Akka源码分析-Cluster-ClusterClient

  ClusterClient可以与某个集群通信,而本身节点不必是集群的一部分。它只需要知道一个或多个节点的位置作为联系节点。它会跟ClusterReceptionist 建立连接,来跟集群中的特定节点发送消息。而且必须把provider改成remote或cluster。receptionist需要在集群所有节点或一组节点内启动,它可以自行启动或通过ClusterReceptionist 扩展来启动。ClusterClient可以进行通信的actor必须是通过ClusterReceptionis扩展注册过的actor。

  看到这里,你是不是想骂人了,这么简单的功能我都可以自己实现了。不过akka就是这样,一些看似非常简单的功能,框架本身提供的功能更加稳定、更加通用,但性能不一定是最优的。废话不多说,我们来看看ClusterClient的具体实现。

/**
 * This actor is intended to be used on an external node that is not member
 * of the cluster. It acts like a gateway for sending messages to actors
 * somewhere in the cluster. From the initial contact points it will establish
 * a connection to a [[ClusterReceptionist]] somewhere in the cluster. It will
 * monitor the connection to the receptionist and establish a new connection if
 * the link goes down. When looking for a new receptionist it uses fresh contact
 * points retrieved from previous establishment, or periodically refreshed
 * contacts, i.e. not necessarily the initial contact points.
 *
 * You can send messages via the `ClusterClient` to any actor in the cluster
 * that is registered in the [[ClusterReceptionist]].
 * Messages are wrapped in [[ClusterClient.Send]], [[ClusterClient.SendToAll]]
 * or [[ClusterClient.Publish]].
 *
 *  Use the factory method [[ClusterClient#props]]) to create the
 * [[akka.actor.Props]] for the actor.
 *
 * If the receptionist is not currently available, the client will buffer the messages
 * and then deliver them when the connection to the receptionist has been established.
 * The size of the buffer is configurable and it can be disabled by using a buffer size
 * of 0. When the buffer is full old messages will be dropped when new messages are sent
 * via the client.
 *
 * Note that this is a best effort implementation: messages can always be lost due to the distributed
 * nature of the actors involved.
 */
final class ClusterClient(settings: ClusterClientSettings) extends Actor with ActorLogging 

   通过ClusterClient的定义和官方注释来看,就是一个普通的actor,它可以集群中的特定actor(ClusterReceptionist)进行通信。它通过初始的联系点(其实就是ActorPath)与集群内的ClusterReceptionist发消息,同时会监控receptionist的链接状态,以确保链接正常。ClusterClient没有重定义preStart,那就看它的主构造函数吧。

sendGetContacts()
scheduleRefreshContactsTick(establishingGetContactsInterval)
self ! RefreshContactsTick

   分别调用了上面三段代码。

def sendGetContacts(): Unit = {
    val sendTo =
      if (contacts.isEmpty) initialContactsSel
      else if (contacts.size == 1) initialContactsSel union contacts
      else contacts
    if (log.isDebugEnabled)
      log.debug(s"""Sending GetContacts to [${sendTo.mkString(",")}]""")
    sendTo.foreach { _ ! GetContacts }
  }

   sendGetContacts很简单就是给当前的联系点发送GetContacts消息。

def scheduleRefreshContactsTick(interval: FiniteDuration): Unit = {
    refreshContactsTask foreach { _.cancel() }
    refreshContactsTask = Some(context.system.scheduler.schedule(
      interval, interval, self, RefreshContactsTick))
  }

   scheduleRefreshContactsTick启动定时器在interval之后,每隔interval时间,给自己发送RefreshContactsTick消息。

  第三段给自己发送了RefreshContactsTick消息。感觉后面两个代码有点重复,定时器第一个参数直接设置成0不就好了?省略了第三段代码的调用。

case RefreshContactsTick ⇒ sendGetContacts()

   收到RefreshContactsTick消息怎么处理?还是调用sendGetContacts。那请问在主构造函数里面调用sendGetContacts干啥呢?

var contactPaths: HashSet[ActorPath] =
    initialContacts.to[HashSet]
  val initialContactsSel =
    contactPaths.map(context.actorSelection)
  var contacts = initialContactsSel

   initialContactsSel、contacts、contactPaths、initialContacts是不是很相似呢?

  其中initialContactsSel最关键,这是把initialContacts给map成了ActorSelection,同时还给initialContacts发送了Identity消息。ActorPath是远程的actor,怎么select呢?还记得上文说过么?必须把provider配置成remote或者cluster,为啥?你猜。

case ActorIdentity(_, Some(receptionist)) ⇒
        log.info("Connected to [{}]", receptionist.path)
        scheduleRefreshContactsTick(refreshContactsInterval)
        sendBuffered(receptionist)
        context.become(active(receptionist) orElse contactPointMessages)
        connectTimerCancelable.foreach(_.cancel())
        failureDetector.heartbeat()
        self ! HeartbeatTick // will register us as active client of the selected receptionist

   收到ActorIdentity之后调用scheduleRefreshContactsTick重新设置定时器,把缓存的消息发送给receptionist ,修改当前行为变成active。至此就可以通过Send、SendToAll、Publish给集群内特定的actor转发消息了。

def active(receptionist: ActorRef): Actor.Receive = {
    case Send(path, msg, localAffinity) ⇒
      receptionist forward DistributedPubSubMediator.Send(path, msg, localAffinity)
    case SendToAll(path, msg) ⇒
      receptionist forward DistributedPubSubMediator.SendToAll(path, msg)
    case Publish(topic, msg) ⇒
      receptionist forward DistributedPubSubMediator.Publish(topic, msg)
    case HeartbeatTick ⇒
      if (!failureDetector.isAvailable) {
        log.info("Lost contact with [{}], reestablishing connection", receptionist)
        reestablish()
      } else
        receptionist ! Heartbeat
    case HeartbeatRsp ⇒
      failureDetector.heartbeat()
    case RefreshContactsTick ⇒
      receptionist ! GetContacts
    case Contacts(contactPoints) ⇒
      // refresh of contacts
      if (contactPoints.nonEmpty) {
        contactPaths = contactPoints.map(ActorPath.fromString).to[HashSet]
        contacts = contactPaths.map(context.actorSelection)
      }
      publishContactPoints()
    case _: ActorIdentity ⇒ // ok, from previous establish, already handled
    case ReceptionistShutdown ⇒
      if (receptionist == sender()) {
        log.info("Receptionist [{}] is shutting down, reestablishing connection", receptionist)
        reestablish()
      }
  }

   总结下ClusterClient的行为,它通过配置的initialContacts给远程的actor(集群内的ClusterReceptionist)发送ActorSelection消息,然后在收到第一个ActorIdentity消息后,就算联系上了集群。(剩下的ActorIdentity消息被忽略,其实就是最快返回的作为联系点)。定时第一个返回ActorIdentity消息的ClusterReceptionist发送GetContacts消息,获取所有的ClusterReceptionist实例的位置。那如何判断第一个联系点失去联系了呢?看到HeartbeatTick了吗?

val heartbeatTask = context.system.scheduler.schedule(
    heartbeatInterval, heartbeatInterval, self, HeartbeatTick)

   我们刚才忽略了heartbeatTask的定义,其实这是一个定时器,每隔heartbeatInterval秒给自己发送HeartbeatTick消息。其实关于在变量定义过程中写代码,我是不喜欢的,不利于分析源码的啊。

  收到HeartbeatTick消息就给receptionist发送了Heartbeat消息,在收到HeartbeatRsp后更新failureDetector当前的心跳信息。如果failureDetector检测到失败则调用reestablish方法,重新建立链接。

  ClusterClient的源码就分析到这里,下面我们来看看Cluster内的ClusterReceptionist的实现,之前说过,我们可以用actorOf启动或者ClusterReceptionist扩展来启动。当然优先看ClusterReceptionist扩展了啊。

object ClusterClientReceptionist extends ExtensionId[ClusterClientReceptionist] with ExtensionIdProvider {
  override def get(system: ActorSystem): ClusterClientReceptionist = super.get(system)

  override def lookup() = ClusterClientReceptionist

  override def createExtension(system: ExtendedActorSystem): ClusterClientReceptionist =
    new ClusterClientReceptionist(system)
}

   上面是ExtensionId的定义,很显然它还扩展了ExtensionIdProvider,也就是说,通过配置这个Extension就可以启动了,无需代码显式的启动。

/**
 * Extension that starts [[ClusterReceptionist]] and accompanying [[akka.cluster.pubsub.DistributedPubSubMediator]]
 * with settings defined in config section `akka.cluster.client.receptionist`.
 * The [[akka.cluster.pubsub.DistributedPubSubMediator]] is started by the [[akka.cluster.pubsub.DistributedPubSub]] extension.
 */
final class ClusterClientReceptionist(system: ExtendedActorSystem) extends Extension 

   有没有发现关于重要的类,官方注释都很清晰?这个扩展启动ClusterReceptionist和DistributedPubSubMediator,而DistributedPubSubMediator由DistributedPubSub扩展启动,关于DistributedPubSub后面再分析。

/**
   * The [[ClusterReceptionist]] actor
   */
  private val receptionist: ActorRef = {
    if (isTerminated)
      system.deadLetters
    else {
      val name = config.getString("name")
      val dispatcher = config.getString("use-dispatcher") match {
        case "" ⇒ Dispatchers.DefaultDispatcherId
        case id ⇒ id
      }
      // important to use val mediator here to activate it outside of ClusterReceptionist constructor
      val mediator = pubSubMediator
      system.systemActorOf(ClusterReceptionist.props(mediator, ClusterReceptionistSettings(config))
        .withDispatcher(dispatcher), name)
    }
  }

  /**
   * Returns the underlying receptionist actor, particularly so that its
   * events can be observed via subscribe/unsubscribe.
   */
  def underlying: ActorRef =
    receptionist
/**
   * Register the actors that should be reachable for the clients in this [[DistributedPubSubMediator]].
   */
  private def pubSubMediator: ActorRef = DistributedPubSub(system).mediator

  ClusterClientReceptionist定义中有上面源码,非常关键,它启动了一个ClusterReceptionist,其他源码都是注册和注销服务的,我们先忽略。

/**
 * [[ClusterClient]] connects to this actor to retrieve. The `ClusterReceptionist` is
 * supposed to be started on all nodes, or all nodes with specified role, in the cluster.
 * The receptionist can be started with the [[ClusterClientReceptionist]] or as an
 * ordinary actor (use the factory method [[ClusterReceptionist#props]]).
 *
 * The receptionist forwards messages from the client to the associated [[akka.cluster.pubsub.DistributedPubSubMediator]],
 * i.e. the client can send messages to any actor in the cluster that is registered in the
 * `DistributedPubSubMediator`. Messages from the client are wrapped in
 * [[akka.cluster.pubsub.DistributedPubSubMediator.Send]], [[akka.cluster.pubsub.DistributedPubSubMediator.SendToAll]]
 * or [[akka.cluster.pubsub.DistributedPubSubMediator.Publish]] with the semantics described in
 * [[akka.cluster.pubsub.DistributedPubSubMediator]].
 *
 * Response messages from the destination actor are tunneled via the receptionist
 * to avoid inbound connections from other cluster nodes to the client, i.e.
 * the `sender()`, as seen by the destination actor, is not the client itself.
 * The `sender()` of the response messages, as seen by the client, is `deadLetters`
 * since the client should normally send subsequent messages via the `ClusterClient`.
 * It is possible to pass the original sender inside the reply messages if
 * the client is supposed to communicate directly to the actor in the cluster.
 *
 */
final class ClusterReceptionist(pubSubMediator: ActorRef, settings: ClusterReceptionistSettings)
  extends Actor with ActorLogging

   加上我们之前的分析和官方注释,这个actor就很好理解了。ClusterClient就是发送GetContracts消息给这个actor的,ClusterReceptionist在集群内所有节点或一组节点启动。它可以通过ClusterClientReceptionist这个扩展启动,或者作为普通的actor启动(actorOf)。ClusterReceptionist把ClusterClient转发的消息再吃给你信转发给DistributedPubSubMediator或注册的DistributedPubSubMediator(也就是我们注册的Service)。目标actor返回的消息通过DistributedPubSubMediator打的“洞”返回给客户端,其实就是修改了sender。

  这个定义也可以看出,它就是一个非常普通的actor。从源码来看,主构造函数和preStart函数都没有需要特别注意的地方,那就直接看receive喽。

case GetContacts ⇒
      // Consistent hashing is used to ensure that the reply to GetContacts
      // is the same from all nodes (most of the time) and it also
      // load balances the client connections among the nodes in the cluster.
      if (numberOfContacts >= nodes.size) {
        val contacts = Contacts(nodes.map(a ⇒ self.path.toStringWithAddress(a))(collection.breakOut))
        if (log.isDebugEnabled)
          log.debug("Client [{}] gets contactPoints [{}] (all nodes)", sender().path, contacts.contactPoints.mkString(","))
        sender() ! contacts
      } else {
        // using toStringWithAddress in case the client is local, normally it is not, and
        // toStringWithAddress will use the remote address of the client
        val a = consistentHash.nodeFor(sender().path.toStringWithAddress(cluster.selfAddress))
        val slice = {
          val first = nodes.from(a).tail.take(numberOfContacts)
          if (first.size == numberOfContacts) first
          else first union nodes.take(numberOfContacts - first.size)
        }
        val contacts = Contacts(slice.map(a ⇒ self.path.toStringWithAddress(a))(collection.breakOut))
        if (log.isDebugEnabled)
          log.debug("Client [{}] gets contactPoints [{}]", sender().path, contacts.contactPoints.mkString(","))
        sender() ! contacts
      }

   对GetContacts消息的处理我们需要特别关注,毕竟ClusterClient就是发送这个消息来获取集群内service信息的。第一个if语句的注释也很明白,有一个一致性hash来保证所有节点对GetContacts消息的返回都是一致的。

  case msg @ (_: Send | _: SendToAll | _: Publish) ⇒
      val tunnel = responseTunnel(sender())
      tunnel ! Ping // keep alive
      pubSubMediator.tell(msg, tunnel)

   上面就是收到Send、SendToAll、Publish消息的处理逻辑。好像就是把消息发送给了pubSubMediator,这里出现了前面注释中说的“打洞”

def responseTunnel(client: ActorRef): ActorRef = {
    val encName = URLEncoder.encode(client.path.toSerializationFormat, "utf-8")
    context.child(encName) match {
      case Some(tunnel) ⇒ tunnel
      case None ⇒
        context.actorOf(Props(classOf[ClientResponseTunnel], client, responseTunnelReceiveTimeout), encName)
    }
  }

   它在干啥,又创建了一个ClientResponseTunnel这个actor?把这个actor作为service消息的返回者?然后还有一个responseTunnelReceiveTimeout超时时间?

/**
     * Replies are tunneled via this actor, child of the receptionist, to avoid
     * inbound connections from other cluster nodes to the client.
     */
    class ClientResponseTunnel(client: ActorRef, timeout: FiniteDuration) extends Actor with ActorLogging {
      context.setReceiveTimeout(timeout)

      private val isAsk = {
        val pathElements = client.path.elements
        pathElements.size == 2 && pathElements.head == "temp" && pathElements.tail.head.startsWith("$")
      }

      def receive = {
        case Ping ⇒ // keep alive from client
        case ReceiveTimeout ⇒
          log.debug("ClientResponseTunnel for client [{}] stopped due to inactivity", client.path)
          context stop self
        case msg ⇒
          client.tell(msg, Actor.noSender)
          if (isAsk)
            context stop self
      }
    }

   这个actor功能很简单,就是给client转发消息,这尼玛有点太绕了啊。在本地给各个client有创建了一个代理actor啊,返回的消息都通过这个actor返回啊,为啥不直接在服务端就把消息发送给client了呢?其实想想这是非常合理且必要的。有可能service所在的节点,与客户端网络是不通的。或者为了安全管理不能直接通信,通过这个代理回送消息就很必要了。不管怎么样吧,akka的都是对的,akka的都是好的。

case Heartbeat ⇒
      if (verboseHeartbeat) log.debug("Heartbeat from client [{}]", sender().path)
      sender() ! HeartbeatRsp
      updateClientInteractions(sender())

   还有就是对客户端发送的Heartbeat消息的处理,处理逻辑很简单,但有一点需要注意,那就是对客户端列表的一个维护。也就是说在每个ClusterReceptionist都是有客户端列表的。其实吧,这一点我是非常不赞同的。毕竟客户端有可能是海量的,光是维护这个列表就非常耗内存了。弄这个列表虽然功能上非常丰富,但容易造成OOM啊。如果客户端不多,说明akka还没有正式被大家所熟知或者被大公司使用啊。

  好了,ClusterClient就分析到这里了。聪明的读者可能会问,我还没有看到消息是如何通过ClusterReceptionist发送给实际的服务actor啊,pubSubMediator.tell(msg, tunnel)这段代码是如何路由消息的呢?嗯,确实,不过别急,这个会在下一章节(DistributedPubSubMediator)讲解。毕竟官方在ClusterClient的文档中,直接推荐用DistributedPubSubMediator来实现类似的功能。我觉得吧,这又是一个坑,既然你都推荐DistributedPubSubMediator了,还提供ClusterClient模块干啥呢?直接废弃掉啊。

posted @ 2018-08-17 14:36  gabry.wu  阅读(759)  评论(0编辑  收藏  举报