[源码解析] Flink的Slot究竟是什么?(2)

时间:2022-07-24
本文章向大家介绍[源码解析] Flink的Slot究竟是什么?(2),主要内容包括其使用实例、应用技巧、基本知识点总结和需要注意事项,具有一定的参考价值,需要的朋友可以参考一下。

[源码解析] Flink 的slot究竟是什么?(2)

0x00 摘要

Flink的Slot概念大家应该都听说过,但是可能很多朋友还不甚了解其中细节,比如具体Slot究竟代表什么?在代码中如何实现?Slot在生成执行图、调度、分配资源、部署、执行阶段分别起到什么作用?本文和上文将带领大家一起分析源码,为你揭开Slot背后的机理。

0x01 前文回顾

书接上回 [源码解析] Flink 的slot究竟是什么?(1)。前文中我们已经从系统架构和数据结构角度来分析了Slot,本文我们将从业务流程角度来分析Slot。我们重新放出系统架构图

和数据结构逻辑关系图

下面我们从几个流程入手一一分析。

0x02 注册/更新Slot

有两个途径会注册Slot/更新Slot状态。

  • 当TaskExecutor注册成功之后会和RM交互进行注册时,一并注册Slot;
  • 定时心跳时,会在心跳payload中附加Slot状态信息;

2.1 TaskExecutor注册成功

当TaskExecutor注册成功之后会和RM交互进行注册。会通过如下的代码调用路径来向ResourceManager(SlotManagerImpl)注册Slot。SlotManagerImpl 在获取消息之后,会更新Slot状态,如果此时已经有如果有pendingSlotRequest,就直接分配,否则就更新freeSlots变量。

  • TaskExecutor#establishResourceManagerConnection;
  • TaskSlotTableImpl#createSlotReport;建立 report
    • 这时候的 report如下: slotReport = {SlotReport@9633} 0 = {SlotStatus@8969} "SlotStatus{slotID=40d390ec-7d52-4f34-af86-d06bb515cc48_0, resourceProfile=ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}, allocationID=null, jobID=null}" slotID = {SlotID@8629} "40d390ec-7d52-4f34-af86-d06bb515cc48_0" resourceProfile = {ResourceProfile@4194} "ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}" allocationID = null jobID = null 1 = {SlotStatus@9638} "SlotStatus{slotID=40d390ec-7d52-4f34-af86-d06bb515cc48_1, resourceProfile=ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}, allocationID=null, jobID=null}" slotID = {SlotID@9643} "40d390ec-7d52-4f34-af86-d06bb515cc48_1" resourceProfile = {ResourceProfile@4194} "ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}" allocationID = null jobID = null
  • ResourceManager#sendSlotReport;通过RPC(resourceManagerGateway.sendSlotReport)调用到RM
  • SlotManagerImpl#registerTaskManager;把TaskManager注册到SlotManager
  • SlotManagerImpl#registerSlot;
  • SlotManagerImpl#createAndRegisterTaskManagerSlot;生成注册了TaskManagerSlot
    • 这时候代码 & 变量如下,我们可以看到,就是把TM的Slot信息注册到SlotManager中: private TaskManagerSlot createAndRegisterTaskManagerSlot(SlotID slotId, ResourceProfile resourceProfile, TaskExecutorConnection taskManagerConnection) { final TaskManagerSlot slot = new TaskManagerSlot( slotId, resourceProfile, taskManagerConnection); slots.put(slotId, slot); return slot; } slot = {TaskManagerSlot@13322} slotId = {SlotID@8629} "40d390ec-7d52-4f34-af86-d06bb515cc48_0" resourceProfile = {ResourceProfile@4194} cpuCores = {CPUResource@11616} "Resource(CPU: 89884656743115785...0)" taskHeapMemory = {MemorySize@11617} "4611686018427387903 bytes" taskOffHeapMemory = {MemorySize@11618} "4611686018427387903 bytes" managedMemory = {MemorySize@11619} "64 mb" networkMemory = {MemorySize@11620} "32 mb" extendedResources = {HashMap@11621} size = 0 taskManagerConnection = {WorkerRegistration@11121} allocationId = null jobId = null assignedSlotRequest = null state = {TaskManagerSlot$State@13328} "FREE"
  • SlotManagerImpl#updateSlot
  • SlotManagerImpl#updateSlotState;如果有pendingSlotRequest,就直接分配
  • SlotManagerImpl#handleFreeSlot;否则就更新freeSlots变量

流程结束后,SlotManager如下,可以看到此时slots个数是两个,freeSlots也是两个,说明都是空闲的:

this = {SlotManagerImpl@11120} 
 scheduledExecutor = {ActorSystemScheduledExecutorAdapter@11125} 
 slotRequestTimeout = {Time@11127} "300000 ms"
 taskManagerTimeout = {Time@11128} "30000 ms"
 slots = {HashMap@11122}  size = 2
  {SlotID@9643} "40d390ec-7d52-4f34-af86-d06bb515cc48_1" -> {TaskManagerSlot@19206} 
  {SlotID@8629} "40d390ec-7d52-4f34-af86-d06bb515cc48_0" -> {TaskManagerSlot@13322} 
 freeSlots = {LinkedHashMap@11129}  size = 2
  {SlotID@8629} "40d390ec-7d52-4f34-af86-d06bb515cc48_0" -> {TaskManagerSlot@13322} 
  {SlotID@9643} "40d390ec-7d52-4f34-af86-d06bb515cc48_1" -> {TaskManagerSlot@19206} 
 taskManagerRegistrations = {HashMap@11130}  size = 1
 fulfilledSlotRequests = {HashMap@11131}  size = 0
 pendingSlotRequests = {HashMap@11132}  size = 0
 pendingSlots = {HashMap@11133}  size = 0
 slotMatchingStrategy = {AnyMatchingSlotMatchingStrategy@11134} "INSTANCE"
 slotRequestTimeoutCheck = {ActorSystemScheduledExecutorAdapter$ScheduledFutureTask@11139} 

2.2 心跳机制更新Slot状态

Flink的心跳机制也会被利用来进行Slots信息的汇报,Slot Report被包括在心跳payload中。

首先在 TE 中建立Slot Report

  • TaskExecutor#heartbeatFromResourceManager
  • HeartbeatManagerImpl#requestHeartbeat
  • TaskExecutor$ResourceManagerHeartbeatListener # retrievePayload
  • TaskSlotTableImpl # createSlotReport

程序运行到 RM,于是 SlotManagerImpl 调用到 reportSlotStatus,进行Slot状态更新。

  • ResourceManager#heartbeatFromTaskManager
  • HeartbeatManagerImpl#receiveHeartbeat
  • ResourceManager$TaskManagerHeartbeatListener#reportPayload
  • SlotManagerImpl#reportSlotStatus,此时的SlotReport如下:
    • slotReport = {SlotReport@8718} slotsStatus = {ArrayList@8717} size = 2 0 = {SlotStatus@9025} "SlotStatus{slotID=d99e16d7-a30c-4e21-b270-f82884b1813f_0, resourceProfile=ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}, allocationID=null, jobID=null}" slotID = {SlotID@9032} "d99e16d7-a30c-4e21-b270-f82884b1813f_0" resourceProfile = {ResourceProfile@4194} "ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}" allocationID = null jobID = null 1 = {SlotStatus@9026} "SlotStatus{slotID=d99e16d7-a30c-4e21-b270-f82884b1813f_1, resourceProfile=ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}, allocationID=null, jobID=null}" slotID = {SlotID@9029} "d99e16d7-a30c-4e21-b270-f82884b1813f_1" resourceProfile = {ResourceProfile@4194} "ResourceProfile{managedMemory=64.000mb (67108864 bytes), networkMemory=32.000mb (33554432 bytes)}" allocationID = null jobID = null
  • SlotManagerImpl#updateSlot
  • SlotManagerImpl#updateSlotState;如果有pendingSlotRequest,就直接分配
  • SlotManagerImpl#handleFreeSlot;否则就更新freeSlots变量
    • freeSlots.put(freeSlot.getSlotId(), freeSlot);

0x03 生成ExecutionGraph阶段

当Job提交之后,经过一系列处理,Scheduler会建立ExecutionGraph。ExecutionGraph 是 JobGraph 的并行版本。而通过一系列的分析,才可以最终把任务分发到相关的任务槽中。槽会根据CPU的数量提前指定出来,这样可以最大限度的利用CPU的计算资源。如果Slot耗尽,也就意味着新分发的作业任务是无法执行的。

ExecutionGraphJobManager根据JobGraph生成的分布式执行图,是调度层最核心的数据结构。

一个JobVertex / ExecutionJobVertex代表的是一个operator,而具体的ExecutionVertex则代表了一个Task。

在生成StreamGraph时候,StreamGraph.addOperator方法就已经确定了operator是什么类型,比如OneInputStreamTask,或者SourceStreamTask等。

假设OneInputStreamTask.class即为生成的StreamNode的vertexClass。这个值会一直传递,当StreamGraph被转化成JobGraph的时候,这个值会被传递到JobVertex的invokableClass。然后当JobGraph被转成ExecutionGraph的时候,这个值被传入到ExecutionJobVertex.TaskInformation.invokableClassName中,最后一直传到Task中。

本系列代码执行序列如下:

  • JobMaster#createScheduler
  • DefaultSchedulerFactory#createInstance
  • DefaultScheduler#init
  • SchedulerBase#init
  • SchedulerBase#createAndRestoreExecutionGraph
  • SchedulerBase#createExecutionGraph
  • ExecutionGraphBuilder#buildGraph
  • ExecutionGraph#attachJobGraph
  • ExecutionJobVertex#init,这里根据并行度来确定要建立多少个Task,即多少个ExecutionVertex。
    • int numTaskVertices = vertexParallelism > 0 ? vertexParallelism : defaultParallelism; this.taskVertices = new ExecutionVertex[numTaskVertices];
  • ExecutionVertex#init,这里会生成Execution。
    • this.currentExecution = new Execution( getExecutionGraph().getFutureExecutor(), this, 0, initialGlobalModVersion, createTimestamp, timeout);

0x04 调度阶段

任务的流程就是通过作业分发到TaskManager,然后再分发到指定的Slot进行执行。

这部分调度阶段的代码只是利用CompletableFuture把程序执行架构搭建起来,可以把认为是自顶之下进行操作

Job开始调度之后,代码执行序列如下:

  • JobMaster#startJobExecution
  • JobMaster#resetAndStartScheduler
  • Future操作
  • JobMaster#startScheduling
  • SchedulerBase#startScheduling
  • DefaultScheduler#startSchedulingInternal
  • LazyFromSourcesSchedulingStrategy#startScheduling,这里开始针对Vertices进行资源分配和部署
    • allocateSlotsAndDeployExecutionVertices(schedulingTopology.getVertices());
  • LazyFromSourcesSchedulingStrategy#allocateSlotsAndDeployExecutionVertices,这里会遍历ExecutionVertex,筛选出Create状态的 & 输入Ready的节点。
    • private void allocateSlotsAndDeployExecutionVertices( final Iterable<? extends SchedulingExecutionVertex<?, ?>> vertices) { // 取出状态是CREATED,且输入Ready的 ExecutionVertex final Set<ExecutionVertexID> verticesToDeploy = IterableUtils.toStream(vertices) .filter(IS_IN_CREATED_EXECUTION_STATE.and(isInputConstraintSatisfied())) .map(SchedulingExecutionVertex::getId) .collect(Collectors.toSet()); // 根据 ExecutionVertex 建立 DeploymentOption final List<ExecutionVertexDeploymentOption> vertexDeploymentOptions = ...; // 分配资源并且部署 schedulerOperations.allocateSlotsAndDeploy(vertexDeploymentOptions); }
  • DefaultScheduler#allocateSlotsAndDeploy

这里来到了本文第一个关键函数 allocateSlotsAndDeploy。其主要功能是:

  1. allocateSlots分配Slot,其实这时候并没有分配,而是建立一系列Future,然后根据Future返回SlotExecutionVertexAssignment列表。
  2. 根据SlotExecutionVertexAssignment建立DeploymentHandle
  3. 根据deploymentHandles进行部署,其实是根据Future把部署搭建起来,具体如何部署需要在slot分配成功之后再执行。
@Override
public void allocateSlotsAndDeploy(final List<ExecutionVertexDeploymentOption> executionVertexDeploymentOptions) {
   validateDeploymentOptions(executionVertexDeploymentOptions);

   final Map<ExecutionVertexID, ExecutionVertexDeploymentOption> deploymentOptionsByVertex =
      groupDeploymentOptionsByVertexId(executionVertexDeploymentOptions);

   final List<ExecutionVertexID> verticesToDeploy = executionVertexDeploymentOptions.stream()
      .map(ExecutionVertexDeploymentOption::getExecutionVertexId)
      .collect(Collectors.toList());

   final Map<ExecutionVertexID, ExecutionVertexVersion> requiredVersionByVertex =
      executionVertexVersioner.recordVertexModifications(verticesToDeploy);

   transitionToScheduled(verticesToDeploy);

   // 分配Slot,其实这时候并没有分配,而是建立一系列Future,然后根据Future返回SlotExecutionVertexAssignment列表
   final List<SlotExecutionVertexAssignment> slotExecutionVertexAssignments =
      allocateSlots(executionVertexDeploymentOptions);

   // 根据SlotExecutionVertexAssignment建立DeploymentHandle
   final List<DeploymentHandle> deploymentHandles = createDeploymentHandles(
      requiredVersionByVertex,
      deploymentOptionsByVertex,
      slotExecutionVertexAssignments);
  
   // 根据deploymentHandles进行部署,其实是根据Future把部署搭建起来,具体如何部署需要在slot分配成功之后再执行
   if (isDeployIndividually()) {
      deployIndividually(deploymentHandles);
   } else {
      waitForAllSlotsAndDeploy(deploymentHandles);
   }
}

接下来 两个小章节我们分别针对 allocateSlots 和 deployIndividually / waitForAllSlotsAndDeploy 进行分析。

0x05 分配资源阶段

注意,此处的入口为 allocateSlotsAndDeploy 的allocateSlots 调用

在分配slot时,首先会在JobMaster中SlotPool中进行分配,具体是先SlotPool中获取所有slot,然后尝试选择一个最合适的slot进行分配,这里的选择有两种策略,即按照位置优先和按照之前已分配的slot优先;若从SlotPool无法分配,则通过RPC请求向ResourceManager请求slot,若此时并未连接上ResourceManager,则会将请求缓存起来,待连接上ResourceManager后再申请。

5.1 CompletableFuture

CompletableFuture 首先是一个 Future,它拥有 Future 所有的功能,包括取得异步执行结果,取消正在执行的任务等,其次是 一个CompleteStage,其最大作用是将回调改为链式调用,从而将 Future 组合起来。

此处生成了执行框架,即通过三个 CompletableFuture 构成了执行框架

我们按照出现顺序命名为 Future 1,Future 2,Future 3。

但是这个反过来说明反而更方便。我们可以看到,'

出现次序是 Future 1,Future 2,Future 3

调用顺序是 Future 3 ---> Future 2 ---> Future 1

5.1.1 Future 3

我们可以称之为 PhysicalSlot Future

类型是:CompletableFuture

生成在:requestNewAllocatedSlot 函数中对 PendingRequest 的生成。PendingRequest 的构造函数中有 new CompletableFuture<>(),这个 Future 3 是 PendingRequest 的成员变量。

用处是:

  • PendingRequest 会 加入到 waitingForResourceManager

回调函数作用是:

  • 在 allocateMultiTaskSlot 的 whenComplete 会把payload赋值给slot,allocatedSlot.tryAssignPayload
  • 进一步回调在 createRootSlot 函数 的 forward . thenApply 语句,会 设置为 Future 3 回调 Future 2 的回调函数

何时回调

  • TM,TE offer Slot的时候,会根据 PendingRequest 间接回调到这里

6.1.2 Future 2

我们可以称之为 allocationFuture

类型是:

  • CompletableFuture ,CompletableFuture 有类型转换

生成在:

  • createRootSlot函数中。final CompletableFuture slotContextFutureAfterRootSlotResolution = new CompletableFuture<>();

用处是:

  • 把 Future 2 设置为 multiTaskSlot 的成员变量 private final CompletableFuture<? extends SlotContext> slotContextFuture;
  • Future 2 其实也就是 SingleTaskSlot 的 parent.getSlotContextFuture(),因为 multiTaskSlot 和 SingleTaskSlot 是父子关系
  • 在 SingleTaskSlot 构造函数 中,Future 2 会赋值给 SingleTaskSlot 的成员变量 singleLogicalSlotFuture。
  • 即 Future 2 实际上是 SingleTaskSlot 的成员变量 singleLogicalSlotFuture
  • SchedulerImpl # allocateSharedSlot 函数,return leaf.getLogicalSlotFuture(); 会被返回 singleLogicalSlotFuture 给外层调用,就是外层看到的 allocationFuture。

回调函数作用是:

  • 在 SingleTaskSlot 构造函数 中,会生成一个 SingleLogicalSlot(未来回调时候会真正生成 )
  • 在 internalAllocateSlot 函数中,会回调 Future 1,allocationResultFuture的回调函数

何时回调

  • 被 Future 3 的回调函数调用

6.1.3 Future 1

我们可以称之为 allocationResultFuture

类型是:

  • CompletableFuture

生成在:

  • SchedulerImpl#allocateSlotInternal,这里生成了第一个 CompletableFuture

用处是:

  • 后续 Deploy 时候会用到 这个 Future 1,会通过 handle 给 Future 1 再加上两个后续调用,是在 Future 1 结束之后的后续调用。

回调函数作用是:

  • allocateSlotsFor 函数中有错误处理
  • 后续 Deploy 时候会用到 这个 Future 1,会通过 handle 给 Future 1 再加上两个后续调用,是在 Future 1 结束之后的后续调用。

何时回调

  • 语句在internalAllocateSlot中,但是在 Future 2 回调函数中调用

5.2 流程图

这里比较复杂,先给出流程图

 *  Run in Job Manager
 *
 *    DefaultScheduler#allocateSlotsAndDeploy 
 *        |
 *        +----> DefaultScheduler#allocateSlots
 *        |     //把ExecutionVertex转化为ExecutionVertexSchedulingRequirements
 *        |     
 *        +----> DefaultExecutionSlotAllocator#allocateSlotsFor( 调用 1 开始 )
 *        |     // 得到 我们的第一个 CompletableFuture,我们称之为 Future 1
 *        |  
 *        |    
 *        +--------------> NormalSlotProviderStrategy#allocateSlot 
 *        |    
 *        |        
 *        +--------------> SchedulerImpl#allocateSlotInternal
 *        |     // 生成了第一个 CompletableFuture,以后称之为 allocationResultFuture
 *        |  
 *  ┌────────────┐   
 *  │  Future 1  │ 生成 allocationResultFuture
 *  └────────────┘ 
 *        │     
 *        │            
 *        +----> SchedulerImpl#internalAllocateSlot( 调用 2 开始 )  
 *        |      // Future 1 做为参数被传进来,这里会继续调用,生成 Future 2, Future 3
 *        |     
 *        |       
 *        +-----------> SchedulerImpl#allocateSharedSlot( 调用 3 开始 )
 *        |    	// 这里涉及到 MultiTaskSlot 和 SingleTaskSlot
 *        |        
 *        +-----------> SchedulerImpl # allocateMultiTaskSlot ( 调用 4 开始 )
 *        |  
 *        |       
 *        +--------------------> SchedulerImpl # requestNewAllocatedSlot
 *        |    
 *        |    
 *        +--------------------> SlotPoolImpl#requestNewAllocatedSlot
 *        |    	// 这里生成一个 PendingRequest
 *        | 	// PendingRequest的构造函数中有 new CompletableFuture<>(),
 *        |     // 所以这里是生成了第三个 Future,注意这里的 Future 是针对 PhysicalSlot  
 *        |  
 *        |  
 *  ┌────────────┐   
 *  │  Future 3  │ 生成 Future<PhysicalSlot>,这个 Future 3 实际是对用户不可见的。
 *  └────────────┘      
 *        |  
 *        | 
 *        +-----------> SchedulerImpl # allocateMultiTaskSlot( 调用 4 结束 )
 *        |      // 回到 ( 调用 4 ) 这里,得倒 Future 3
 *        |      // 这里得倒了第三个 Future<PhysicalSlot> 
 *        |      // 第三是因为从用户角度看,它是第三个出现的     
 *        |     
 *        +-----------------------> slotSharingManager # createRootSlot  
 *        |      // 把 Future 3 做为参数传进去     
 *        |      // 这里马上生成 Future 2
 *        |      // Future 2 被设置为 multiTaskSlot 的成员变量 slotContextFuture;  
 *        |      // 然后forward . thenApply 语句 会 设置为 Future 3 回调 Future 2 的回调函数
 *        |     
 *        |       
 *        +-----------> SchedulerImpl#allocateSharedSlot
 *        |    	// 回到 ( 调用 3 ) 这里   
 *        |     
 *        | 
 *        +-----------------------> SlotSharingManager#allocateSingleTaskSlo  
 *        |  // 在 rootMultiTaskSlot 之上生成一个 SingleTaskSlot leaf加入到allTaskSlots。    
 *        |  // leaf.getLogicalSlotFuture(); 这个就是Future 2,设置好的
 *        |   
 *        |        
 *        +-----------> SchedulerImpl#allocateSharedSlot
 *        |    	// 还在 ( 调用 3 ) 这里  
 *        |     // return leaf.getLogicalSlotFuture(); 返回 Future 2   
 *        | 
 *        |       
 *  ┌────────────┐   
 *  │  Future 2  │
 *  └────────────┘  
 *        |     
 *        |       
 *        |       
 *        +----> SchedulerImpl#internalAllocateSlot    
 *        |      // 回到 ( 调用 2 ) 这里 
 *        |      // 设置,在 Future 2 的回调函数中会调用 Future 1    
 *        |  
 *        |      
 *        +----> DefaultExecutionSlotAllocator#allocateSlotsFor 
 *        |     // 回到 ( 调用 1 ) 这里 
 *        |
 *        |  
 *        |       
 *  ┌────────────┐   
 *  │  Future 1  │ 
 *  └────────────┘    
 *        |  
 *        |      
 *        +---->  createDeploymentHandles    
 *        |  // 生成 DeploymentHandle
 *        |     
 *        |       
 *        +-----------> deployIndividually(deploymentHandles);    
 *        |           // 这里会给 Future 1 再加上两个 回调函数,作为 部署回调
 *        | 

下图是为了手机阅读。

5.3 具体执行路径

默认情况下,Flink 允许subtasks共享slot,条件是它们都来自同一个Job的不同task的subtask。结果可能一个slot持有该job的整个pipeline。允许slot共享有以下两点好处:

  • Flink 集群所需的task slots数与job中最高的并行度一致。也就是说我们不需要再去计算一个程序总共会起多少个task了。
  • 更容易获得更充分的资源利用。如果没有slot共享,那么非密集型操作source/flatmap就会占用同密集型操作 keyAggregation/sink 一样多的资源。如果有slot共享,将基线的2个并行度增加到6个,能充分利用slot资源,同时保证每个TaskManager能平均分配到重的subtasks。

此处执行路径大致如下:

  • DefaultScheduler#allocateSlotsAndDeploy
  • DefaultScheduler#allocateSlots;该过程会把ExecutionVertex转化为ExecutionVertexSchedulingRequirements,会封装包含一些location信息、sharing信息、资源信息等
  • DefaultExecutionSlotAllocator#allocateSlotsFor;我们小节实际是从这里开始分析,这里会进行一系列操作,一层层调用下去。首先这个函数会得到我们的第一个 CompletableFuture,我们称之为 allocationResultFuture,这个名字的由来后续就会知道。这个 slotFuture 会赋值给 SlotExecutionVertexAssignment,然后传递给外面。后续 Deploy 时候会用到 这个 slotFuture,会通过 handle 给 slotFuture 再加上两个后续调用,是在slotFuture结束之后的后续调用。
    • public List<SlotExecutionVertexAssignment> allocateSlotsFor(...) { for (ExecutionVertexSchedulingRequirements schedulingRequirements : executionVertexSchedulingRequirements) { // 得到第一个 CompletableFuture,具体是在 calculatePreferredLocations 中通过 CompletableFuture<LogicalSlot> slotFuture = calculatePreferredLocations(...).thenCompose(...) -> slotProviderStrategy.allocateSlot( // 函数里面生成了第一个CompletableFuture slotRequestId, new ScheduledUnit(...), SlotProfile.priorAllocation(...))); SlotExecutionVertexAssignment slotExecutionVertexAssignment = new SlotExecutionVertexAssignment(executionVertexId, slotFuture); slotFuture.whenComplete( (ignored, throwable) -> { // 第一个CompletableFuture的回调函数,里其实只是异常处理,后续有人会调用到这里 pendingSlotAssignments.remove(executionVertexId); if (throwable != null) { slotProviderStrategy.cancelSlotRequest(slotRequestId, slotSharingGroupId, throwable); } }); slotExecutionVertexAssignments.add(slotExecutionVertexAssignment); } return slotExecutionVertexAssignments; }
  • NormalSlotProviderStrategy#allocateSlot(slotProviderStrategy.allocateSlot)
  • SchedulerImpl#allocateSlotInternal,这里生成了第一个 CompletableFuture,我们可以称之为 allocationResultFuture
    • private CompletableFuture<LogicalSlot> allocateSlotInternal(...) { // 这里生成了第一个 CompletableFuture,我们以后称之为 allocationResultFuture final CompletableFuture<LogicalSlot> allocationResultFuture = new CompletableFuture<>(); // allocationResultFuture 会传送进去继续处理 internalAllocateSlot(allocationResultFuture, slotRequestId, scheduledUnit, slotProfile, allocationTimeout); // 返回 allocationResultFuture return allocationResultFuture; }
  • SchedulerImpl#allocateSlot
  • SchedulerImpl#internalAllocateSlot,该方法会根据vertex是否共享slot来分配singleSlot/SharedSlot。这里得到第二个 CompletableFuture,我们以后成为 allocationFuture
    • private void internalAllocateSlot( CompletableFuture<LogicalSlot> allocationResultFuture, ...) { // 这里得到第二个 CompletableFuture,我们以后称为 allocationFuture,注意目前只是得到,不是生成。 CompletableFuture<LogicalSlot> allocationFuture = scheduledUnit.getSlotSharingGroupId() == null ? allocateSingleSlot(slotRequestId, slotProfile, allocationTimeout) : allocateSharedSlot(slotRequestId, scheduledUnit, slotProfile, allocationTimeout); // 第二个Future,allocationFuture的回调函数。注意,CompletableFuture可以连续调用多个whenComplete。 allocationFuture.whenComplete((LogicalSlot slot, Throwable failure) -> { if (failure != null) { // 异常处理 cancelSlotRequest(...); allocationResultFuture.completeExceptionally(failure); } else { allocationResultFuture.complete(slot); // 它将回调第一个 allocationResultFuture的回调函数 } }); }
  • SchedulerImpl#allocateSharedSlot,这里也比较复杂,涉及到 MultiTaskSlot 和 SingleTaskSlot
    • private CompletableFuture<LogicalSlot> allocateSharedSlot(...) { // allocate slot with slot sharing final SlotSharingManager multiTaskSlotManager = slotSharingManagers.computeIfAbsent( scheduledUnit.getSlotSharingGroupId(), id -> new SlotSharingManager(id,slotPool,this)); // 生成 SlotSharingManager final SlotSharingManager.MultiTaskSlotLocality multiTaskSlotLocality; if (scheduledUnit.getCoLocationConstraint() != null) { multiTaskSlotLocality = allocateCoLocatedMultiTaskSlot(...); } else { multiTaskSlotLocality = allocateMultiTaskSlot(...); // 这里生成 MultiTaskSlot } // 这里生成 SingleTaskSlot final SlotSharingManager.SingleTaskSlot leaf = multiTaskSlotLocality.getMultiTaskSlot().allocateSingleTaskSlot(...); return leaf.getLogicalSlotFuture(); // 返回 SingleTaskSlot 的 future,就是第二个Future,具体生成我们在下面会详述 }
  • SchedulerImpl # allocateMultiTaskSlot,这里是一个难点函数。因为这里生成了第三个 Future ,这里把第三个 Future 提前说明第三是因为从用户角度看,它是第三个出现的
    • private SlotSharingManager.MultiTaskSlotLocality allocateMultiTaskSlot(...) { SlotSharingManager.MultiTaskSlot multiTaskSlot = slotSharingManager.getUnresolvedRootSlot(groupId); if (multiTaskSlot == null) { // requestNewAllocatedSlot 会调用 SlotPoolImpl 的同名函数 // 得到第 三 个 Future,注意,这个 Future 针对的是 PhysicalSlot final CompletableFuture<PhysicalSlot> slotAllocationFuture = requestNewAllocatedSlot(...); // 使用 第 三 个 Future 来构建 multiTaskSlot multiTaskSlot = slotSharingManager.createRootSlot(...,slotAllocationFuture,...); // 第 三 个 Future的回调函数,这里会把payload赋值给slot slotAllocationFuture.whenComplete( (PhysicalSlot allocatedSlot, Throwable throwable) -> { final SlotSharingManager.TaskSlot taskSlot = slotSharingManager.getTaskSlot(multiTaskSlotRequestId); if (taskSlot != null) { // 会把payload赋值给slot if (!allocatedSlot.tryAssignPayload(((SlotSharingManager.MultiTaskSlot) taskSlot))) {...} } }); } return SlotSharingManager.MultiTaskSlotLocality.of(multiTaskSlot, Locality.UNKNOWN); }
  • SchedulerImpl # requestNewAllocatedSlot 会调用 SlotPoolImpl 的同名函数
  • SlotPoolImpl#requestNewAllocatedSlot,这里生成一个 PendingRequest
    • public CompletableFuture<PhysicalSlot> requestNewAllocatedSlot(...) { // 生成 PendingRequest final PendingRequest pendingRequest = PendingRequest.createStreamingRequest(slotRequestId, resourceProfile); // 添加 PendingRequest 到 waitingForResourceManager,然后返回Future return requestNewAllocatedSlotInternal(pendingRequest) .thenApply((Function.identity())); }
    • PendingRequest的构造函数中有 new CompletableFuture<>(),所以这里是生成了第三个 Future,注意这里的 Future 是针对 PhysicalSlot
    • requestNewAllocatedSlotInternal
      • private CompletableFuture<AllocatedSlot> requestNewAllocatedSlotInternal(PendingRequest pendingRequest) { if (resourceManagerGateway == null) { // 就是把 pendingRequest 加到 waitingForResourceManager 之中 stashRequestWaitingForResourceManager(pendingRequest); } else { requestSlotFromResourceManager(resourceManagerGateway, pendingRequest); } return pendingRequest.getAllocatedSlotFuture(); // 第三个Future }
  • SlotSharingManager#createRootSlot,这里才是生成 第二个 Future 的地方
    • MultiTaskSlot createRootSlot( SlotRequestId slotRequestId, CompletableFuture<? extends SlotContext> slotContextFuture, // 参数是第三个Future SlotRequestId allocatedSlotRequestId) { // 生成第二个Future<SlotContext> final CompletableFuture<SlotContext> slotContextFutureAfterRootSlotResolution = new CompletableFuture<>(); final MultiTaskSlot rootMultiTaskSlot = createAndRegisterRootSlot(... slotContextFutureAfterRootSlotResolution); // 第二个Future 在 createAndRegisterRootSlot 函数中 被赋值为 MultiTaskSlot的 slotContextFuture 成员变量 FutureUtils.forward( slotContextFuture.thenApply( // 第三个Future进一步回调时候,会回调第二个Future (SlotContext slotContext) -> { // add the root node to the set of resolved root nodes once the SlotContext future has // been completed and we know the slot's TaskManagerLocation tryMarkSlotAsResolved(slotRequestId, slotContext); return slotContext; }), slotContextFutureAfterRootSlotResolution); // 在这里回调第二个Future return rootMultiTaskSlot; }
  • SlotSharingManager#allocateSingleTaskSlot,这里的目的是在 rootMultiTaskSlot 之上生成一个 SingleTaskSlot leaf加入到allTaskSlots。
    • SingleTaskSlot allocateSingleTaskSlot( SlotRequestId slotRequestId, ResourceProfile resourceProfile, AbstractID groupId, Locality locality) { final SingleTaskSlot leaf = new SingleTaskSlot( slotRequestId, resourceProfile, groupId, this, locality); children.put(groupId, leaf); // register the newly allocated slot also at the SlotSharingManager allTaskSlots.put(slotRequestId, leaf); reserveResource(resourceProfile); return leaf; }
  • 最后回到 SchedulerImpl # allocateSharedSlot 函数,return leaf.getLogicalSlotFuture(); 这里也是一个难点,即 getLogicalSlotFuture 返回的是一个 CompletableFuture(就是第二个 Future),但是这个 SingleLogicalSlot 是未来回调时候才会生成。
    • public final class SingleTaskSlot extends TaskSlot { private final MultiTaskSlot parent; // future containing a LogicalSlot which is completed once the underlying SlotContext future is completed private final CompletableFuture<SingleLogicalSlot> singleLogicalSlotFuture; private SingleTaskSlot() { singleLogicalSlotFuture = parent.getSlotContextFuture() .thenApply( (SlotContext slotContext) -> { return new SingleLogicalSlot( // 未来回调时候才会生成 slotRequestId, slotContext, slotSharingGroupId, locality, slotOwner); }); } CompletableFuture<LogicalSlot> getLogicalSlotFuture() { return singleLogicalSlotFuture.thenApply(Function.identity()); } }

0x06 Deploy阶段

注意,此处的入口为 allocateSlotsAndDeploy函数中 的 deployIndividually / waitForAllSlotsAndDeploy 语句

此处执行路径大致如下:

  • DefaultScheduler#allocateSlotsAndDeploy
  • DefaultScheduler#allocateSlots;得到 SlotExecutionVertexAssignment 列表,上节已经详细介绍(该过程会ExecutionVertex转化为ExecutionVertexSchedulingRequirements,会封装包含一些location信息、sharing信息、资源信息等)
  • List deploymentHandles = createDeploymentHandles() 根据SlotExecutionVertexAssignment建立DeploymentHandle
  • DefaultScheduler#deployIndividually 根据deploymentHandles进行部署,其实是根据Future把部署搭建起来,具体如何部署需要在slot分配成功之后再执行。我们小节实际是从这里开始分析,具体代码可以看出,取出了 Future 1 进行一些列操作
    • private void deployIndividually(final List<DeploymentHandle> deploymentHandles) { for (final DeploymentHandle deploymentHandle : deploymentHandles) { FutureUtils.assertNoException( deploymentHandle .getSlotExecutionVertexAssignment() .getLogicalSlotFuture() .handle(assignResourceOrHandleError(deploymentHandle)) .handle(deployOrHandleError(deploymentHandle))); } }
  • DefaultScheduler#assignResourceOrHandleError;就是返回函数,以备后续回调使用
    • private BiFunction<LogicalSlot, Throwable, Void> assignResourceOrHandleError(final DeploymentHandle deploymentHandle) { final ExecutionVertexVersion requiredVertexVersion = deploymentHandle.getRequiredVertexVersion(); final ExecutionVertexID executionVertexId = deploymentHandle.getExecutionVertexId(); return (logicalSlot, throwable) -> { if (throwable == null) { final ExecutionVertex executionVertex = getExecutionVertex(executionVertexId); final boolean sendScheduleOrUpdateConsumerMessage = deploymentHandle.getDeploymentOption().sendScheduleOrUpdateConsumerMessage(); executionVertex .getCurrentExecutionAttempt() .registerProducedPartitions(logicalSlot.getTaskManagerLocation(), sendScheduleOrUpdateConsumerMessage); executionVertex.tryAssignResource(logicalSlot); } else { handleTaskDeploymentFailure(executionVertexId, maybeWrapWithNoResourceAvailableException(throwable)); } return null; }; }
  • deployOrHandleError 就是返回函数,以备后续回调使用
    • private BiFunction<Object, Throwable, Void> deployOrHandleError(final DeploymentHandle deploymentHandle) { final ExecutionVertexVersion requiredVertexVersion = deploymentHandle.getRequiredVertexVersion(); final ExecutionVertexID executionVertexId = requiredVertexVersion.getExecutionVertexId(); return (ignored, throwable) -> { if (throwable == null) { deployTaskSafe(executionVertexId); } else { handleTaskDeploymentFailure(executionVertexId, throwable); } return null; }; }

0x07 RM分配资源

之前的工作基本都是在 JM 之中。通过 Scheduler 和 SlotPool 来完成申请资源和部署阶段。目前 SlotPool 之中已经积累了一个 PendingRequest,等 SlotPool 连接上 RM,就可以开始向 RM 申请资源了。

当ResourceManager收到申请slot请求时,若发现该JobManager未注册,则直接抛出异常;否则将请求转发给SlotManager处理,SlotManager中维护了集群所有空闲的slot(TaskManager会向ResourceManager上报自己的信息,在ResourceManager中由SlotManager保存Slot和TaskManager对应关系),并从其中找出符合条件的slot,然后向TaskManager发送RPC请求申请对应的slot。

代码执行路径如下:

  • JobMaster # establishResourceManagerConnection 程序执行在 JM 之中
  • SlotPoolImpl # connectToResourceManager
  • SlotPoolImpl # requestSlotFromResourceManager,这里 Pool 会向 RM 进行 RPC 请求。
    • private void requestSlotFromResourceManager( final ResourceManagerGateway resourceManagerGateway, final PendingRequest pendingRequest) { // 生成一个 AllocationID,这个会传到 TM 那里,注册到 TaskSlot上。 final AllocationID allocationId = new AllocationID(); // 生成一个SlotRequest,并且向 RM 进行 RPC 请求。 CompletableFuture<Acknowledge> rmResponse = resourceManagerGateway.requestSlot( jobMasterId, new SlotRequest(jobId, allocationId, pendingRequest.getResourceProfile(), jobManagerAddress), rpcTimeout); }
  • RPC
  • ResourceManager # requestSlot 程序切换到 RM 之中
  • SlotManagerImpl # registerSlotRequest。registerSlotRequest方法会先执行checkDuplicateRequest判断是否有重复,没有重复的话,则将该slotRequest维护到pendingSlotRequests,然后调用internalRequestSlot进行分配,如果出现异常则从pendingSlotRequests中异常,然后抛出SlotManagerException。
    • pendingSlotRequests.put
  • SlotManagerImpl # internalRequestSlot
  • SlotManagerImpl # findMatchingSlot
  • SlotManagerImpl # internalAllocateSlot,此时是没有资源的,需要向 TM 要求资源
    • private void internalRequestSlot(PendingSlotRequest pendingSlotRequest) throws ResourceManagerException { final ResourceProfile resourceProfile = pendingSlotRequest.getResourceProfile(); OptionalConsumer.of(findMatchingSlot(resourceProfile)) .ifPresent(taskManagerSlot -> allocateSlot(taskManagerSlot, pendingSlotRequest)) .ifNotPresent(() -> fulfillPendingSlotRequestWithPendingTaskManagerSlot(pendingSlotRequest)); }
  • SlotManagerImpl # allocateSlot,向task manager要求资源。TaskExecutorGateway接口用来通过RPC分配任务槽,或者说分配任务的资源。
    • TaskExecutorGateway gateway = taskExecutorConnection.getTaskExecutorGateway(); CompletableFuture<Acknowledge> requestFuture = gateway.requestSlot( slotId, pendingSlotRequest.getJobId(), allocationId, pendingSlotRequest.getResourceProfile(), pendingSlotRequest.getTargetAddress(), resourceManagerId, taskManagerRequestTimeout);
  • RPC
  • TaskExecutor # requestSlot,程序切换到 TE
  • TaskSlotTableImpl # allocateSlot,分配资源,更新task slot map,把slot加入到 set of job slots 中。
    • public boolean allocateSlot(int index, JobID jobId, AllocationID allocationId, ResourceProfile resourceProfile,Time slotTimeout) { taskSlot = new TaskSlot<>(index, resourceProfile, memoryPageSize, jobId, allocationId); taskSlots.put(index, taskSlot); allocatedSlots.put(allocationId, taskSlot); slots.add(allocationId); }

0x08 Offer资源阶段

此阶段是由 TE,TM 开始,就是TE 向 RM 提供 Slot,然后 RM 通知 JM 可以运行 Job。也可以认为这部分是从底向上的执行。

等待所有的slot申请完成后,然后会将ExecutionVertex对应的Execution分配给对应的Slot,即从Slot中分配对应的资源给Execution,完成分配后可开始部署作业。

这里两个关键点是:

  • 当 JM 收到 SlotOffer时候,就会根据 RPC传递过来的 taskManagerId 参数,构建一个 taskExecutorGateway,然后这个 taskExecutorGateway 被赋予为 AllocatedSlot . taskManagerGateway。这样就把 JM 范畴的 Slot 和 Slot 所在的 taskManager 联系起来
  • Execution 部署时候,是 从 SingleLogicalSlot ---> AllocatedSlot ---> TaskManagerGateway 这个顺序获取了 TaskManager 的 RPC 网关,然后通过 taskManagerGateway.submitTask 才能提交任务的。这样就把 Execution 部署阶段和执行阶段联系起来了
---------- Task Executor ----------
       │ 
       │ 
┌─────────────┐   
│  TaskSlot   │  requestSlot
└─────────────┘     
       │ 
       │                  
┌──────────────┐   
│  SlotOffer   │  offerSlotsToJobManager
└──────────────┘       
       │ 
       │      
------------- Job Manager -------------
       │ 
       │       
┌──────────────┐   
│  SlotOffer   │  JobMaster#offerSlots(taskManagerId,slots)
└──────────────┘     
       │ //taskManager = registeredTaskManagers.get(taskManagerId);     
       │ //taskManagerLocation = taskManager.f0;     
       │ //taskExecutorGateway = taskManager.f1;     
       │    
       │       
┌──────────────┐   
│  SlotOffer   │  SlotPoolImpl#offerSlots
└──────────────┘       
       │ 
       │      
┌───────────────┐   
│ AllocatedSlot │  SlotPoolImpl#offerSlot
└───────────────┘      
       │ 
       │      
┌───────────────┐   
│ 回调 Future 3  │ SlotSharingManager#createRootSlot
└───────────────┘      
       │ 
       │      
┌───────────────┐   
│ 回调 Future 2  │  SingleTaskSlot#SingleTaskSlot 
└───────────────┘      
       │ 
       │      
┌───────────────────┐   
│ SingleLogicalSlot │ new SingleLogicalSlot
└───────────────────┘    
       │ 
       │     
┌───────────────────┐   
│ SingleLogicalSlot │  
│ 回调 Future 1      │ allocationResultFuture.complete()
└───────────────────┘   
       │    
       │        
┌───────────────────────────────┐  
│     SingleLogicalSlot         │  
│回调 assignResourceOrHandleError│ 
└───────────────────────────────┘
       │    
       │        
┌────────────────┐   
│ ExecutionVertex│ tryAssignResource
└────────────────┘    
       │    
       │        
┌────────────────┐   
│    Execution   │ tryAssignResource
└────────────────┘       
       │    
       │        
┌──────────────────┐   
│ SingleLogicalSlot│ tryAssignPayload
└──────────────────┘  
       │    
       │        
┌───────────────────────┐   
│   SingleLogicalSlot   │      
│ 回调deployOrHandleError│ 
└───────────────────────┘   
       │    
       │       
┌────────────────┐   
│ ExecutionVertex│ deploy
└────────────────┘    
       │    
       │        
┌────────────────┐   
│    Execution   │ deploy // 关键点
└────────────────┘        
       │  
       │    
       │        
 ---------- Task Executor ----------
       │    
       │     
┌────────────────┐   
│  TaskExecutor  │ submitTask
└────────────────┘     
       │    
       │        
┌────────────────┐   
│  TaskExecutor  │ startTaskThread
└────────────────┘         

执行路径如下:

  • TaskExecutor # establishJobManagerConnection
  • TaskExecutor # offerSlotsToJobManager,这里就是遍历已经分配的TaskSlot,然后每个TaskSlot会生成一个SlotOffer(里面是allocationId,slotIndex,resourceProfile),这个会通过RPC发给 JM。
    • private void offerSlotsToJobManager(final JobID jobId) { final Iterator<TaskSlot<Task>> reservedSlotsIterator = taskSlotTable.getAllocatedSlots(jobId); final JobMasterId jobMasterId = jobManagerConnection.getJobMasterId(); final Collection<SlotOffer> reservedSlots = new HashSet<>(2); while (reservedSlotsIterator.hasNext()) { SlotOffer offer = reservedSlotsIterator.next().generateSlotOffer(); reservedSlots.add(offer); } // 把 SlotOffer 通过RPC发给 JM CompletableFuture<Collection<SlotOffer>> acceptedSlotsFuture = jobMasterGateway.offerSlots( getResourceID(), reservedSlots, taskManagerConfiguration.getTimeout()); }
  • RPC
  • JobMaster # offerSlots 。程序执行到 JM。当 JM 收到 SlotOffer时候,就会根据 RPC传递过来的 taskManagerId 参数,构建一个 taskExecutorGateway,然后这个 taskExecutorGateway 被赋予为 AllocatedSlot . taskManagerGateway。这样就把 JM 范畴的 Slot 和 Slot 所在的 taskManager 联系起来
    • public CompletableFuture<Collection<SlotOffer>> offerSlots( final ResourceID taskManagerId, final Collection<SlotOffer> slots, final Time timeout) { Tuple2<TaskManagerLocation, TaskExecutorGateway> taskManager = registeredTaskManagers.get(taskManagerId); final TaskManagerLocation taskManagerLocation = taskManager.f0; final TaskExecutorGateway taskExecutorGateway = taskManager.f1; final RpcTaskManagerGateway rpcTaskManagerGateway = new RpcTaskManagerGateway(taskExecutorGateway, getFencingToken()); return CompletableFuture.completedFuture( slotPool.offerSlots( taskManagerLocation, rpcTaskManagerGateway, slots)); }
  • SlotPoolImpl # offerSlots
  • SlotPoolImpl # offerSlot,这里根据 SlotOffer 的信息生成一个 AllocatedSlot,对于 AllocatedSlot 来说,有效信息就是 slotIndex, resourceProfile。提醒,AllocatedSlot implements PhysicalSlot。
    • boolean offerSlot( final TaskManagerLocation taskManagerLocation, final TaskManagerGateway taskManagerGateway, final SlotOffer slotOffer) { // 根据 SlotOffer 的信息生成一个 AllocatedSlot,对于 AllocatedSlot 来说,有效信息就是 slotIndex, resourceProfile final AllocatedSlot allocatedSlot = new AllocatedSlot( allocationID, taskManagerLocation, slotOffer.getSlotIndex(), slotOffer.getResourceProfile(), taskManagerGateway); allocatedSlots.add(pendingRequest.getSlotRequestId(), allocatedSlot); if (pendingRequest != null) { allocatedSlots.add(pendingRequest.getSlotRequestId(), allocatedSlot); // 这里取出了 pendingRequest 的 Future, 就是我们之前的 Future 3,进行回调 if (!pendingRequest.getAllocatedSlotFuture().complete(allocatedSlot)) { // we could not complete the pending slot future --> try to fulfill another pending request allocatedSlots.remove(pendingRequest.getSlotRequestId()); tryFulfillSlotRequestOrMakeAvailable(allocatedSlot); } } }
  • 开始回调 Future 3,代码在 SlotSharingManager # createRootSlot 这里
    • FutureUtils.forward( slotContextFuture.thenApply( (SlotContext slotContext) -> { // add the root node to the set of resolved root nodes once the SlotContext future has // been completed and we know the slot's TaskManagerLocation tryMarkSlotAsResolved(slotRequestId, slotContext); // 运行到这里 return slotContext; }), slotContextFutureAfterRootSlotResolution); // 然后到这里
  • 开始回调 Future 2,代码在 SingleTaskSlot 构造函数 ,因为有 PhysicalSlot extends SlotContext, 所以这里就把 物理Slot 映射成了一个 逻辑Slot
    • singleLogicalSlotFuture = parent.getSlotContextFuture() .thenApply( (SlotContext slotContext) -> { return new SingleLogicalSlot( // 回调生成了 SingleLogicalSlot slotRequestId, slotContext, slotSharingGroupId, locality, slotOwner); });
  • 开始回调 Future 1,代码在这里,调用到 后续 Deploy 时候设置的回调函数
    • allocationFuture.whenComplete((LogicalSlot slot, Throwable failure) -> { if (failure != null) { cancelSlotRequest( slotRequestId, scheduledUnit.getSlotSharingGroupId(), failure); allocationResultFuture.completeExceptionally(failure); } else { allocationResultFuture.complete(slot); // 代码在这里 } });
  • 继续回调到 Deploy 阶段设置的回调函数 assignResourceOrHandleError,就是分配资源
    • private BiFunction<LogicalSlot, Throwable, Void> assignResourceOrHandleError(final DeploymentHandle deploymentHandle) { return (logicalSlot, throwable) -> { if (executionVertexVersioner.isModified(requiredVertexVersion)) { if (throwable == null) { final ExecutionVertex executionVertex = getExecutionVertex(executionVertexId); final boolean sendScheduleOrUpdateConsumerMessage = deploymentHandle.getDeploymentOption().sendScheduleOrUpdateConsumerMessage(); executionVertex .getCurrentExecutionAttempt() .registerProducedPartitions(logicalSlot.getTaskManagerLocation(), sendScheduleOrUpdateConsumerMessage); executionVertex.tryAssignResource(logicalSlot); // 运行到这里 } return null; }; }
    • 回调函数会深入调用 executionVertex.tryAssignResource,
    • ExecutionVertex # tryAssignResource
    • Execution # tryAssignResource
    • SingleLogicalSlot# tryAssignPayload(this),这里会把 Execution 自己 赋值给Slot.payload,最后 Execution 在 runtime 的变量举例如下:
      • payload = {Execution@10669} "Attempt #0 (CHAIN DataSource (at getDefaultTextLineDataSet(WordCountData.java:47) (org.apache.flink.api.java.io.CollectionInputFormat)) -> FlatMap (FlatMap at main(WordCount.java:64)) -> Combine (SUM(1), at main(WordCount.java:67) (1/1)) @ org.apache.flink.runtime.jobmaster.slotpool.SingleLogicalSlot@61c7928f - [SCHEDULED]" executor = {ScheduledThreadPoolExecutor@5928} "java.util.concurrent.ScheduledThreadPoolExecutor@6a2c6c71[Running, pool size = 3, active threads = 0, queued tasks = 1, completed tasks = 2]" vertex = {ExecutionVertex@10534} "CHAIN DataSource (at getDefaultTextLineDataSet(WordCountData.java:47) (org.apache.flink.api.java.io.CollectionInputFormat)) -> FlatMap (FlatMap at main(WordCount.java:64)) -> Combine (SUM(1), at main(WordCount.java:67) (1/1)" attemptId = {ExecutionAttemptID@10792} "2f8b6c7297527225ee4c8036c457ba27" globalModVersion = 1 stateTimestamps = {long[9]@10793} attemptNumber = 0 rpcTimeout = {Time@5924} "18000000 ms" partitionInfos = {ArrayList@10794} size = 0 terminalStateFuture = {CompletableFuture@10795} "java.util.concurrent.CompletableFuture@2eb8f94c[Not completed]" releaseFuture = {CompletableFuture@10796} "java.util.concurrent.CompletableFuture@7c794914[Not completed]" taskManagerLocationFuture = {CompletableFuture@10797} "java.util.concurrent.CompletableFuture@2e11ac18[Not completed]" state = {ExecutionState@10789} "SCHEDULED" assignedResource = {SingleLogicalSlot@10507} failureCause = null taskRestore = null assignedAllocationID = null accumulatorLock = {Object@10798} userAccumulators = null ioMetrics = null producedPartitions = {LinkedHashMap@10799} size = 1
  • 继续回调到 Deploy 阶段设置的回调函数 deployOrHandleError,就是部署
    • private BiFunction<Object, Throwable, Void> deployOrHandleError(final DeploymentHandle deploymentHandle) { return (ignored, throwable) -> { if (executionVertexVersioner.isModified(requiredVertexVersion)) { if (throwable == null) { deployTaskSafe(executionVertexId); // 在这里部署 } else { handleTaskDeploymentFailure(executionVertexId, throwable); } return null; }; }
    • 回调函数深入调用其他函数
    • DefaultScheduler # deployTaskSafe
    • ExecutionVertex # deploy
    • Execution # deploy。每次调度ExecutionVertex,都会有一个Execution,在此阶段会将Execution的状态变更为DEPLOYING状态,并且为该ExecutionVertex生成对应的部署描述信息,然后从对应的slot中获取对应的TaskManagerGateway,以便向对应的TaskManager提交Task。其中,ExecutionVertex.createDeploymentDescriptor方法中,包含了从Execution Graph到真正物理执行图的转换。如将IntermediateResultPartition转化成ResultPartition,ExecutionEdge转成InputChannelDeploymentDescriptor(最终会在执行时转化成InputGate)。
      • // 这里一个关键点是:Execution 部署时候,是 从 SingleLogicalSlot ---> AllocatedSlot ---> TaskManagerGateway 这个顺序获取了 TaskManager 的 RPC 网关,然后通过 taskManagerGateway.submitTask 才能提交任务的。这样就把 Execution 部署阶段和执行阶段联系起来了 public void deploy() throws JobException { final TaskDeploymentDescriptor deployment = TaskDeploymentDescriptorFactory .fromExecutionVertex(vertex, attemptNumber) .createDeploymentDescriptor( slot.getAllocationId(), slot.getPhysicalSlotNumber(), taskRestore, producedPartitions.values()); // 这里就是关键点 final TaskManagerGateway taskManagerGateway = slot.getTaskManagerGateway(); // 在这里通过RPC提交task给了TaskManager CompletableFuture.supplyAsync(() -> taskManagerGateway.submitTask(deployment, rpcTimeout), executor).thenCompose(Function.identity()) }
  • TaskExecutor # submitTask, 程序执行到 TE,这就是正式执行了。TaskManager(TaskExecutor)在接收到提交Task的请求后,会经过一些初始化(如从BlobServer拉取文件,反序列化作业和Task信息、LibaryCacheManager等),然后这些初始化的信息会用于生成Task(Runnable对象),然后启动该Task,其代码调用路径如下 Task#startTaskThread(启动Task线程)-> Task#run(将ExecutionVertex状态变更为RUNNING状态,此时在FLINK web前台查看顶点状态会变更为RUNNING状态,另外还会生成了一个AbstractInvokable对象,该对象是FLINK衔接执行用户代码的关键。
    • // 这个方法会创建真正的Task,然后调用task.startTaskThread();开始task的执行。 public CompletableFuture<Acknowledge> submitTask( TaskDeploymentDescriptor tdd, JobMasterId jobMasterId, Time timeout) { // taskSlot.getMemoryManager(); 会获取slot的内存管理器,这里就是分割内存的部分功能 memoryManager = taskSlotTable.getTaskMemoryManager(tdd.getAllocationId()); // 在Task构造函数中,会根据输入的参数,创建InputGate, ResultPartition, ResultPartitionWriter等。 Task task = new Task( jobInformation, taskInformation, tdd.getExecutionAttemptId(), tdd.getAllocationId(), tdd.getSubtaskIndex(), tdd.getAttemptNumber(), tdd.getProducedPartitions(), tdd.getInputGates(), tdd.getTargetSlotNumber(), memoryManager, taskExecutorServices.getIOManager(), taskExecutorServices.getShuffleEnvironment(), taskExecutorServices.getKvStateService(), taskExecutorServices.getBroadcastVariableManager(), taskExecutorServices.getTaskEventDispatcher(), taskStateManager, taskManagerActions, inputSplitProvider, checkpointResponder, aggregateManager, blobCacheService, libraryCache, fileCache, taskManagerConfiguration, taskMetricGroup, resultPartitionConsumableNotifier, partitionStateChecker, getRpcService().getExecutor()); taskAdded = taskSlotTable.addTask(task); task.startTaskThread(); }
    • 开始了线程了。而startTaskThread方法,则会执行executingThread.start,从而调用Task.run方法。
      • public void startTaskThread() { executingThread.start(); }
  • 最后会执行到 Task,就是调用用户代码。这里的invokable即为operator对象实例,通过反射创建。具体地,即为OneInputStreamTask,或者SourceStreamTask等。以OneInputStreamTask为例,Task的核心执行代码即为OneInputStreamTask.invoke方法,它会调用StreamTask.run方法,这是个抽象方法,最终会调用其派生类的run方法,即OneInputStreamTask, SourceStreamTask等。
    • // 这里的invokable即为operator对象实例,通过反射创建。 private void doRun() { AbstractInvokable invokable = null; invokable = loadAndInstantiateInvokable(userCodeClassLoader, nameOfInvokableClass, env); // run the invokable invokable.invoke(); }
  • tryFulfillSlotRequestOrMakeAvailable

0x09 Slot发挥作用

有人可能有一个疑问:Slot分配之后,在运行时候怎么发挥作用呢?

这里我们就用WordCount示例来看看。

示例代码就是WordCount。只不过做了一些配置:

  • taskmanager.numberOfTaskSlots 是为了设置有几个taskmanager。
  • 其他是为了调试,加长了心跳时间或者超时时间。
public class WordCount {

    public static void main(String[] args) throws Exception {

        Configuration conf = new Configuration();
        conf.setString("heartbeat.timeout", "18000000");
        conf.setString("resourcemanager.job.timeout", "18000000");
        conf.setString("resourcemanager.taskmanager-timeout", "18000000");
        conf.setString("slotmanager.request-timeout", "18000000");
        conf.setString("slotmanager.taskmanager-timeout", "18000000");
        conf.setString("slot.request.timeout", "18000000");
        conf.setString("slot.idle.timeout", "18000000");
        conf.setString("akka.ask.timeout", "18000000");
        conf.setString("taskmanager.numberOfTaskSlots", "1");

        final LocalEnvironment env = ExecutionEnvironment.createLocalEnvironment(conf);
        final MultipleParameterTool params = MultipleParameterTool.fromArgs(args);
        env.getConfig().setGlobalJobParameters(params);

        // get input data
        DataSet<String> text = null;
        if (params.has("input")) {
            // union all the inputs from text files
            for (String input : params.getMultiParameterRequired("input")) {
                if (text == null) {
                    text = env.readTextFile(input);
                } else {
                    text = text.union(env.readTextFile(input));
                }
            }
        } else {
            // get default test text data
            text = WordCountData.getDefaultTextLineDataSet(env);
        }

        DataSet<Tuple2<String, Integer>> counts =
                // split up the lines in pairs (2-tuples) containing: (word,1)
                text.flatMap(new Tokenizer())
                        // group by the tuple field "0" and sum up tuple field "1"
                        .groupBy(0)
                        .sum(1);

        // emit result
        if (params.has("output")) {
            counts.writeAsCsv(params.get("output"), "n", " ");
            env.execute("WordCount Example");
        } else {
            counts.print();
        }
    }

    // *************************************************************************
    //     USER FUNCTIONS
    // *************************************************************************
    public static final class Tokenizer implements FlatMapFunction<String, Tuple2<String, Integer>> {

        @Override
        public void flatMap(String value, Collector<Tuple2<String, Integer>> out) {
            // normalize and split the line
            String[] tokens = value.toLowerCase().split("\W+");

            // emit the pairs
            for (String token : tokens) {
                if (token.length() > 0) {
                    out.collect(new Tuple2<>(token, 1));
                }
            }
        }
    }
}

9.1 部署阶段

这里 Slot 起到了一个承接作用,把具体提交部署和执行阶段联系起来

前面提到,当TE 提交一个Slot之后,RM会在这个Slot上提交Task。具体逻辑如下:

每次调度ExecutionVertex,都会有一个Execution。在 Execution # deploy 函数中。

  • 会将Execution的状态变更为DEPLOYING状态,并且为该ExecutionVertex生成对应的部署描述信息。其中,ExecutionVertex.createDeploymentDescriptor方法中,包含了从Execution Graph到真正物理执行图的转换。
    • 如将IntermediateResultPartition转化成ResultPartition
    • ExecutionEdge转成InputChannelDeploymentDescriptor(最终会在执行时转化成InputGate)。
  • 然后从对应的slot中获取对应的TaskManagerGateway,以便向对应的TaskManager提交Task。这里一个关键点是:Execution 部署时候,是 从 SingleLogicalSlot ---> AllocatedSlot ---> TaskManagerGateway 这个顺序获取了 TaskManager 的 RPC 网关。
  • 最后通过 taskManagerGateway.submitTask 提交 Task。

具体代码如下:

// 这里一个关键点是:Execution 部署时候,是 从 SingleLogicalSlot ---> AllocatedSlot ---> TaskManagerGateway 这个顺序获取了 TaskManager 的 RPC 网关,然后通过 taskManagerGateway.submitTask 才能提交任务的。这样就把 Execution 部署阶段和执行阶段联系起来了
public void deploy() throws JobException {
			final TaskDeploymentDescriptor deployment = TaskDeploymentDescriptorFactory
			.fromExecutionVertex(vertex, attemptNumber)
				.createDeploymentDescriptor(
				slot.getAllocationId(),
					slot.getPhysicalSlotNumber(),
					taskRestore,
					producedPartitions.values());

   	// 这里就是关键点
  		final TaskManagerGateway taskManagerGateway = slot.getTaskManagerGateway();

      // 在这里通过RPC提交task给了TaskManager
  		CompletableFuture.supplyAsync(() -> taskManagerGateway.submitTask(deployment, rpcTimeout), executor).thenCompose(Function.identity())
}

9.2 运行阶段

这里仅以Split为例子说明,Slot在其中也起到了连接作用,用户从Slot中可以得到其 TaskManager 的host,然后Split会根据这个host继续操作

当 Source 读取输入之后,可能涉及到分割输入,Flink就会进行输入分片的切分。

9.2.1 FileInputSplit 的由来

Flink 一般把文件按并行度拆分成FileInputSplit的个数,当然并不是完全有几个并行度就生成几个FileInputSplit对象,根据具体算法得到,但是FileInputSplit个数,一定是(并行度个数,或者并行度个数+1)。因为计算FileInputSplit个数时,参照物是文件大小 / 并行度 ,如果没有余数,刚好整除,那么FileInputSplit个数一定是并行度,如果有余数,FileInputSplit个数就为是(并行度个数,或者并行度个数+1)。

Flink在生成阶段,会把JobVertex 转化为ExecutionJobVertex,调用new ExecutionJobVertex(),ExecutionJobVertex中存了inputSplits,所以会根据并行并来计算inputSplits的个数。

ExecutionJobVertex 构造函数中有如下代码,这些代码作用是生成 InputSplit,赋值到 ExecutionJobVertex 的成员变量 inputSplits 中,这样就知道了从哪里得倒 Split:

// set up the input splits, if the vertex has any
try {
			InputSplitSource<InputSplit> splitSource = (InputSplitSource<InputSplit>) jobVertex.getInputSplitSource();

			if (splitSource != null) {
				try {
					inputSplits = splitSource.createInputSplits(numTaskVertices);
					if (inputSplits != null) {
						splitAssigner = splitSource.getInputSplitAssigner(inputSplits);
					}
				} 
}
            
// 此时splitSource如下:
splitSource = {CollectionInputFormat@7603} "[To be, or not to be,--that is the question:--, Whether 'tis nobler in the mind to suffer, The slings and arrows of outrageous fortune, ...]"
 serializer = {StringSerializer@7856} 
 dataSet = {ArrayList@7857}  size = 35
 iterator = null
 partitionNumber = 0
 runtimeContext = null    

9.2.2 File Split

这里以网上文章Flink-1.10.0中的readTextFile解读内容为例,给大家看看文件切片大致流程。当然他介绍的是Stream类型。

readTextFile分成两个阶段,一个Source,一个Split Reader。这两个阶段可以分为多个线程,不一定是2个线程。因为Split Reader的并行度时根据配置文件或者启动参数来决定的。

Source的执行流程如下,Source的是用来构建输入切片的,不做数据的读取操作。这里是按照本地运行模式整理的。

Task.run()
 |-- invokable.invoke()
 |    |-- StreamTask.invoke()
 |    |    |-- beforeInvoke()
 |    |    |    |-- init()
 |    |    |    |    |-- SourceStreamTask.init()
 |    |    |    |-- initializeStateAndOpen()
 |    |    |    |    |-- operator.initializeState()
 |    |    |    |    |-- operator.open()
 |    |    |    |    |    |-- SourceStreamTask.LegacySourceFunctionThread.run()
 |    |    |    |    |    |    |-- StreamSource.run()
 |    |    |    |    |    |    |    |-- userFunction.run(ctx)
 |    |    |    |    |    |    |    |    |-- ContinuousFileMonitoringFunction.run()
 |    |    |    |    |    |    |    |    |    |-- RebalancePartitioner.selectChannel()
 |    |    |    |    |    |    |    |    |    |-- RecordWriter.emit()

Split Reader的代码执行流程如下:

Task.run()
 |-- invokable.invoke()
 |    |-- StreamTask.invoke()
 |    |    |-- beforeInvoke()
 |    |    |    |-- init()
 |    |    |    |    |--OneInputStreamTask.init()
 |    |    |    |-- initializeStateAndOpen()
 |    |    |    |    |-- operator.initializeState()
 |    |    |    |    |    |-- ContinuousFileReaderOperator.initializeState()
 |    |    |    |    |-- operator.open()
 |    |    |    |    |    |-- ContinuousFileReaderOperator.open()
 |    |    |    |    |    |    |-- ContinuousFileReaderOperator.SplitReader.run()
 |    |    |-- runMailboxLoop()
 |    |    |    |-- StreamTask.processInput()
 |    |    |    |    |-- StreamOneInputProcessor.processInput()
 |    |    |    |    |    |-- StreamTaskNetworkInput.emitNext() while循环不停的处理输入数据
 |    |    |    |    |    |    |-- ContinuousFileReaderOperator.processElement()
 |    |    |-- afterInvoke()    

9.2.3 Slot的使用

针对本文示例,我们重点介绍Slot在其中的使用。

调用路径如下:

  • DataSourceTask # invoke,此时运行在 TE
  • DataSourceTask # hasNext
    • while (!this.taskCanceled && splitIterator.hasNext())
  • RpcInputSplitProvider # getNextInputSplit
    • CompletableFuture<SerializedInputSplit> futureInputSplit = jobMasterGateway.requestNextInputSplit( jobVertexID, executionAttemptID);
  • RPC
  • 来到 JM
  • JobMaster # requestNextInputSplit
  • SchedulerBase # requestNextInputSplit,这里会从 executionGraph 获取 Execution,然后从 Execution 获取 InputSplit
    • public SerializedInputSplit requestNextInputSplit(JobVertexID vertexID, ExecutionAttemptID executionAttempt) throws IOException { final Execution execution = executionGraph.getRegisteredExecutions().get(executionAttempt); final ExecutionJobVertex vertex = executionGraph.getJobVertex(vertexID); final InputSplit nextInputSplit = execution.getNextInputSplit(); final byte[] serializedInputSplit = InstantiationUtil.serializeObject(nextInputSplit); return new SerializedInputSplit(serializedInputSplit); }
    • 这里 execution.getNextInputSplit() 就会调用 Slot,可以看到,这里先获取Slot,然后从Slot获取其 TaskManager 的host。再从 Vertiex 获取 InputSplit
      • public InputSplit getNextInputSplit() { final LogicalSlot slot = this.getAssignedResource(); final String host = slot != null ? slot.getTaskManagerLocation().getHostname() : null; return this.vertex.getNextInputSplit(host); }
      • public InputSplit getNextInputSplit(String host) { final int taskId = getParallelSubtaskIndex(); synchronized (inputSplits) { final InputSplit nextInputSplit = jobVertex.getSplitAssigner().getNextInputSplit(host, taskId); if (nextInputSplit != null) { inputSplits.add(nextInputSplit); } return nextInputSplit; } } // runtime 信息如下 inputSplits = {GenericInputSplit[1]@13113} 0 = {GenericInputSplit@13121} "GenericSplit (0/1)" partitionNumber = 0 totalNumberOfPartitions = 1
  • 回到 SchedulerBase # requestNextInputSplit,返回 return new SerializedInputSplit(serializedInputSplit);
  • RPC
  • 返回 算子 Task,TE,获取到了 InputSplit,就可以继续处理输入。
    • final InputSplit split = splitIterator.next(); final InputFormat<OT, InputSplit> format = this.format; // open input format // open还没开始真正的读数据,只是定位,设置当前切片信息(切片的开始位置,切片长度),和定位开始位置。把第一个换行符,分到前一个分片,自己从第二个换行符开始读取数据 format.open(split);

0xFF 参考

一文了解 Apache Flink 的资源管理机制

Flink5:Flink运行架构(Slot和并行度)

Flink Slot详解与Job Execution Graph优化

聊聊flink的slot.request.timeout配置

Apache Flink 源码解析(三)Flink on Yarn (2) Resource Manager

Flink on Yarn模式下的TaskManager个数

Flink on YARN时,如何确定TaskManager数

Flink】Flink作业调度流程分析

Flink原理与实现:如何生成ExecutionGraph及物理执行图

Flink源码走读(一):Flink工程目录

flink分析使用之七任务的启动

flink源码解析3 ExecutionGraph的形成与物理执行

Flink 内部原理之作业与调度

Flink之用户代码生成调度层图结构

3. Flink Slot申请

Flink 任务和调度

Flink的Slot是如何做到平均划分TM内存的?

Flink-1.10.0中的readTextFile解读

Flink1.7.2 Dataset 文件切片计算方式和切片数据读取源码分析

flink任务提交流程分析

Flink Parallelism和Slot理解