聊聊PowerJob的MapProcessor

序

本文主要研究一下PowerJob的MapProcessor

MapProcessor

tech/powerjob/worker/core/processor/sdk/MapProcessor.java

public interface MapProcessor extends BasicProcessor {

    Logger log = LoggerFactory.getLogger(MapProcessor.class);

    int RECOMMEND_BATCH_SIZE = 200;

    /**
     * 分发子任务
     * @param taskList 子任务，再次执行时可通过 TaskContext#getSubTask 获取
     * @param taskName 子任务名称，即子任务处理器中 TaskContext#getTaskName 获取到的值
     * @throws PowerJobCheckedException map 失败将抛出异常
     */
    default void map(List<?> taskList, String taskName) throws PowerJobCheckedException {

        if (CollectionUtils.isEmpty(taskList)) {
            return;
        }

        TaskDO task = ThreadLocalStore.getTask();
        WorkerRuntime workerRuntime = ThreadLocalStore.getRuntimeMeta();

        if (taskList.size() > RECOMMEND_BATCH_SIZE) {
            log.warn("[Map-{}] map task size is too large, network maybe overload... please try to split the tasks.", task.getInstanceId());
        }

        // 修复 map 任务命名和根任务名或者最终任务名称一致导致的问题（无限生成子任务或者直接失败）
        if (TaskConstant.ROOT_TASK_NAME.equals(taskName) || TaskConstant.LAST_TASK_NAME.equals(taskName)) {
            log.warn("[Map-{}] illegal map task name : {}! please do not use 'OMS_ROOT_TASK' or 'OMS_LAST_TASK' as map task name. as a precaution, it will be renamed 'X-{}' automatically." ,task.getInstanceId() ,taskName , taskName);
            taskName ="X-"+taskName;
        }

        // 1. 构造请求
        ProcessorMapTaskRequest req = new ProcessorMapTaskRequest(task, taskList, taskName);

        // 2. 可靠发送请求（任务不允许丢失，需要使用 ask 方法，失败抛异常）
        boolean requestSucceed = TransportUtils.reliableMapTask(req, task.getAddress(), workerRuntime);

        if (requestSucceed) {
            log.info("[Map-{}] map task[name={},num={}] successfully!", task.getInstanceId(), taskName, taskList.size());
        }else {
            throw new PowerJobCheckedException("map failed for task: " + taskName);
        }
    }

    /**
     * 是否为根任务
     * @return true -> 根任务 / false -> 非根任务
     */
    default boolean isRootTask() {
        TaskDO task = ThreadLocalStore.getTask();
        return TaskConstant.ROOT_TASK_NAME.equals(task.getTaskName());
    }
}

MapProcessor接口继承了BasicProcessor，它提供了默认的map方法用于分发子任务，它主要是构造了ProcessorMapTaskRequest，通过TransportUtils.reliableMapTask发送请求；它还提供了isRootTask方法用于判断当前任务是不是根任务

ProcessorMapTaskRequest

tech/powerjob/worker/pojo/request/ProcessorMapTaskRequest.java

@Getter
@NoArgsConstructor
public class ProcessorMapTaskRequest implements PowerSerializable {

    private Long instanceId;
    private Long subInstanceId;

    private String taskName;
    private List<SubTask> subTasks;

    @Getter
    @NoArgsConstructor
    @AllArgsConstructor
    public static class SubTask {
        private String taskId;
        private byte[] taskContent;
    }

    public ProcessorMapTaskRequest(TaskDO parentTask, List<?> subTaskList, String taskName) {

        this.instanceId = parentTask.getInstanceId();
        this.subInstanceId = parentTask.getSubInstanceId();
        this.taskName = taskName;
        this.subTasks = Lists.newLinkedList();

        subTaskList.forEach(subTask -> {
            // 同一个 Task 内部可能多次 Map，因此还是要确保线程级别的唯一
            String subTaskId = parentTask.getTaskId() + "." + ThreadLocalStore.getTaskIDAddr().getAndIncrement();
            // 写入类名，方便反序列化
            subTasks.add(new SubTask(subTaskId, SerializerUtils.serialize(subTask)));
        });
    }
}

ProcessorMapTaskRequest的构造器将subTaskList转换为一系列的SubTask，它使用SerializerUtils.serialize序列化(Kryo)了用户定义的subTask

reliableMapTask

tech/powerjob/worker/common/utils/TransportUtils.java

    public static boolean reliableMapTask(ProcessorMapTaskRequest req, String address, WorkerRuntime workerRuntime) throws PowerJobCheckedException {
        try {
            return reliableAsk(ServerType.WORKER, WTT_PATH, WTT_HANDLER_MAP_TASK, address, req, workerRuntime.getTransporter()).isSuccess();
        } catch (Throwable throwable) {
            throw new PowerJobCheckedException(throwable);
        }
    }

    private static AskResponse reliableAsk(ServerType t, String rootPath, String handlerPath, String address, PowerSerializable req, Transporter transporter) throws Exception {
        final URL url = easyBuildUrl(t, rootPath, handlerPath, address);
        final CompletionStage<AskResponse> completionStage = transporter.ask(url, req, AskResponse.class);
        return completionStage
                .toCompletableFuture()
                .get(RemoteConstant.DEFAULT_TIMEOUT_MS, TimeUnit.MILLISECONDS);
    }    

reliableMapTask方法通过reliableAsk往taskTracker/mapTask接口发送请求，默认是5s超时

onReceiveProcessorMapTaskRequest

tech/powerjob/worker/actors/TaskTrackerActor.java

    /**
     * 子任务 map 处理器
     */
    @Handler(path = WTT_HANDLER_MAP_TASK)
    public AskResponse onReceiveProcessorMapTaskRequest(ProcessorMapTaskRequest req) {

        HeavyTaskTracker taskTracker = HeavyTaskTrackerManager.getTaskTracker(req.getInstanceId());
        if (taskTracker == null) {
            log.warn("[TaskTrackerActor] receive ProcessorMapTaskRequest({}) but system can't find TaskTracker.", req);
            return null;
        }

        boolean success = false;
        List<TaskDO> subTaskList = Lists.newLinkedList();

        try {

            req.getSubTasks().forEach(originSubTask -> {
                TaskDO subTask = new TaskDO();

                subTask.setTaskName(req.getTaskName());
                subTask.setSubInstanceId(req.getSubInstanceId());

                subTask.setTaskId(originSubTask.getTaskId());
                subTask.setTaskContent(originSubTask.getTaskContent());

                subTaskList.add(subTask);
            });

            success = taskTracker.submitTask(subTaskList);
        }catch (Exception e) {
            log.warn("[TaskTrackerActor] process map task(instanceId={}) failed.", req.getInstanceId(), e);
        }

        AskResponse response = new AskResponse();
        response.setSuccess(success);
        return response;
    }

TaskTrackerActor提供了onReceiveProcessorMapTaskRequest方法处理ProcessorMapTaskRequest，它将入参的subTasks转换为一系列的TaskDO，然后通过taskTracker.submitTask提交

submitTask

tech/powerjob/worker/core/tracker/task/heavy/HeavyTaskTracker.java

    /**
     * 提交Task任务(MapReduce的Map，Broadcast的广播)，上层保证 batchSize，同时插入过多数据可能导致失败
     *
     * @param newTaskList 新增的子任务列表
     */
    public boolean submitTask(List<TaskDO> newTaskList) {
        if (finished.get()) {
            return true;
        }
        if (CollectionUtils.isEmpty(newTaskList)) {
            return true;
        }
        // 基础处理（多循环一次虽然有些浪费，但分布式执行中，这点耗时绝不是主要占比，忽略不计！）
        newTaskList.forEach(task -> {
            task.setInstanceId(instanceId);
            task.setStatus(TaskStatus.WAITING_DISPATCH.getValue());
            task.setFailedCnt(0);
            task.setLastModifiedTime(System.currentTimeMillis());
            task.setCreatedTime(System.currentTimeMillis());
            task.setLastReportTime(-1L);
        });

        log.debug("[TaskTracker-{}] receive new tasks: {}", instanceId, newTaskList);
        return taskPersistenceService.batchSave(newTaskList);
    }

submitTask方法先填充一些字段信息，比如设置status为TaskStatus.WAITING_DISPATCH，然后调用taskPersistenceService.batchSave保存

batchSave

tech/powerjob/worker/persistence/TaskPersistenceService.java

    public boolean batchSave(List<TaskDO> tasks) {
        if (CollectionUtils.isEmpty(tasks)) {
            return true;
        }
        try {
            return execute(() -> taskDAO.batchSave(tasks));
        }catch (Exception e) {
            log.error("[TaskPersistenceService] batchSave tasks({}) failed.", tasks, e);
        }
        return false;
    }

    private static  <T> T execute(SupplierPlus<T> executor) throws Exception {
        return CommonUtils.executeWithRetry(executor, RETRY_TIMES, RETRY_INTERVAL_MS);
    }    

batchSave通过taskDAO.batchSave报错，它针对异常会重试3次，每次间隔100ms

TaskDAOImpl.batchSave

tech/powerjob/worker/persistence/TaskDAOImpl.java

    public boolean batchSave(Collection<TaskDO> tasks) throws SQLException {
        String insertSql = "insert into task_info(task_id, instance_id, sub_instance_id, task_name, task_content, address, status, result, failed_cnt, created_time, last_modified_time, last_report_time) values (?,?,?,?,?,?,?,?,?,?,?,?)";
        boolean originAutoCommitFlag ;
        try (Connection conn = connectionFactory.getConnection()) {
            originAutoCommitFlag = conn.getAutoCommit();
            conn.setAutoCommit(false);
            try ( PreparedStatement ps = conn.prepareStatement(insertSql)) {
                for (TaskDO task : tasks) {
                    fillInsertPreparedStatement(task, ps);
                    ps.addBatch();
                }
                ps.executeBatch();
                return true;
            } catch (Throwable e) {
                conn.rollback();
                throw e;
            } finally {
                conn.setAutoCommit(originAutoCommitFlag);
            }
        }
    }

TaskDAOImpl的batchSave直接通过jdbc的executeBatch进行批量保存

Dispatcher

tech/powerjob/worker/core/tracker/task/heavy/HeavyTaskTracker.java

    /**
     * 定时扫描数据库中的task（出于内存占用量考虑，每次最多获取100个），并将需要执行的任务派发出去
     */
    protected class Dispatcher implements Runnable {

        // 数据库查询限制，每次最多查询几个任务
        private static final int DB_QUERY_LIMIT = 100;

        @Override
        public void run() {

            if (finished.get()) {
                return;
            }

            Stopwatch stopwatch = Stopwatch.createStarted();

            // 1. 获取可以派发任务的 ProcessorTracker
            List<String> availablePtIps = ptStatusHolder.getAvailableProcessorTrackers();

            // 2. 没有可用 ProcessorTracker，本次不派发
            if (availablePtIps.isEmpty()) {
                log.debug("[TaskTracker-{}] no available ProcessorTracker now.", instanceId);
                return;
            }

            // 3. 避免大查询，分批派发任务
            long currentDispatchNum = 0;
            long maxDispatchNum = availablePtIps.size() * instanceInfo.getThreadConcurrency() * 2L;
            AtomicInteger index = new AtomicInteger(0);

            // 4. 循环查询数据库，获取需要派发的任务
            while (maxDispatchNum > currentDispatchNum) {

                int dbQueryLimit = Math.min(DB_QUERY_LIMIT, (int) maxDispatchNum);
                List<TaskDO> needDispatchTasks = taskPersistenceService.getTaskByStatus(instanceId, TaskStatus.WAITING_DISPATCH, dbQueryLimit);
                currentDispatchNum += needDispatchTasks.size();

                needDispatchTasks.forEach(task -> {
                    // 获取 ProcessorTracker 地址，如果 Task 中自带了 Address，则使用该 Address
                    String ptAddress = task.getAddress();
                    if (StringUtils.isEmpty(ptAddress) || RemoteConstant.EMPTY_ADDRESS.equals(ptAddress)) {
                        ptAddress = availablePtIps.get(index.getAndIncrement() % availablePtIps.size());
                    }
                    dispatchTask(task, ptAddress);
                });

                // 数量不足 或 查询失败，则终止循环
                if (needDispatchTasks.size() < dbQueryLimit) {
                    break;
                }
            }

            log.debug("[TaskTracker-{}] dispatched {} tasks,using time {}.", instanceId, currentDispatchNum, stopwatch.stop());
        }
    }

HeavyTaskTracker每5s调度一次Dispatcher，其run方法先通过ptStatusHolder.getAvailableProcessorTrackers()获取可以派发任务的ProcessorTracker，之后循环从数据库拉取一批状态为TaskStatus.WAITING_DISPATCH的任务，通过轮询的方式进行dispatchTask

dispatchTask

tech/powerjob/worker/core/tracker/task/heavy/HeavyTaskTracker.java

    protected void dispatchTask(TaskDO task, String processorTrackerAddress) {

        // 1. 持久化，更新数据库（如果更新数据库失败，可能导致重复执行，先不处理）
        TaskDO updateEntity = new TaskDO();
        updateEntity.setStatus(TaskStatus.DISPATCH_SUCCESS_WORKER_UNCHECK.getValue());
        // 写入处理该任务的 ProcessorTracker
        updateEntity.setAddress(processorTrackerAddress);
        boolean success = taskPersistenceService.updateTask(instanceId, task.getTaskId(), updateEntity);
        if (!success) {
            log.warn("[TaskTracker-{}] dispatch task(taskId={},taskName={}) failed due to update task status failed.", instanceId, task.getTaskId(), task.getTaskName());
            return;
        }

        // 2. 更新 ProcessorTrackerStatus 状态
        ptStatusHolder.getProcessorTrackerStatus(processorTrackerAddress).setDispatched(true);
        // 3. 初始化缓存
        taskId2BriefInfo.put(task.getTaskId(), new TaskBriefInfo(task.getTaskId(), TaskStatus.DISPATCH_SUCCESS_WORKER_UNCHECK, -1L));

        // 4. 任务派发
        TaskTrackerStartTaskReq startTaskReq = new TaskTrackerStartTaskReq(instanceInfo, task, workerRuntime.getWorkerAddress());
        TransportUtils.ttStartPtTask(startTaskReq, processorTrackerAddress, workerRuntime.getTransporter());

        log.debug("[TaskTracker-{}] dispatch task(taskId={},taskName={}) successfully.", instanceId, task.getTaskId(), task.getTaskName());
    }

dispatchTask先更新status为DISPATCH_SUCCESS_WORKER_UNCHECK，更新处理该任务的ProcessorTracker，之后构建TaskTrackerStartTaskReq，通过TransportUtils.ttStartPtTask派送执行task的请求

onReceiveTaskTrackerStartTaskReq

tech/powerjob/worker/actors/ProcessorTrackerActor.java

    @Handler(path = RemoteConstant.WPT_HANDLER_START_TASK, processType = ProcessType.NO_BLOCKING)
    public void onReceiveTaskTrackerStartTaskReq(TaskTrackerStartTaskReq req) {

        Long instanceId = req.getInstanceInfo().getInstanceId();

        // 创建 ProcessorTracker 一定能成功
        ProcessorTracker processorTracker = ProcessorTrackerManager.getProcessorTracker(
                instanceId,
                req.getTaskTrackerAddress(),
                () -> new ProcessorTracker(req, workerRuntime));

        TaskDO task = new TaskDO();

        task.setTaskId(req.getTaskId());
        task.setTaskName(req.getTaskName());
        task.setTaskContent(req.getTaskContent());
        task.setFailedCnt(req.getTaskCurrentRetryNums());
        task.setSubInstanceId(req.getSubInstanceId());

        processorTracker.submitTask(task);
    }

ProcessorTrackerActor的onReceiveTaskTrackerStartTaskReq主要是获取或者创建processorTracker，然后执行其submitTask提交任务

ProcessorTracker.submitTask

tech/powerjob/worker/core/tracker/processor/ProcessorTracker.java

    public void submitTask(TaskDO newTask) {

        // 一旦 ProcessorTracker 出现异常，所有提交到此处的任务直接返回失败，防止形成死锁
        // 死锁分析：TT创建PT，PT创建失败，无法定期汇报心跳，TT长时间未收到PT心跳，认为PT宕机（确实宕机了），无法选择可用的PT再次派发任务，死锁形成，GG斯密达 T_T
        if (lethal) {
            ProcessorReportTaskStatusReq report = new ProcessorReportTaskStatusReq()
                    .setInstanceId(instanceId)
                    .setSubInstanceId(newTask.getSubInstanceId())
                    .setTaskId(newTask.getTaskId())
                    .setStatus(TaskStatus.WORKER_PROCESS_FAILED.getValue())
                    .setResult(lethalReason)
                    .setReportTime(System.currentTimeMillis());

            TransportUtils.ptReportTask(report, taskTrackerAddress, workerRuntime);
            return;
        }

        boolean success = false;
        // 1. 设置值并提交执行
        newTask.setInstanceId(instanceInfo.getInstanceId());
        newTask.setAddress(taskTrackerAddress);

        HeavyProcessorRunnable heavyProcessorRunnable = new HeavyProcessorRunnable(instanceInfo, taskTrackerAddress, newTask, processorBean, omsLogger, statusReportRetryQueue, workerRuntime);
        try {
            threadPool.submit(heavyProcessorRunnable);
            success = true;
        } catch (RejectedExecutionException ignore) {
            log.warn("[ProcessorTracker-{}] submit task(taskId={},taskName={}) to ThreadPool failed due to ThreadPool has too much task waiting to process, this task will dispatch to other ProcessorTracker.",
                    instanceId, newTask.getTaskId(), newTask.getTaskName());
        } catch (Exception e) {
            log.error("[ProcessorTracker-{}] submit task(taskId={},taskName={}) to ThreadPool failed.", instanceId, newTask.getTaskId(), newTask.getTaskName(), e);
        }

        // 2. 回复接收成功
        if (success) {
            ProcessorReportTaskStatusReq reportReq = new ProcessorReportTaskStatusReq();
            reportReq.setInstanceId(instanceId);
            reportReq.setSubInstanceId(newTask.getSubInstanceId());
            reportReq.setTaskId(newTask.getTaskId());
            reportReq.setStatus(TaskStatus.WORKER_RECEIVED.getValue());
            reportReq.setReportTime(System.currentTimeMillis());

            TransportUtils.ptReportTask(reportReq, taskTrackerAddress, workerRuntime);

            log.debug("[ProcessorTracker-{}] submit task(taskId={}, taskName={}) success, current queue size: {}.",
                    instanceId, newTask.getTaskId(), newTask.getTaskName(), threadPool.getQueue().size());
        }
    }

ProcessorTracker的submitTask方法创建HeavyProcessorRunnable，提交到threadPool执行，之后回复ProcessorReportTaskStatusReq，告知status为TaskStatus.WORKER_RECEIVED

小结

• MapProcessor接口继承了BasicProcessor，它提供了默认的map方法用于分发子任务，它主要是构造了ProcessorMapTaskRequest，通过TransportUtils.reliableMapTask发送请求，它先把task保存下来，初始的status为TaskStatus.WAITING_DISPATCH；
• HeavyTaskTracker每5s调度一次Dispatcher，其run方法先通过ptStatusHolder.getAvailableProcessorTrackers()获取可以派发任务的ProcessorTracker，之后循环从数据库拉取一批状态为TaskStatus.WAITING_DISPATCH的任务，通过轮询的方式进行dispatchTask；
• dispatchTask先更新status为DISPATCH_SUCCESS_WORKER_UNCHECK，更新处理该任务的ProcessorTracker，之后构建TaskTrackerStartTaskReq，通过TransportUtils.ttStartPtTask派送执行task的请求；
• ProcessorTracker的submitTask方法创建HeavyProcessorRunnable，提交到threadPool执行，之后回复ProcessorReportTaskStatusReq，告知status为TaskStatus.WORKER_RECEIVED；
• 最后通过HeavyProcessorRunnable调用对应的processor.process方法执行具体的任务