基于Flink 1.12
Flink底层RPC是通过AKKA实现的,AKKA是基于Actor模型实现的框架。下面,将大致介绍一下actor模型。
在Actor模型中,一切事物都是actor,一个actor是一个基本的计算单元,每个actor是完全隔离的,不会共享内存,也就不会有共享数据带来的并发问题;它们是自己维护自身的状态,该状态不会被其他actor直接修改。
整体模型大致是:多个actor同时运行,每个actor接收消息,并根据消息做出相应的反应。消息本身是通过异步的形式发送给actor的,消息会被存储在一个叫做“邮箱(mailbox)”的地方,actor会顺序的处理收到的信息,避免锁的使用。从描述可以了解到actor模型中,消息的发送者和已发送消息解耦,是以并发的形式处理数据的。
RPC作用是让远程调用像本地调用,封装调用的细节。
Flink定义了各个组件的Gateway,通过回调的方式隐藏实现细节,将业务本身和通信解绑了,方便RPC调用。目前,Flink的RPC请求的底层通信是通过AKKA的实现的。
public abstract class RpcEndpoint implements RpcGateway, AutoCloseableAsyn{ //启动sever和获取RPC Gateway /** RPC service to be used to start the RPC server and to obtain rpc gateways. */ private final RpcService rpcService; //RpcServer用于启动和连接到RpcEndpoint, 连接到rpc服务器将返回一个RpcGateway,为RpcService提供RPC服务/连接远程Server /** Interface to access the underlying rpc server. */ protected final RpcServer rpcServer; }
是可以发送心跳和请求心跳相应组件接口,是对具备心跳能力对象的一种抽象。
HeartbeatTarget的函数具备以下两种动作:
HeartbeatMonitor管理HeartbeatTarget的心跳状态。当在指定时间内未收到心跳信息时,monitor将会通知对应的HeartbeatListener,收到心跳信息后会重置其定时器。其工厂接口如下:
HeartbeatMonitor<O> createHeartbeatMonitor( ResourceID resourceID, HeartbeatTarget<O> heartbeatTarget, ScheduledExecutor mainThreadExecutor, HeartbeatListener<?, O> heartbeatListener, //用于处理心跳信息 long heartbeatTimeoutIntervalMs);
HeartbeatListener是和HeartbeatManager交互的接口,Flink的业务的处理逻辑需要继承该接口以处理心跳结果,其三个回调函数如下:
心跳的管理者,用于开始/停止对HeartbeatTarget的心跳监控,以及会处理某个节点的心跳超时。
HeartbeatManager继承了HeartbeatTarget,其具有了HeartbeatTarget的函数功能以外,该接口还有以下四种函数:
核心接口交互的大致过程:HeartbeatManager将HeartbeatTarget放入到监控列表中,当心跳超时时,HeartbeatMonitor回通知HeartbeatListener处理,通过对HeartbeatListener的实现,完成相关处理心跳超时的逻辑。
下面通过分析1.3.1中核心接口的实现类,来具体分析心跳处理的过程。
该manager维护了一个heartbeat 的监控对象(HeartbeatMonitor)和资源ID信息,当收到新的心跳信息是,monitor对象将会被更新;心跳超时时,将会通知HeartbeatListenter对象。
public class HeartbeatManagerImpl<I, O> implements HeartbeatManager<I, O> { //心跳间隔 /** Heartbeat timeout interval in milli seconds. */ private final long heartbeatTimeoutIntervalMs; //心跳 /** Heartbeat listener with which the heartbeat manager has been associated. */ private final HeartbeatListener<I, O> heartbeatListener; //使用一个map存放资源-心跳的monitor信息,其monitorTarget方法就是将对应信息放入该map中 /** Map containing the heartbeat monitors associated with the respective resource ID. */ private final ConcurrentHashMap<ResourceID, HeartbeatMonitor<O>> heartbeatTargets; /** Running state of the heartbeat manager. */ protected volatile boolean stopped;
HearbeatManagerImpl实现的主要函数有:
继承于HearbeatManagerImpl,HeartbeatManagerSenderImpl向其监控的heartbeatTarget对象请求心跳的响应,属于主动触发心跳请求。实现了Runnable接口,在其run方法中,会遍历heartbeatMonitor,通过requestHeartbeat()方法向节点获取心跳信息。
public class HeartbeatManagerSenderImpl<I, O> extends HeartbeatManagerImpl<I, O> implements Runnable { @Override public void run() { if (!stopped) { log.debug("Trigger heartbeat request."); for (HeartbeatMonitor<O> heartbeatMonitor : getHeartbeatTargets().values()) { requestHeartbeat(heartbeatMonitor); } // 周期性调度,事件周期可配 getMainThreadExecutor().schedule(this, heartbeatPeriod, TimeUnit.MILLISECONDS); } } // 主动发起心跳请求 private void requestHeartbeat(HeartbeatMonitor<O> heartbeatMonitor) { O payload = getHeartbeatListener().retrievePayload(heartbeatMonitor.getHeartbeatTargetId()); final HeartbeatTarget<O> heartbeatTarget = heartbeatMonitor.getHeartbeatTarget(); // 调用Target的 requestHeartbeat函数 heartbeatTarget.requestHeartbeat(getOwnResourceID(), payload); } }
HeartbeatMonitor管理心跳目标,它在初始化会启动一个ScheduledExecutor。
public class HeartbeatMonitorImpl<O> implements HeartbeatMonitor<O>, Runnable { /** Resource ID of the monitored heartbeat target. */ private final ResourceID resourceID; //监控的资源ID /** Associated heartbeat target. */ private final HeartbeatTarget<O> heartbeatTarget; //心跳目标 private final ScheduledExecutor scheduledExecutor; /** Listener which is notified about heartbeat timeouts. */ private final HeartbeatListener<?, ?> heartbeatListener; HeartbeatMonitorImpl( ResourceID resourceID, HeartbeatTarget<O> heartbeatTarget, ScheduledExecutor scheduledExecutor, HeartbeatListener<?, O> heartbeatListener, long heartbeatTimeoutIntervalMs) { this.resourceID = Preconditions.checkNotNull(resourceID); this.heartbeatTarget = Preconditions.checkNotNull(heartbeatTarget); this.scheduledExecutor = Preconditions.checkNotNull(scheduledExecutor); this.heartbeatListener = Preconditions.checkNotNull(heartbeatListener); Preconditions.checkArgument( heartbeatTimeoutIntervalMs > 0L, "The heartbeat timeout interval has to be larger than 0."); this.heartbeatTimeoutIntervalMs = heartbeatTimeoutIntervalMs; lastHeartbeat = 0L; //初始化的时候,就启动一个定时任务 resetHeartbeatTimeout(heartbeatTimeoutIntervalMs); } @Override public void run() { // The heartbeat has timed out if we're in state running if (state.compareAndSet(State.RUNNING, State.TIMEOUT)) { //通知heartbeatListener处理 heartbeatListener.notifyHeartbeatTimeout(resourceID); } } void resetHeartbeatTimeout(long heartbeatTimeout) { if (state.get() == State.RUNNING) { cancelTimeout(); //重新开启新的定时任务 futureTimeout = scheduledExecutor.schedule(this, heartbeatTimeout, TimeUnit.MILLISECONDS); // Double check for concurrent accesses (e.g. a firing of the scheduled future) if (state.get() != State.RUNNING) { cancelTimeout(); } } } }
HeartbeatServices为所有需要心跳服务的创建heartbeat receivers and heartbeat senders。
public class HeartbeatServices { /** * 创建 heartbeat receivers * Creates a heartbeat manager which does not actively send heartbeats. */ public <I, O> HeartbeatManager<I, O> createHeartbeatManager( ResourceID resourceId, HeartbeatListener<I, O> heartbeatListener, ScheduledExecutor mainThreadExecutor, Logger log) { return new HeartbeatManagerImpl<>( heartbeatTimeout, resourceId, heartbeatListener, mainThreadExecutor, log); } /** * 创建 heartbeat sender * Creates a heartbeat manager which actively sends heartbeats to monitoring targets. */ public <I, O> HeartbeatManager<I, O> createHeartbeatManagerSender( ResourceID resourceId, HeartbeatListener<I, O> heartbeatListener, ScheduledExecutor mainThreadExecutor, Logger log) { return new HeartbeatManagerSenderImpl<>( heartbeatInterval, heartbeatTimeout, resourceId, heartbeatListener, mainThreadExecutor, log); } // 从配置文件配置心跳间隔时间和心跳超时时间 //两者的关系 0 < 心跳间隔时间 < 心跳超时时间 public static HeartbeatServices fromConfiguration(Configuration configuration) { long heartbeatInterval = configuration.getLong(HeartbeatManagerOptions.HEARTBEAT_INTERVAL); long heartbeatTimeout = configuration.getLong(HeartbeatManagerOptions.HEARTBEAT_TIMEOUT); return new HeartbeatServices(heartbeatInterval, heartbeatTimeout); } }
在一个Flink集群中只有一个ResourceManager(RM),和一个或多个TaskManager(TM)。两者的交互过程为:TM启动时会向RM注册,注册成功之后,RM会主动要求TM上报心跳信息。通过RM和TM的心跳信息,双方知道对方是否存活。
在2.2.4小节总,我们知道HeartbeatManagerSenderImpl属于Sender,HeartbeatManagerImpl属于Receiver。sender要对心跳目标上报心跳信息,receiver收到信息请求后返回一个response。
public abstract class ResourceManager<WorkerType extends ResourceIDRetrievable> extends FencedRpcEndpoint<ResourceManagerId> implements ResourceManagerGateway, LeaderContender { // RM启动时运行的方法 @Override public final void onStart() throws Exception { try { // 启动RMServices startResourceManagerServices(); } catch (Throwable t) { final ResourceManagerException exception = new ResourceManagerException( String.format("Could not start the ResourceManager %s", getAddress()), t); onFatalError(exception); throw exception; } } }
StandaloneLeaderElectionService#start | ResourceManager#grantLeadership | ResourceManager#tryAcceptLeadership | ResourceManager#startServicesOnLeadership //其具体实现如下 private void startServicesOnLeadership() { //启动心跳服务 startHeartbeatServices(); //slotManager是RM中管理slot的组件,其具体过程后续博客分析 slotManager.start(getFencingToken(), getMainThreadExecutor(), new ResourceActionsImpl()); //周期性判断是否存在未满足的slot请求 onLeadership(); }启动心跳服务,就是创建分别创建了taskManagerHeartbeatManager和jobManagerHeartbeatManager用于RM和TM、RM和JM的心跳服务
private void startHeartbeatServices() { taskManagerHeartbeatManager = heartbeatServices.createHeartbeatManagerSender( resourceId, new TaskManagerHeartbeatListener(), getMainThreadExecutor(), log); jobManagerHeartbeatManager = heartbeatServices.createHeartbeatManagerSender( resourceId, new JobManagerHeartbeatListener(), getMainThreadExecutor(), log); }
结合2.2.2小节,RM在心跳服务在和TM与JM的心跳过程中,充当的是请求心跳请求的发起方,即RM是主动去拉取心跳信息的。
TaskExecutor在创建时,就初始化了心跳组件。
public TaskExecutor( RpcService rpcService, TaskManagerConfiguration taskManagerConfiguration, HighAvailabilityServices haServices, TaskManagerServices taskExecutorServices, ExternalResourceInfoProvider externalResourceInfoProvider, HeartbeatServices heartbeatServices, TaskManagerMetricGroup taskManagerMetricGroup, @Nullable String metricQueryServiceAddress, BlobCacheService blobCacheService, FatalErrorHandler fatalErrorHandler, TaskExecutorPartitionTracker partitionTracker, BackPressureSampleService backPressureSampleService) { //创建HeartbeatManagerImpl,对JM的心跳进行相应 this.jobManagerHeartbeatManager = createJobManagerHeartbeatManager(heartbeatServices, resourceId); // 创建HeartbeatManagerImpl,对RM的心跳进行相应 this.resourceManagerHeartbeatManager = createResourceManagerHeartbeatManager(heartbeatServices, resourceId); }
TaskExecutor#onStart | TaskExecutor#startTaskExecutorServices | StandaloneLeaderRetrievalService#start //以standalone模式分析 | |//在standalone模式下,已知晓JobManager的地址,会直接去链接RM TaskExecutor.ResourceManagerLeaderListener#notifyLeaderAddress | TaskExecutor#notifyOfNewResourceManagerLeader | TaskExecutor#reconnectToResourceManager | |//在该方法中会主动调用TaskExecutorToResourceManagerConnection类的start方法去链接RM TaskExecutor#connectToResourceManager | | //在该函数的createNewRegistration方法中的回调函数,处理注册成功后的逻辑 RegisteredRpcConnection#start | |//z在该方法中会先链接RM,然后连接成功后发起注册请求 RetryingRegistration#startRegistration | RetryingRegistration#register | TaskExecutorToResourceManagerConnection#invokeRegistration
到此,TM向RM发起了注册,通过AKKA RPC,请求来到了RM中。
ResourceManager#registerTaskExecutor | |// 该方法的返回值是RegistrationResponse,在该方法中会将调用taskManagerHeartbeatManager.monitorTarget,监控节点的心跳信息 ResourceManager#registerTaskExecutorInternal | return new TaskExecutorRegistrationSuccess( registration.getInstanceID(), resourceId, clusterInformation)
//注册成功后将会走start方法中createNewRegistration创建registration时的回调函数 RegisteredRpcConnection#start | TaskExecutorToResourceManagerConnection#onRegistrationSuccess | TaskExecutor#onRegistrationSuccess | | //和RM建立联系,开始监控RM TaskExecutor#establishResourceManagerConnection | resourceManagerHeartbeatManager#monitorTarget
下面主要分析心跳交互过程。
public void start() { checkState(!closed, "The RPC connection is already closed"); checkState( !isConnected() && pendingRegistration == null, "The RPC connection is already started"); //会在创建newRegistration时,定义链接成功后处理逻辑 final RetryingRegistration<F, G, S, R> newRegistration = createNewRegistration(); if (REGISTRATION_UPDATER.compareAndSet(this, null, newRegistration)) { // 启动注册 newRegistration.startRegistration(); } else { // concurrent start operation newRegistration.cancel(); } }
在此过程中有好多回调,需要慢慢的体会。
由RM的初始化的分析,我们了解到,RM会主动要求TM上报心跳,其过程如下:
// 在该该方法中会创建一个HeartbeatManagerSenderImpl ResourceManager#startHeartbeatServices | | //这里会一步步调用构造函数中,在该构造函数中,会将心跳检查加入周期性任务列表中 HeartbeatManagerImpl | | //在任务启动时,会调用HeartbeatManagerSenderImpl的run方法,在该方法中会循环遍历HeartbeatMonitor,通过requestHeartbeat要求target上报心跳信息 HeartbeatManagerSenderImpl#run | | //该调用会跑到ResourceManager#TaskManagerHeartbeatListener中,这里返回为null是因为RM不是任何组件的receiver,即不会有组件向RM请求心跳信息,并要求其返回心跳。 getHeartbeatListener().retrievePayload | | //这里会调用TM向RM注册时指定的requestHeartbeat heartbeatTarget.requestHeartbeat | | //ResouceManager#registerTaskExecutorInternal taskExecutorGateway.heartbeatFromResourceManager
通过RPC调用,请求来到了TM中,其过程如下:
TaskExecutor#heartbeatFromResourceManager | HeartbeatManagerImpl#requestHeartbeat | HeartbeatMonitorImpl#reportHeartbeat | | //在该方法中,判断若是running则会取消之前的Timeout定时任务ScheduledFuture,重新开始检查是否timeout超时的定时任务。 HeartbeatMonitorImpl#resetHeartbeatTimeout | | //因为从RM发来的请求中heartbeatPayload为null,则TM直接走返回心跳反应的流程 HeartbeatMonitorImpl#reportHeartbea->heartbeatTarget.receiveHeartbeat | | //这里生成TM的心跳信息,包括slot信息 TaskExecutor.ResourceManagerHeartbeatListener#retrievePayload | | //通过在TM向RM注册过程中定义的receiveHeartbeat方法来实现调用RM中方法 TaskExecutor#establishResourceManagerConnection-> resourceManagerGateway.heartbeatFromTaskManager
RM收到TM的心跳信息,主要做了两件事:重置RM的Monitor线程;解析TM上报信息
ResourceManager#heartbeatFromTaskManager | HeartbeatManagerImpl#reportHeartbeat | | //和TM一样,重置了monitor线程 HeartbeatManagerImpl#reportHeartbeat->reportHeartbeat | | //在该方法中处理上报的slot信息, ResourceManager.TaskManagerHeartbeatListener#reportPayload
周期性心跳的具体分析过程见上述流程中的注释。