Akka-Cluster（6）- Cluster-Sharding：集群分片，分布式交互程序核心方式

在前面几篇讨论里我们介绍了在集群环境里的一些编程模式、分布式数据结构及具体实现方式。到目前为止，我们已经实现了把程序任务分配给处于很多服务器上的actor，能够最大程度的利用整体系统的硬件资源。这是因为通过akka-cluster能够把很多服务器组合成一个虚拟的整体系统，编程人员不需要知道负责运算的actor具体在那台服务器上运行。当然，我所指的整体系统是一种分布式的系统，实质底层还是各集群节点作为完整个体独立运行的，所以核心理念还是需要将程序分割成能独立运算的任务，然后分派给可能分布在很多服务器上的actor去运算。在上一篇的cluster-load-balance里我们采用了一种fire-and-forget模式把多项独立任务分配给集群节点上的actor，然后任由它们各自完成运算，中途不做任何交互、控制。这也是一种典型的无内部状态的运算模式。对外界来讲就是开始、完成，中间没有关于运算进展或当前状态的交流需要。但在现实里，很多任务是无法完全进行独立细分的，或者再细分会影响系统效率。比如网上购物网站每个客户的购物车：它记录了客户在网上的所有商品拣选过程，每一个拣选动作都代表更新的购物车状态，直到完成结算。那么在一个可能有几十万用户同时在线购物的网站，保留在内存的购物车状态应该是任何机器都无法容纳的，只有回到传统的数据库模式了，还是要面对无法解决的多并发系统效率问题。这么分析，集群分片技术可能是最好的解决方法了。

简单讲：集群分片技术就是把一堆带唯一标识identifier的actor，即entity分布到集群节点上去。控制程序可以通过唯一ID与entityr进行交互，控制整个运算过程。这样，我们可以把程序分成相对合理的包含多个过程状态的细分任务。这些细分任务是由分布在集群节点上的entity来运算的，产生的状态当然也使用的是各集群节点上的资源，如此解决上面所提到的内存容量问题。akka-cluster提供的actor位置透明化机制能在系统崩溃、增减集群节点时自动重新部署所有的actor以达到负责均衡。而用户通过固定的ID就能联络目标entity，无论它被转移到任何集群节点上。

集群分片由分片管理ShardRegion和分片定位ShardCoordinator共同协作实现，目标是把消息正确传递给指定ID的entity。分片定位负责确定分片所在集群节点，分片管理则对每个集群节点上分片内的entity进行定位。ShardCoordinator是个cluster-singleton，而ShardRegion则必须部署在每个集群节点上。每个分片内的entity必须是一个类型的actor。发给entity的消息内部必须包含分片编号和entity ID。通过从消息中解析位置信息后由ShardCoordinator确定负责传递消息的ShardRegion，相关的ShardRegion按ID把消息发送至目标entity。

每个节点上的ShardRegion是通过下面这个start函数构建的：

  /**
   * Scala API: Register a named entity type by defining the [[akka.actor.Props]] of the entity actor
   * and functions to extract entity and shard identifier from messages. The [[ShardRegion]] actor
   * for this type can later be retrieved with the [[#shardRegion]] method.
   *
   * Some settings can be configured as described in the `akka.cluster.sharding` section
   * of the `reference.conf`.
   *
   * @param typeName the name of the entity type
   * @param entityProps the `Props` of the entity actors that will be created by the `ShardRegion`
   * @param settings configuration settings, see [[ClusterShardingSettings]]
   * @param extractEntityId partial function to extract the entity id and the message to send to the
   *   entity from the incoming message, if the partial function does not match the message will
   *   be `unhandled`, i.e. posted as `Unhandled` messages on the event stream
   * @param extractShardId function to determine the shard id for an incoming message, only messages
   *   that passed the `extractEntityId` will be used
   * @param allocationStrategy possibility to use a custom shard allocation and
   *   rebalancing logic
   * @param handOffStopMessage the message that will be sent to entities when they are to be stopped
   *   for a rebalance or graceful shutdown of a `ShardRegion`, e.g. `PoisonPill`.
   * @return the actor ref of the [[ShardRegion]] that is to be responsible for the shard
   */
  def start(
    typeName:           String,
    entityProps:        Props,
    settings:           ClusterShardingSettings,
    extractEntityId:    ShardRegion.ExtractEntityId,
    extractShardId:     ShardRegion.ExtractShardId,
    allocationStrategy: ShardAllocationStrategy,
    handOffStopMessage: Any): ActorRef = {...}

这个函数登记了名称为typeName类型entity的分片。函数返回ActorRef，说明ShardRegion是在本节点上的一个actor。下面是调用示范：

     ClusterSharding(system).start(
        typeName = Counter.shardName,
        entityProps = Counter.props(),
        settings = ClusterShardingSettings(system),
        extractEntityId = Counter.idExtractor,
        extractShardId = Counter.shardResolver)
...

object Counter {

  trait Command
  case object Increment extends Command
  case object Decrement extends Command
  case object Get extends Command
  case object Stop extends Command

  trait Event
  case class CounterChanged(delta: Int) extends Event

  // Sharding Name
  val shardName: String = "Counter"

  // outside world if he want to send message to sharding should use this message
  case class CounterMessage(id: Long, cmd: Command)

  // id extrator
  val idExtractor: ShardRegion.ExtractEntityId = {
    case CounterMessage(id, msg) => (id.toString, msg)
  }

  // shard resolver
  val shardResolver: ShardRegion.ExtractShardId = {
    case CounterMessage(id, msg) => (id % ).toString
  }

  def props() = Props[Counter]

}

entityProps是ShardRegion用来重构entity的。typeName是用来查找ShardRegion的，如下：

val counterRegion: ActorRef = ClusterSharding(system).shardRegion("Counter")
counterRegion ! Get()

用"Counter"获得ShardRegion的ActorRef后所有本节点的消息都是通过这个ShardRegion actor来定位，转达。所以每个ShardRegion都必须具备消息目的地entity的分片编号及entityID的解析方法：extractShardId和extractEntityId。在有些情况下由于节点角色的关系在某个节点不部署任何entity，但本节点需要向其它节点的entity发送消息，这时需要构建一个中介ProxyOnlyShardRegion：

  /**
   * Java/Scala API: Register a named entity type `ShardRegion` on this node that will run in proxy only mode,
   * i.e. it will delegate messages to other `ShardRegion` actors on other nodes, but not host any
   * entity actors itself. The [[ShardRegion]] actor for this type can later be retrieved with the
   * [[#shardRegion]] method.
   *
   * Some settings can be configured as described in the `akka.cluster.sharding` section
   * of the `reference.conf`.
   *
   * @param typeName the name of the entity type
   * @param role specifies that this entity type is located on cluster nodes with a specific role.
   *   If the role is not specified all nodes in the cluster are used.
   * @param messageExtractor functions to extract the entity id, shard id, and the message to send to the
   *   entity from the incoming message
   * @return the actor ref of the [[ShardRegion]] that is to be responsible for the shard
   */
  def startProxy(
    typeName:         String,
    role:             Optional[String],
    messageExtractor: ShardRegion.MessageExtractor): ActorRef = {...}

还有一个重要问题是如何弃用passivate entity，以释放占用资源。akka-cluster提供的方法是通过定义一个空转时间值idle-timeout，如果空转超出此时间段则可以进行passivate。下面是一段应用示范：两分钟空转就passivate entity

class ABC extends Actor {
...
 // passivate the entity when no activity
  context.setReceiveTimeout(.minutes)

...

override def receive .....

  override def receiveCommand: Receive = {
    case Increment      ⇒ persist(CounterChanged(+))(updateState)
    case Decrement      ⇒ persist(CounterChanged(-))(updateState)
    case Get(_)         ⇒ sender() ! count
    case ReceiveTimeout ⇒ context.parent ! Passivate(stopMessage = Stop)
    case Stop           ⇒ context.stop(self)
  }
/* 或者
  override def unhandled(msg: Any): Unit = msg match {
    case ReceiveTimeout => context.parent ! Passivate(stopMessage = PoisonPill)
    case _              => super.unhandled(msg)
  }
*/
}

又或者通过设定配置来实现自动的passivation：

在配置文件中设定：akka.cluster.sharding.passivate-idle-entity-after = 120 s   // off to disable

下面是官网提供的一个说明passivation-stop-message的示范代码：

trait CounterCommand
case object Increment extends CounterCommand
final case class GetValue(replyTo: ActorRef[Int]) extends CounterCommand

case object Idle extends CounterCommand
case object GoodByeCounter extends CounterCommand

def counter2(shard: ActorRef[ClusterSharding.ShardCommand], entityId: String): Behavior[CounterCommand] = {
  Behaviors.setup { ctx ⇒

    def become(value: Int): Behavior[CounterCommand] =
      Behaviors.receiveMessage[CounterCommand] {
        case Increment ⇒
          become(value + )
        case GetValue(replyTo) ⇒
          replyTo ! value
          Behaviors.same
        case Idle ⇒
          // after receive timeout
          shard ! ClusterSharding.Passivate(ctx.self)
          Behaviors.same
        case GoodByeCounter ⇒
          // the stopMessage, used for rebalance and passivate
          Behaviors.stopped
      }

    ctx.setReceiveTimeout(.seconds, Idle)
    become()
  }
}

sharding.init(Entity(
  typeKey = TypeKey,
  createBehavior = ctx ⇒ counter2(ctx.shard, ctx.entityId))
  .withStopMessage(GoodByeCounter))

实际上是向主管ShardRegion发送Passivation消息，并指定停止方式。

还有必须注意的是如果使用BackoffSupervisor监控entity：必须使用Backoff.OnStop，因为persist异常会直接停掉entity。Backoff.OnStop策略会重构entity(BackoffSupervisedEntity)，再启动。那么如果实施passivation时真的需要停止entity呢？我们可以如下操作：

    case "stop" =>
      context.stop(self)
      context.parent ! PoisonPill

context.parent是BackoffSupervisor，需要同时停掉。

下面我们就设计一个例子来示范集群分片应用。为了更贴近现实，在例子使用了event-sourcing,persistentActor等尚未完整介绍的技术和工具。我会在接着的讨论里介绍它们的原理和使用方式。这个例子模仿一个水果店收银业务：有三台pos机，顾客到任何pos机前录入商品、数量，然后结账。这个示范的主要目的是任何时间如果后端服务器出现故障，正在录入过程中的销售单状态都能得到完整恢复。

我们先看看这个pos前端的源代码：

import akka.actor._
import akka.cluster._
import akka.persistence._
import akka.pattern._
import scala.concurrent.duration._

object POSTerminal {
  case class Fruit(code: String, name: String, price: Double)
  case class Item(fruit: Fruit, qty: Int)

  sealed trait Command {
  }
  case class Checkout(fruit: Fruit, qty: Int) extends Command
  case object ShowTotol extends Command
  case class PayCash(amount: Double) extends Command
  case object Shutdown extends Command

  sealed trait Event {}
  case class ItemScanned(fruit: Fruit, qty: Int) extends Event
  case object Paid extends Event

  case class Items(items: List[Item] = Nil) {
    def itemAdded(evt: Event): Items = evt match {
      case ItemScanned(fruit,qty) =>
        copy( Item(fruit,qty) :: items )   //append item

      case _ => this     //nothing happens
    }
    def billPaid = copy(Nil)     //clear all items
    override def toString = items.reverse.toString()
  }

  def termProps = Props(new POSTerminal())

  //backoff suppervisor  must use onStop mode
  def POSProps: Props = {
    val options = Backoff.onStop(
      childProps = termProps,
      childName = "posterm",
      minBackoff =  second,
      maxBackoff =  seconds,
      randomFactor = 0.20
    )
    BackoffSupervisor.props(options)
  }

}

class POSTerminal extends PersistentActor with ActorLogging {
  import POSTerminal._
  val cluster = Cluster(context.system)

  // self.path.parent.name is the type name (utf-8 URL-encoded)
  // self.path.name is the entry identifier (utf-8 URL-encoded)  but entity has a supervisor
  override def persistenceId: String = self.path.parent.parent.name + "-" + self.path.parent.name

  var currentItems = Items()

  override def receiveRecover: Receive = {
    case evt: Event => currentItems = currentItems.itemAdded(evt)
      log.info(s"*****  ${persistenceId} recovering events ...  ********")
    case SnapshotOffer(_,loggedItems: Items) =>
      log.info(s"*****  ${persistenceId} recovering snapshot ...  ********")
      currentItems = loggedItems
  }

  override def receiveCommand: Receive = {
    case Checkout(fruit,qty) =>
      log.info(s"*********${persistenceId} is scanning item: $fruit, qty: $qty *********")
      persist(ItemScanned(fruit,qty))(evt =>  currentItems = currentItems.itemAdded(evt))

    case ShowTotol =>
      log.info(s"*********${persistenceId} on ${cluster.selfAddress} has current scanned items: *********")
      if (currentItems.items == Nil)
        log.info(s"**********${persistenceId} None transaction found! *********")
      else
        currentItems.items.reverse.foreach (item =>
          log.info(s"*********${persistenceId}: ${item.fruit.name} ${item.fruit.price} X ${item.qty} = ${item.fruit.price * item.qty} *********"))

    case PayCash(amt) =>
      log.info(s"**********${persistenceId} paying $amt to settle ***********")
      persist(Paid) { _ =>
        currentItems = currentItems.billPaid
        saveSnapshot(currentItems)     //no recovery
      }

    //shutdown this node to validate entity relocation and proper state recovery
    case Shutdown =>
      log.info(s"******** node ${cluster.selfAddress} is leaving cluster ... *******")
      cluster.leave(cluster.selfAddress)
  }
}

我用下面几项来总结一下：

1、POSTerminal是具体的业务运算前端，包裹在BackoffSupervisor里。能保证这个entity在因异常如持久化失败造成停顿时能进行重试。所以，使用了Backoff.onStop方式。

2、persistenceId=self.path.parent.parent.name+"-"+self.path.parent.name 代表: 店号-机号 如： 1-1021。actor.path.name的产生是由ShardRegion具体操作的，其实就是ExtactShardId-ExtractEntityId。

3、注意这个状态类型Item，它的方法itemAdded(evt): Item 即返回新状态。所以必须谨记用currentItems=itemAdded(evt)这样的语法。

下面是构建和启动ClusterSharding的源代码：

object POSShard {
 import POSTerminal._

 val shardName = "POSManager"
 case class POSCommand(id: Long, cmd: Command) {
   def shopId = id.toString.head.toString
   def posId = id.toString
 }

 val getPOSId: ShardRegion.ExtractEntityId =  {
   case posCommand: POSCommand => (posCommand.posId,posCommand.cmd)
 }
 val getShopId: ShardRegion.ExtractShardId = {
   case posCommand: POSCommand => posCommand.shopId
 }

 def create(port: Int) = {
   val config = ConfigFactory.parseString(s"akka.remote.netty.tcp.port=$port")
     .withFallback(ConfigFactory.load())
   val system = ActorSystem("posSystem",config)

   ClusterSharding(system).start(
     typeName = shardName,
     entityProps = POSProps,
     settings = ClusterShardingSettings(system),
     extractEntityId = getPOSId,
     extractShardId = getShopId
   )
 }

}

用下面的代码来测试：

object POSDemo extends App {
  POSShard.create()
  Thread.sleep()
  POSShard.create()
  POSShard.create()
  val posref = POSShard.create()
  scala.io.StdIn.readLine()

  val apple = Fruit("","high grade apple",10.5)
  val orange = Fruit("","sunkist orage",12.0)
  val grape = Fruit("","xinjiang red grape",15.8)

  posref ! POSCommand(, Checkout(apple,))
  posref ! POSCommand(,Checkout(grape,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Shutdown)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Checkout(orange,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Checkout(orange,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

}

运算结果如下：

[akka.tcp://posSystem@127.0.0.1:2551*********1-1021 is scanning item: Fruit(0001,high grade apple,10.5), qty: 2 *********
[akka.tcp://posSystem@127.0.0.1:2551*********1-1021 is scanning item: Fruit(0003,xinjiang red grape,15.8), qty: 1 *********
[akka.tcp://posSystem@127.0.0.1:2551*********1-1021 on akka.tcp://posSystem@127.0.0.1:2551 has current scanned items: *********
[akka.tcp://posSystem@127.0.0.1:2551*********1-1021: high grade apple 10.5 X 2 = 21.0 *********
[akka.tcp://posSystem@127.0.0.1:2551*********1-1021: xinjiang red grape 15.8 X 1 = 15.8 *********

[akka.tcp://posSystem@127.0.0.1:2551******** node akka.tcp://posSystem@127.0.0.1:2551 is leaving cluster ... *******
[akka.tcp://posSystem@127.0.0.1:2551/system/remoting-terminator] Remoting shut down.

[akka.tcp://posSystem@127.0.0.1:2552*****  1-1021 recovering events ...  ********
[akka.tcp://posSystem@127.0.0.1:2552*****  1-1021 recovering events ...  ********
[akka.tcp://posSystem@127.0.0.1:2552********1-1021 is scanning item: Fruit(0002,sunkist orage,12.0), qty: 10 *********
[akka.tcp://posSystem@127.0.0.1:2552*********1-1021 on akka.tcp://posSystem@127.0.0.1:2552 has current scanned items: *********
[akka.tcp://posSystem@127.0.0.1:2552*********1-1021: high grade apple 10.5 X 2 = 21.0 *********
[akka.tcp://posSystem@127.0.0.1:2552*********1-1021: xinjiang red grape 15.8 X 1 = 15.8 *********
[akka.tcp://posSystem@127.0.0.1:2552*********1-1021: sunkist orage 12.0 X 10 = 120.0 *********

从结果显示看到：一开始1-1021是在2551节点上运行的。我们用Shutdown关停2551后ClusterSharding立即在2552上重构了1-1021并且恢复了之前的状态。能够在系统出现故障无法使用的情况下自动对运行中的actor进行迁移、状态恢复，正是我们这次讨论的核心内容。

下面是本次示范的源代码：

build.sbt

name := "akka-cluster-sharding"

version := "0.2"

scalaVersion := "2.12.8"

libraryDependencies := Seq(
  "com.typesafe.akka" %% "akka-cluster-sharding" % "2.5.19",
  "com.typesafe.akka" %% "akka-persistence" % "2.5.19",
  "com.typesafe.akka" %% "akka-persistence-cassandra" % "0.92",
  "com.typesafe.akka" %% "akka-persistence-cassandra-launcher" % "0.92" % Test
)

resources/application.conf

akka.actor.warn-about-java-serializer-usage = off
akka.log-dead-letters-during-shutdown = off
akka.log-dead-letters = off

akka {
  loglevel = INFO
  actor {
    provider = "cluster"
  }

  remote {
    log-remote-lifecycle-events = off
    netty.tcp {
      hostname = "127.0.0.1"
      port =
    }
  }

  cluster {
    seed-nodes = [
      "akka.tcp://posSystem@127.0.0.1:2551"]
    log-info = off
  }

  persistence {
    journal.plugin = "cassandra-journal"
    snapshot-store.plugin = "cassandra-snapshot-store"
  }

}

Entities.scala

import akka.actor._
import akka.cluster._
import akka.persistence._
import akka.pattern._
import scala.concurrent.duration._

object POSTerminal {
  case class Fruit(code: String, name: String, price: Double)
  case class Item(fruit: Fruit, qty: Int)

  sealed trait Command {
  }
  case class Checkout(fruit: Fruit, qty: Int) extends Command
  case object ShowTotol extends Command
  case class PayCash(amount: Double) extends Command
  case object Shutdown extends Command

  sealed trait Event {}
  case class ItemScanned(fruit: Fruit, qty: Int) extends Event
  case object Paid extends Event

  case class Items(items: List[Item] = Nil) {
    def itemAdded(evt: Event): Items = evt match {
      case ItemScanned(fruit,qty) =>
        copy( Item(fruit,qty) :: items )   //append item

      case _ => this     //nothing happens
    }
    def billPaid = copy(Nil)     //clear all items
    override def toString = items.reverse.toString()
  }

  def termProps = Props(new POSTerminal())

  //backoff suppervisor  must use onStop mode
  def POSProps: Props = {
    val options = Backoff.onStop(
      childProps = termProps,
      childName = "posterm",
      minBackoff =  second,
      maxBackoff =  seconds,
      randomFactor = 0.20
    )
    BackoffSupervisor.props(options)
  }

}

class POSTerminal extends PersistentActor with ActorLogging {
  import POSTerminal._
  val cluster = Cluster(context.system)

  // self.path.parent.name is the type name (utf-8 URL-encoded)
  // self.path.name is the entry identifier (utf-8 URL-encoded)  but entity has a supervisor
  override def persistenceId: String = self.path.parent.parent.name + "-" + self.path.parent.name

  var currentItems = Items()

  override def receiveRecover: Receive = {
    case evt: Event => currentItems = currentItems.itemAdded(evt)
      log.info(s"*****  ${persistenceId} recovering events ...  ********")
    case SnapshotOffer(_,loggedItems: Items) =>
      log.info(s"*****  ${persistenceId} recovering snapshot ...  ********")
      currentItems = loggedItems
  }

  override def receiveCommand: Receive = {
    case Checkout(fruit,qty) =>
      log.info(s"*********${persistenceId} is scanning item: $fruit, qty: $qty *********")
      persist(ItemScanned(fruit,qty))(evt =>  currentItems = currentItems.itemAdded(evt))

    case ShowTotol =>
      log.info(s"*********${persistenceId} on ${cluster.selfAddress} has current scanned items: *********")
      if (currentItems.items == Nil)
        log.info(s"**********${persistenceId} None transaction found! *********")
      else
        currentItems.items.reverse.foreach (item =>
          log.info(s"*********${persistenceId}: ${item.fruit.name} ${item.fruit.price} X ${item.qty} = ${item.fruit.price * item.qty} *********"))

    case PayCash(amt) =>
      log.info(s"**********${persistenceId} paying $amt to settle ***********")
      persist(Paid) { _ =>
        currentItems = currentItems.billPaid
        saveSnapshot(currentItems)     //no recovery
      }

    //shutdown this node to validate entity relocation and proper state recovery
    case Shutdown =>
      log.info(s"******** node ${cluster.selfAddress} is leaving cluster ... *******")
      cluster.leave(cluster.selfAddress)
  }
}

Shards.scala

import akka.actor._
import akka.cluster.sharding._
import com.typesafe.config.ConfigFactory

object POSShard {
 import POSTerminal._

 val shardName = "POSManager"
 case class POSCommand(id: Long, cmd: Command) {
   def shopId = id.toString.head.toString
   def posId = id.toString
 }

 val getPOSId: ShardRegion.ExtractEntityId =  {
   case posCommand: POSCommand => (posCommand.posId,posCommand.cmd)
 }
 val getShopId: ShardRegion.ExtractShardId = {
   case posCommand: POSCommand => posCommand.shopId
 }

 def create(port: Int) = {
   val config = ConfigFactory.parseString(s"akka.remote.netty.tcp.port=$port")
     .withFallback(ConfigFactory.load())
   val system = ActorSystem("posSystem",config)

   ClusterSharding(system).start(
     typeName = shardName,
     entityProps = POSProps,
     settings = ClusterShardingSettings(system),
     extractEntityId = getPOSId,
     extractShardId = getShopId
   )
 }

}

POSDemo.scala

import POSTerminal._
import POSShard._

object POSDemo extends App {
  POSShard.create()
  Thread.sleep()
  POSShard.create()
  POSShard.create()
  val posref = POSShard.create()
  scala.io.StdIn.readLine()

  val apple = Fruit("","high grade apple",10.5)
  val orange = Fruit("","sunkist orage",12.0)
  val grape = Fruit("","xinjiang red grape",15.8)

  posref ! POSCommand(, Checkout(apple,))
  posref ! POSCommand(,Checkout(grape,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Shutdown)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Checkout(orange,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

  posref ! POSCommand(,Checkout(orange,))

  posref ! POSCommand(,ShowTotol)
  scala.io.StdIn.readLine()

}