Flink源码之TaskManager启动流程

从启动命令flink-daemon.sh可以看出TaskManger入口类为org.apache.flink.runtime.taskexecutor.TaskManagerRunner

TaskManagerRunner::main
TaskManagerRunner::runTaskManagerProcessSecurely
TaskManagerRunner::runTaskManager //构造TaskManagerRunner并调用start()方法
TaskManagerRunner::new //核心

在TaskManagerRunner构造函数中，可以看出与JobManger类似，也是先构造出一些公共服务：

highAvailabilityServices//用于获取JobManger的地址
rpcService //将TaskExecutor包装为AkkaActor提供RPC服务
heartbeatServices //心跳服务，与JobManger通信
metricRegistry //metric服务，提供metric注册和查询
blobCacheService //缓存Blob

这些服务在构造TaskExecutor时作为构造函数参数传入

构造TaskExecutor前会先构造TaskManagerServices辅助TaskExecutor实现其核心功能

TaskManagerRunner::createTaskExecutorService
TaskManagerRunner::startTaskManager // 构造MetricGroup和相关服务
TaskManagerServices.fromConfiguration//读取TaskManger的配置信息启动TaskManager相关服务
TaskExecutor::new  //核心

启动TaskEexector后会与ResouceManager建立连接，将自身信息注册到RM后发送Slot报告给RM，具体调用链路如下：

TaskManagerRunner::start
TaskExecutorToServiceAdapter::start
TaskExecutor::start
TaskExecutor::onStart
TaskExecutor::startTaskExecutorServices //获取ResourceManager地址后与ResourceManager建立连接，发送Slot报告
ResourceManagerLeaderListener::notifyLeaderAddress
TaskExecutor::notifyOfNewResourceManagerLeader
TaskExecutor::reconnectToResourceManager
TaskExecutor::tryConnectToResourceManager
TaskExecutor::connectToResourceManager
TaskExecutorToResourceManagerConnection::start
RegisteredRpcConnection::start
RegisteredRpcConnection::createNewRegistration
TaskExecutorToResourceManagerConnection::generateRegistration
RetryingRegistration::startRegistration //与resourcemanager建立连接
RetryingRegistration::register
ResourceManagerRegistration::invokeRegistration //向ResourceManager注册TaskExecutorRegistration信息
ResourceManagerGateway.registerTaskExecutor
TaskExecutorToResourceManagerConnection::onRegistrationSuccess
ResourceManagerRegistrationListener::onRegistrationSuccess
TaskExecutor::establishResourceManagerConnection 
    ResourceManagerGateway.sendSlotReport //发送自身slot信息给ResourceManager
    HeartbeatManagerImpl::monitorTarget//与RM建立心跳连接，当接到来自RM的心跳请求时，就会将SlotReport发送给RM作为心跳回应

TaskExecutor提供了以下两个核心方法：

 //RM将Slot分配给JobMaster请求TM将具体Slot信息发送给JobMaster
  CompletableFuture<Acknowledge> requestSlot(
        SlotID slotId,
        JobID jobId,
        AllocationID allocationId,
        ResourceProfile resourceProfile,
        String targetAddress,
        ResourceManagerId resourceManagerId,
        @RpcTimeout Time timeout);
 
 //执行JobMaster提交的物理Task       
CompletableFuture<Acknowledge> submitTask(
        TaskDeploymentDescriptor tdd, JobMasterId jobMasterId, @RpcTimeout Time timeout);    

TaskManager中管理Slot的实现类TaskSlotTableImpl，该实例记录了Slot的分配信息。

HeartBeat

在TaskExecutor构造函数中有两个HeartbeatManager，实现类都是HeartbeatManagerImpl，此类是接受心跳请求，发送心跳响应：

ResourceManagerHeartbeatManager //响应RM的心跳请求，心跳响应中带上SlotReport
JobManagerHeartbeatManager  //响应JobMaster的心跳请求, 心跳响应中带上AccumulatorReport

调用HeartbeatManagerImpl.monitorTarget(ResourceID resourceID, HeartbeatTarget heartbeatTarget) 与目标对象建立心跳连接。

HeartbeatManager还有个实现类是HeartbeatManagerSenderImpl，用于主动向监控目标发送心跳请求，比如在ResourceManager中创建的就是HeartbeatManagerSenderImpl，TaskManager启动时向ResourceManager注册后，RM就会调用HeartbeatManagerSenderImpl.monitor监控TM, 并定时向TM的HeartbeatManagerImpl发送心跳请求。同样，在JobMaster中创建的也是HeartbeatManagerSenderImpl，JobMaster定时向执行当前Job的TM发送心跳请求，TM响应与该Job相关信息。

综上，TM启动后向RM注册，与TM通过心跳信息同步Slot分配状况，接受RM的Slot分配请求向JobMaster提供Slot后，就可以接受JobMaster 执行具体的物理Task了。