private void registry() {
	WorkerHeartBeat workerHeartBeat = workerHeartBeatTask.getHeartBeat();
	String workerZKPath = workerConfig.getWorkerRegistryPath();
	// remove before persist
	registryClient.remove(workerZKPath);
	registryClient.persistEphemeral(workerZKPath, JSONUtils.tojsonString(workerHeartBeat));
	log.info("Worker node: {} registry to ZK {} successfully", workerConfig.getWorkerAddress(), workerZKPath);
	while (!registryClient.checkNodeExists(workerConfig.getWorkerAddress(), NodeType.WORKER)) {
		ThreadUtils.sleep(SLEEP_TIME_MILLIS);
	}
	// sleep 1s, waiting master failover remove
	ThreadUtils.sleep(Constants.SLEEP_TIME_MILLIS);
	workerHeartBeatTask.start();
	log.info("Worker node: {} registry finished", workerConfig.getWorkerAddress());
}

// /nodes/worker/+ip:listenPortworkerConfig.setWorkerRegistryPath(REGISTRY_DOLPHINSCHEDULER_WORKERS + "/" + workerConfig.getWorkerAddress());

@Override
public WorkerHeartBeat getHeartBeat() {
	double loadAverage = OSUtils.loadAverage();
	double cpuUsage = OSUtils.cpuUsage();
	int maxCpuLoadAvg = workerConfig.getMaxCpuLoadAvg();
	double reservedMemory = workerConfig.getReservedMemory();
	double availablePhysicalMemorySize = OSUtils.availablePhysicalMemorySize();
	int execThreads = workerConfig.getExecThreads();
	int workerWaitingTaskCount = this.workerWaitingTaskCount.get();
	int serverStatus = getServerStatus(loadAverage, maxCpuLoadAvg, availablePhysicalMemorySize, reservedMemory,
			execThreads, workerWaitingTaskCount);
	return WorkerHeartBeat.builder()
			.startupTime(ServerLifeCycleManager.getServerStartupTime())
			.reportTime(System.currentTimeMillis())
			.cpuUsage(cpuUsage)
			.loadAverage(loadAverage)
			.availablePhysicalMemorySize(availablePhysicalMemorySize)
			.maxCpuloadAvg(maxCpuLoadAvg)
			.memoryUsage(OSUtils.memoryUsage())
			.reservedMemory(reservedMemory)
			.diskAvailable(OSUtils.diskAvailable())
			.processId(processId)
			.workerHostWeight(workerConfig.getHostWeight())
			.workerWaitingTaskCount(this.workerWaitingTaskCount.get())
			.workerExecThreadCount(workerConfig.getExecThreads())
			.serverStatus(serverStatus)
			.build();
}

//node
registryClient.subscribe(REGISTRY_DOLPHINSCHEDULER_NODE, new MasterRegistryDataListener());

@Override
public void notify(Event event) {
	final String path = event.path();
	if (Strings.isNullOrEmpty(path)) {
		return;
	}
	//monitor master
	if (path.startsWith(REGISTRY_DOLPHINSCHEDULER_MASTERS + Constants.SINGLE_SLASH)) {
		handleMasterEvent(event);
	} else if (path.startsWith(REGISTRY_DOLPHINSCHEDULER_WORKERS + Constants.SINGLE_SLASH)) {
		//monitor worker
		handleWorkerEvent(event);
	}
}

private void handleWorkerEvent(Event event) {
	final String path = event.path();
	switch (event.type()) {
		case ADD:
			logger.info("worker node added : {}", path);
			break;
		case REMOVE:
			logger.info("worker node deleted : {}", path);
			masterRegistryClient.removeWorkerNodePath(path, NodeType.WORKER, true);
			break;
		default:
			break;
	}
}

public void removeWorkerNodePath(String path, NodeType nodeType, boolean failover) {
	logger.info("{} node deleted : {}", nodeType, path);
	try {
                //获取节点信息
		String serverHost = null;
		if (!StringUtils.isEmpty(path)) {
			serverHost = registryClient.getHostByEventDataPath(path);
			if (StringUtils.isEmpty(serverHost)) {
				logger.error("server down error: unknown path: {}", path);
				return;
			}
			if (!registryClient.exists(path)) {
				logger.info("path: {} not exists", path);
			}
		}
		// failover server
		if (failover) {
			failoverService.failoverServerWhenDown(serverHost, nodeType);
		}
	} catch (Exception e) {
		logger.error("{} server failover failed", nodeType, e);
	}
}

public void failoverWorker(@NonNull String workerHost) {
	LOGGER.info("Worker[{}] failover starting", workerHost);
	final StopWatch failoverTimeCost = StopWatch.createStarted();
	//获取需要容错worker节点的启动时间，用于后续判断worker节点是否还在下线状态，或者是否已经重新启动
	// we query the task instance from cache, so that we can directly update the cache
	final Optional<Date> needFailoverWorkerStartTime =
			getServerStartupTime(registryClient.getServerList(NodeType.WORKER), workerHost);
	//根据异常的worker的信息查询需要容错的任务实例，获取只属于当前master节点需要容错的任务实例信息，这里也是和master不同的，并且容错没加锁的原因。
	final List<TaskInstance> needFailoverTaskInstanceList = getNeedFailoverTaskInstance(workerHost);
	if (CollectionUtils.isEmpty(needFailoverTaskInstanceList)) {
		LOGGER.info("Worker[{}] failover finished there are no taskInstance need to failover", workerHost);
		return;
	}
	LOGGER.info(
			"Worker[{}] failover there are {} taskInstance may need to failover, will do a deep check, taskInstanceIds: {}",
			workerHost,
			needFailoverTaskInstanceList.size(),
			needFailoverTaskInstanceList.stream().map(TaskInstance::getId).collect(Collectors.toList()));
	final Map<Integer, ProcessInstance> processInstanceCacheMap = new HashMap<>();
	for (TaskInstance taskInstance : needFailoverTaskInstanceList) {
		LoggerUtils.setWorkflowAndTaskInstanceIDMDC(taskInstance.getProcessInstanceId(), taskInstance.getId());
		try {
			ProcessInstance processInstance = processInstanceCacheMap.computeIfAbsent(
					taskInstance.getProcessInstanceId(), k -> {
						WorkflowExecuteRunnable workflowExecuteRunnable = cacheManager.getByProcessInstanceId(
								taskInstance.getProcessInstanceId());
						if (workflowExecuteRunnable == null) {
							return null;
						}
						return workflowExecuteRunnable.getProcessInstance();
					});
			//这里注意的是需要容错的任务是在worker重新启动之前的任务，之后worker异常重启后分配的新任务不要容错
			if (!checkTaskInstanceNeedFailover(needFailoverWorkerStartTime, processInstance, taskInstance)) {
				LOGGER.info("Worker[{}] the current taskInstance doesn't need to failover", workerHost);
				continue;
			}
			LOGGER.info(
					"Worker[{}] failover: begin to failover taskInstance, will set the status to NEED_FAULT_TOLERANCE",
					workerHost);
			failoverTaskInstance(processInstance, taskInstance);
			LOGGER.info("Worker[{}] failover: Finish failover taskInstance", workerHost);
		} catch (Exception ex) {
			LOGGER.info("Worker[{}] failover taskInstance occur exception", workerHost, ex);
		} finally {
			LoggerUtils.removeWorkflowAndTaskInstanceIdMDC();
		}
	}
	failoverTimeCost.stop();
	LOGGER.info("Worker[{}] failover finished, useTime:{}ms",
			workerHost,
			failoverTimeCost.getTime(TimeUnit.MILLISECONDS));
}

private void failoverTaskInstance(@NonNull ProcessInstance processInstance, @NonNull TaskInstance taskInstance) {
	TaskMetrics.incTaskInstanceByState("failover");
	boolean isMasterTask = TaskProcessorFactory.isMasterTask(taskInstance.getTaskType());
	taskInstance.setProcessInstance(processInstance);
	if (!isMasterTask) {
		LOGGER.info("The failover taskInstance is not master task");
		TaskExecutionContext taskExecutionContext = TaskExecutionContextBuilder.get()
				.buildTaskInstanceRelatedInfo(taskInstance)
				.buildProcessInstanceRelatedInfo(processInstance)
				.buildProcessDefinitionRelatedInfo(processInstance.getProcessDefinition())
				.create();
		if (masterConfig.isKillYarnJobWhenTaskFailover()) {
			// only kill yarn job if exists , the local thread has exited
			LOGGER.info("TaskInstance failover begin kill the task related yarn job");
			ProcessUtils.killYarnJob(loGClient, taskExecutionContext);
		}
	} else {
		LOGGER.info("The failover taskInstance is a master task");
	}
	taskInstance.setState(TaskExecutionStatus.NEED_FAULT_TOLERANCE);
	taskInstance.setFlag(Flag.NO);
	processService.saveTaskInstance(taskInstance);
        //提交event
	TaskStateEvent stateEvent = TaskStateEvent.builder()
			.processInstanceId(processInstance.getId())
			.taskInstanceId(taskInstance.getId())
			.status(TaskExecutionStatus.NEED_FAULT_TOLERANCE)
			.type(StateEventType.TASK_STATE_CHANGE)
			.build();
	workflowExecuteThreadPool.submitStateEvent(stateEvent);
}

public void submitStateEvent(StateEvent stateEvent) {
	WorkflowExecuteRunnable workflowExecuteThread =
			processInstanceExecCacheManager.getByProcessInstanceId(stateEvent.getProcessInstanceId());
	if (workflowExecuteThread == null) {
		logger.warn("Submit state event error, cannot from workflowExecuteThread from cache manager, stateEvent:{}",
				stateEvent);
		return;
	}
	workflowExecuteThread.addStateEvent(stateEvent);
	logger.info("Submit state event success, stateEvent: {}", stateEvent);
}

private final ConcurrentLinkedQueue<StateEvent> stateEvents = new ConcurrentLinkedQueue<>();

public void handleEvents() {
	if (!isStart()) {
		logger.info(
				"The workflow instance is not started, will not handle its state event, current state event size: {}",
				stateEvents);
		return;
	}
	StateEvent stateEvent = null;
	while (!this.stateEvents.isEmpty()) {
		try {
			stateEvent = this.stateEvents.peek();
			LoggerUtils.setWorkflowAndTaskInstanceIDMDC(stateEvent.getProcessInstanceId(),
					stateEvent.getTaskInstanceId());
			// if state handle success then will remove this state, otherwise will retry this state next time.
			// The state should always handle success except database error.
			checkProcessInstance(stateEvent);
			StateEventHandler stateEventHandler =
					StateEventHandlerManager.getStateEventHandler(stateEvent.getType())
							.orElseThrow(() -> new StateEventHandleError(
									"Cannot find handler for the given state event"));
			logger.info("Begin to handle state event, {}", stateEvent);
			if (stateEventHandler.handleStateEvent(this, stateEvent)) {
				this.stateEvents.remove(stateEvent);
			}
		} catch (StateEventHandleError stateEventHandleError) {
			logger.error("State event handle error, will remove this event: {}", stateEvent, stateEventHandleError);
			this.stateEvents.remove(stateEvent);
			ThreadUtils.sleep(Constants.SLEEP_TIME_MILLIS);
		} catch (StateEventHandleException stateEventHandleException) {
			logger.error("State event handle error, will retry this event: {}",
					stateEvent,
					stateEventHandleException);
			ThreadUtils.sleep(Constants.SLEEP_TIME_MILLIS);
		} catch (Exception e) {
			// we catch the exception here, since if the state event handle failed, the state event will still keep
			// in the stateEvents queue.
			logger.error("State event handle error, get a unknown exception, will retry this event: {}",
					stateEvent,
					e);
			ThreadUtils.sleep(Constants.SLEEP_TIME_MILLIS);
		} finally {
			LoggerUtils.removeWorkflowAndTaskInstanceIdMDC();
		}
	}
}

 if (task.getState().isFinished()) {
		if (completeTaskMap.containsKey(task.getTaskCode())
				&& completeTaskMap.get(task.getTaskCode()) == task.getId()) {
			logger.warn("The task instance is already complete, stateEvent: {}", stateEvent);
			return true;
		}
		workflowExecuteRunnable.taskFinished(task);
		if (task.getTaskGroupId() > 0) {
			logger.info("The task instance need to release task Group: {}", task.getTaskGroupId());
			workflowExecuteRunnable.releaseTaskGroup(task);
		}
		return true;
	}

public boolean isFinished() {
	return isSuccess() || isKill() || isFailure() || isPause();
}
public boolean isFailure() {
	return this == TaskExecutionStatus.FAILURE || this == NEED_FAULT_TOLERANCE;
}

} else if (taskInstance.taskCanRetry() && !processInstance.getState().isReadyStop()) {
			// retry task
			logger.info("Retry taskInstance taskInstance state: {}", taskInstance.getState());
			retryTaskInstance(taskInstance);
}
//判断了是否容错的状态，前面对其已经进行了更新
public boolean taskCanRetry() {
	if (this.isSubProcess()) {
		return false;
	}
	if (this.getState() == TaskExecutionStatus.NEED_FAULT_TOLERANCE) {
		return true;
	}
	return this.getState() == TaskExecutionStatus.FAILURE && (this.getRetryTimes() < this.getMaxRetryTimes());
}

private void retryTaskInstance(TaskInstance taskInstance) throws StateEventHandleException {
	if (!taskInstance.taskCanRetry()) {
		return;
	}
	TaskInstance newTaskInstance = cloneRetryTaskInstance(taskInstance);
	if (newTaskInstance == null) {
		logger.error("Retry task fail because new taskInstance is null, task code:{}, task id:{}",
				taskInstance.getTaskCode(),
				taskInstance.getId());
		return;
	}
	waitToRetryTaskInstanceMap.put(newTaskInstance.getTaskCode(), newTaskInstance);
	if (!taskInstance.retryTaskIntervalOverTime()) {
		logger.info(
				"Failure task will be submitted, process id: {}, task instance code: {}, state: {}, retry times: {} / {}, interval: {}",
				processInstance.getId(), newTaskInstance.getTaskCode(),
				newTaskInstance.getState(), newTaskInstance.getRetryTimes(), newTaskInstance.getMaxRetryTimes(),
				newTaskInstance.getRetryInterval());
		stateWheelExecuteThread.addTask4TimeoutCheck(processInstance, newTaskInstance);
		stateWheelExecuteThread.addTask4RetryCheck(processInstance, newTaskInstance);
	} else {
		addTaskToStandByList(newTaskInstance);
		submitStandByTask();
		waitToRetryTaskInstanceMap.remove(newTaskInstance.getTaskCode());
	}
}

addTaskToStandByList(newTaskInstance);
submitStandByTask();