Kafka 0.8 sever：源代码High level分析

本文主要介绍了Kafka High level的代码架构和主要的类。

这张图是0.8版本的架构

Boker 架构

1 network layer

Kafka使用NIO自己实现了网络层的代码， 而不是采用netty, mina等第三方的网络框架。从性能上来讲，这一块的代码不是性能的瓶颈。

它采用IO多路复用和多线程下的Reactor模式,主要实现类包括SocketServer, Acceptor, Processor和RequestChannel。

Kafka的服务器由SocketServer实现,它是一个NIO的服务器，线程模型如下：

• 1个Acceptor线程负责处理新连接
• N个Processor线程， 每个processor都有自己的selector，负责从socket中读取请求和发送response
• M个Handler线程处理请求，并产生response给processor线程

可以从上面的图形中看到Acceptor, Processor和Handler的功能。

1.1 a. Boker在启动的时候会调用SocketServer的startup方法。

def startup() {
    ......
    for(i <- 0 until numProcessorThreads) {
      processors(i) = new Processor(i,
                                    time,
                                    maxRequestSize,
                                    aggregateIdleMeter,
                                    newMeter("IdlePercent", "percent", TimeUnit.NANOSECONDS, Map("networkProcessor" -> i.toString)),
                                    numProcessorThreads,
                                    requestChannel,
                                    quotas,
                                    connectionsMaxIdleMs)
      Utils.newThread("kafka-network-thread-%d-%d".format(port, i), processors(i), false).start()
    }
    ......
    // start accepting connections
    this.acceptor = new Acceptor(host, port, processors, sendBufferSize, recvBufferSize, quotas)
    Utils.newThread("kafka-socket-acceptor-%s-%d".format(protocol.toString, endpoint.port), acceptor, false).start()
    acceptor.awaitStartup
    ......
  }

1.2 b. 它为每个Processor生成一个线程并启动，然后启动一个Acceptor线程。

Acceptor是一个典型NIO 处理新连接的方法类：

private[kafka] class Acceptor(...) extends AbstractServerThread(connectionQuotas) {
	val serverChannel = openServerSocket(host, port)
	def run() {
		serverChannel.register(selector, SelectionKey.OP_ACCEPT);
		......
		while(isRunning) {
		  val ready = selector.select(500)
		  if(ready > 0) {
			val keys = selector.selectedKeys()
			val iter = keys.iterator()
			while(iter.hasNext && isRunning) {
				......
				accept(key, processors(currentProcessor))
				......
				currentProcessor = (currentProcessor + 1) % processors.length
			}
		  }
		}
		......
	}
}

1.3 c. 它会将新的连接均匀地分配给一个Processor。通过accept方法配置网络参数，并交给processor读写数据。

def accept(key: SelectionKey, processor: Processor) {
    val serverSocketChannel = key.channel().asInstanceOf[ServerSocketChannel]
    val socketChannel = serverSocketChannel.accept()
    try {
      connectionQuotas.inc(socketChannel.socket().getInetAddress)
      socketChannel.configureBlocking(false)
      socketChannel.socket().setTcpNoDelay(true)
      socketChannel.socket().setSendBufferSize(sendBufferSize)

      processor.accept(socketChannel)
    } catch {
      case e: TooManyConnectionsException =>
        info("Rejected connection from %s, address already has the configured maximum of %d connections.".format(e.ip, e.count))
        close(socketChannel)
    }
}

1.4 d. Processor的accept方法将新连接加入它的新连接待处理队列中

在configureNewConnections方法中注册OP_READ。

def accept(socketChannel: SocketChannel) {
	newConnections.add(socketChannel)
	wakeup()
}
private def configureNewConnections() {
	while(newConnections.size() > 0) {
	  val channel = newConnections.poll()
	  debug("Processor " + id + " listening to new connection from " + channel.socket.getRemoteSocketAddress)
	  channel.register(selector, SelectionKey.OP_READ)
	}
  }

1.5 e. Processor线程的主处理逻辑如下， 这是一个死循环，会一直处理这些连接的读写

override def run() {
    startupComplete()
    while (isRunning) {
      try {
        // setup any new connections that have been queued up // 为新连接注册OP_READ
        configureNewConnections()
        // register any new responses for writing
        // 为新的response注册OP_WRITE， 它从requestChannel.receiveResponse(processor's id)读取response
        processNewResponses()
        poll()
        processCompletedReceives()
        processCompletedSends()
        processDisconnected()
      } catch {
        ...
      }
    }

    debug("Closing selector - processor " + id)
    swallowError(closeAll())
    shutdownComplete()
  }

这也是一个标准的NIO的处理代码。

1.6 f. 我们看看read和write是怎么实现的。<这个和0.10的代码对应不上，这个类是修改了的。>

因为Kafka的消息前四个字节代表(一个int)为后续消息的size,所以首先读取size,接着把一个完整的消息读取出来。

如果读取出来一个完整的Request,则将它放到requestChannel中。

具体的Kafka消息的格式可以参考 A Guide To The Kafka Protocol

我们再看看write方法的实现, 直到写完一个response,才讲Ops设为OP_READ,否则一直尝试写。

以上是网络层的主要代码逻辑，主要负责Kafka消息的读写。

2.API layer

API层的主要功能是由KafkaApis类实现的。

根据配置Kafka生成了一组KafkaRequestHandler线程，叫做KafkaRequestHandlerPool:

class KafkaRequestHandlerPool(......) extends Logging with KafkaMetricsGroup {
  ......
  val threads = new Array[Thread](numThreads)
  val runnables = new Array[KafkaRequestHandler](numThreads)
  for(i <- 0 until numThreads) {
    runnables(i) = new KafkaRequestHandler(i, brokerId, aggregateIdleMeter, numThreads, requestChannel, apis)
    threads(i) = Utils.daemonThread("kafka-request-handler-" + i, runnables(i))
    threads(i).start()
  }
  .....
}

KafkaRequestHandler不断的从requestChannel队列里面取出request交给apis处理。

class KafkaRequestHandler(......) extends Runnable with Logging {
   def run() {
    while(true) {
      try {
        var req : RequestChannel.Request = null
        while (req == null) {
          req = requestChannel.receiveRequest(300)
        }
        if(req eq RequestChannel.AllDone) {
          return
        }
        ......
        apis.handle(req)
      } catch {
        ......
      }
    }
  }

}

apis根据不同的请求类型调用不同的方法进行处理。

def handle(request: RequestChannel.Request) {
    try{
      request.requestId match {
        case RequestKeys.ProduceKey => handleProducerRequest(request)
        case RequestKeys.FetchKey => handleFetchRequest(request)
        case RequestKeys.OffsetsKey => handleOffsetRequest(request)
        case RequestKeys.MetadataKey => handleTopicMetadataRequest(request)
        case RequestKeys.LeaderAndIsrKey => handleLeaderAndIsrRequest(request)
        case RequestKeys.StopReplicaKey => handleStopReplicaRequest(request)
        case RequestKeys.UpdateMetadataKey => handleUpdateMetadataRequest(request)
        case RequestKeys.ControlledShutdownKey => handleControlledShutdownRequest(request)
        case RequestKeys.OffsetCommitKey => handleOffsetCommitRequest(request)
        case RequestKeys.OffsetFetchKey => handleOffsetFetchRequest(request)
        case RequestKeys.ConsumerMetadataKey => handleConsumerMetadataRequest(request)
        case RequestKeys.JoinGroupKey => handleJoinGroupRequest(request)
        case RequestKeys.HeartbeatKey => handleHeartbeatRequest(request)
        case requestId => throw new KafkaException("Unknown api code " + requestId)
      }
    } catch {

    } finally
      request.apiLocalCompleteTimeMs = SystemTime.milliseconds
}

显然，此处处理的速度影响Kafka整体的消息处理的速度。

这里我们分析一个处理方法handleProducerRequest。

def handleProducerRequest(request: RequestChannel.Request) {
    val produceRequest = request.body.asInstanceOf[ProduceRequest]
    val numBytesAppended = request.header.sizeOf + produceRequest.sizeOf

    val (authorizedRequestInfo, unauthorizedRequestInfo) = produceRequest.partitionRecords.asScala.partition {
      case (topicPartition, _) => authorize(request.session, Write, new Resource(Topic, topicPartition.topic))
    }

    // the callback for sending a produce response
    def sendResponseCallback(responseStatus: Map[TopicPartition, PartitionResponse]) {
    }

    if (authorizedRequestInfo.isEmpty)
      sendResponseCallback(Map.empty)
    else {
      val internalTopicsAllowed = request.header.clientId == AdminUtils.AdminClientId

      // Convert ByteBuffer to ByteBufferMessageSet
      val authorizedMessagesPerPartition = authorizedRequestInfo.map {
        case (topicPartition, buffer) => (topicPartition, new ByteBufferMessageSet(buffer))
      }

      // call the replica manager to append messages to the replicas
      replicaManager.appendMessages(
        produceRequest.timeout.toLong,
        produceRequest.acks,
        internalTopicsAllowed,
        authorizedMessagesPerPartition,
        sendResponseCallback)

      // if the request is put into the purgatory, it will have a held reference
      // and hence cannot be garbage collected; hence we clear its data here in
      // order to let GC re-claim its memory since it is already appended to log
      produceRequest.clearPartitionRecords()
    }
  }

这里会调用replicaManager.appendMessages处理Kafka message的保存和备份,也就是leader和备份节点上。

3.Replication subsystem

我们进入replicaManager.appendMessages的代码。

这个方法会将消息放到leader分区上，并复制到备份分区上。在超时或者根据required acks的值及时返回response。

def appendMessages(......) {
    if (isValidRequiredAcks(requiredAcks)) {
	  val localProduceResults = appendToLocalLog(internalTopicsAllowed, messagesPerPartition, requiredAcks)
      val produceStatus = localProduceResults.map { case (topicAndPartition, result) =>
        topicAndPartition ->
                ProducePartitionStatus(
                  result.info.lastOffset + 1, // required offset
                  ProducerResponseStatus(result.errorCode, result.info.firstOffset)) // response status
      }
      if (delayedRequestRequired(requiredAcks, messagesPerPartition, localProduceResults)) {
        // create delayed produce operation
        val produceMetadata = ProduceMetadata(requiredAcks, produceStatus)
        val delayedProduce = new DelayedProduce(timeout, produceMetadata, this, responseCallback)
        // create a list of (topic, partition) pairs to use as keys for this delayed produce operation
        val producerRequestKeys = messagesPerPartition.keys.map(new TopicPartitionOperationKey(_)).toSeq
        // try to complete the request immediately, otherwise put it into the purgatory
        // this is because while the delayed produce operation is being created, new
        // requests may arrive and hence make this operation completable.
        delayedProducePurgatory.tryCompleteElseWatch(delayedProduce, producerRequestKeys)
      } else {
        // we can respond immediately
        val produceResponseStatus = produceStatus.mapValues(status => status.responseStatus)
        responseCallback(produceResponseStatus)
      }
    } else {
      // If required.acks is outside accepted range, something is wrong with the client
      // Just return an error and don't handle the request at all
      val responseStatus = messagesPerPartition.map {
        case (topicAndPartition, messageSet) =>
          (topicAndPartition ->
                  ProducerResponseStatus(Errors.INVALID_REQUIRED_ACKS.code,
                    LogAppendInfo.UnknownLogAppendInfo.firstOffset))
      }
      responseCallback(responseStatus)
    }
  }

注意复制是ReplicaFetcherManager通过ReplicaFetcherThread线程完成的。

To publish a message to a partition, the client first finds the leader of the partition from Zookeeper and sends the message to the leader. The leader writes the message to its local log. Each follower constantly pulls new messages from the leader using a single socket channel. That way, the follower receives all messages in the same order as written in the leader. The follower writes each received message to its own log and sends an acknowledgment back to the leader. Once the leader receives the acknowledgment from all replicas in ISR, the message is committed. The leader advances the HW and sends an acknowledgment to the client. For better performance, each follower sends an acknowledgment after the message is written to memory. So, for each committed message, we guarantee that the message is stored in multiple replicas in memory. However, there is no guarantee that any replica has persisted the commit message to disks though. Given that correlated failures are relatively rare, this approach gives us a good balance between response time and durability. In the future, we may consider adding options that provide even stronger guarantees. The leader also periodically broadcasts the HW to all followers. The broadcasting can be piggybacked on the return value of the fetch requests from the followers. From time to time, each replica checkpoints its HW to its disk.

4. Log subsystem

LogManager负责管理Kafka的Log(Kafka消息)， 包括log/Log文件夹的创建，获取和清理。它也会通过定时器检查内存中的log是否要缓存到磁盘中。

重要的类包括LogManager(@threadsafe)和Log。

5.offsetManager

负责管理offset，提供offset的读写。（kafka.client.ClientUtils#channelToOffsetManager）

6.DynamicConfigManager

它负责动态改变Topic\Client的配置属性。（已经改成了kafka.server.DynamicConfigManager,）

如果某个topic的配置属性改变了，Kafka会在ZooKeeper上创建一个类似/kafka10/config/changes/config_change_13321的节点， DynamicConfigManager会监控这些节点， 获得属性改变的topics并处理.

7.其它类

还有一些其它的重要的类， 包括KafkaController, KafkaScheduler,ConsumerCoordinator,KafkaHealthcheck等。

将这四个类都做一遍解析。

二、Metrics

kafka/metrics，Kafka使用metrics进行性能的度量。原先是yammer metrics,现在独立成dropwizard metrics.目前这个框架的package名字比较乱，但是性能监控的功能却是非常的强大。

---

三、Producer

0.8版本（可能是线程安全）

0.8版本的kafka.producer.Producer定义了两种类型的Producer:

• sync
• async
• 基本上都是通过 eventHandler.handle(messages)处理消息, 只不过async会通过一个线程， 以LinkedBlockingQueue为缓冲发送消息。

0.10 版本（线程安全）

The producer is thread safe and sharing a single producer instance across threads will generally be faster than having multiple instances.

1. send方法是异步的，batch.size是缓存partition的大小，每个partition对应一个batch size。
2. linger.ms是Producer用来等待批量数据的到来的时间，在这个时间内，它期待有新的数据到来。我的理解是这是一个超时的时间用来控制发送条件的。（1. 消息达到batch 2.时间超时）
3. buffer.memory 代表Producer缓存的所有partition的内存大小。这个buffer是Producer端的，如果填满，将阻塞这么长的时间max.block.ms。超过这个时间，如果buffer仍然是full的，将抛出TimeoutExeception

证实：0.10版本Producer send只有异步的，没有同步的方法。就算在send之后里面调用get()方法，也就是模拟阻塞，并不是同步。（区分好 异步阻塞、异步非阻塞）

---

四、Consumer

0.8版本（可能，线程安全）

kafka.consumer.SimpleConsumer提供了Simple Consumer API.它通过一个BlockingChannel发送消息，接收Response完成任务。

kafka.javaapi.consumer.SimpleConsumer则提供了java接口。

High level consumer实际由ZookeeperConsumerConnector完成，它将consumer信息记录在zookeeper中，提供KafkaStream获取Kafka消息。

OLS中对Consumer的使用是一个Consumer示例，创建了多线程去读取Consumer的List<KafkaStream<byte[], byte[]>的。

0.10版本（线程不安全）

Consumer 通过方法poll 拉取数据。提交Consumer Offset的方法如下：（后面两种是手动管理的Offset）

• 自动提交
• 阻塞提交：commitSync
• 非阻塞提交：commitAsync

官方的代码 还没看文档。下次看看 2016-11-08