Kafka源码之Consumer分析_MQ, Tuxedo及OLTP讨论区_Weblogic技术|Tuxedo技术|中间件技术|Oracle论坛|JAVA论坛|Linux/Unix技术|hadoop论坛_联动北方技术论坛  
网站首页 | 关于我们 | 服务中心 | 经验交流 | 公司荣誉 | 成功案例 | 合作伙伴 | 联系我们 |
联动北方-国内领先的云技术服务提供商
»  游客             当前位置:  论坛首页 »  自由讨论区 »  MQ, Tuxedo及OLTP讨论区 »
总帖数
1
每页帖数
101/1页1
返回列表
0
发起投票  发起投票 发新帖子
查看: 4212 | 回复: 0   主题: Kafka源码之Consumer分析        上一篇   下一篇 
风影子
注册用户
等级:新兵
经验:61
发帖:6
精华:0
注册:2012-3-9
状态:离线
发送短消息息给风影子 加好友    发送短消息息给风影子 发消息
发表于: IP:您无权察看 2018-5-9 10:12:51 | [全部帖] [楼主帖] 楼主

本章我们将重点跟踪Consumer的源码细节。

Consumer的配置文件如下:

Kafka Consumer配置:
group.id: 指定consumer所属的consumer group
consumer.id: 如果不指定会自动生成
socket.timeout.ms: 网络请求的超时设定
socket.receive.buffer.bytes: Socket的接收缓存大小
fetch.message.max.bytes: 试图获取的消息大小之和(bytes)
num.consumer.fetchers: 该消费去获取data的总线程数
auto.commit.enable: 如果是true, 定期向zk中更新Consumer已经获取的last message offset(所获取的最后一个batch的first message offset)
auto.commit.interval.ms: Consumer向ZK中更新offset的时间间隔
queued.max.message.chunks: 默认为2
rebalance.max.retries: 在rebalance时retry的最大次数,默认为4
fetch.min.bytes: 对于一个fetch request, Broker Server应该返回的最小数据大小,达不到该值request会被block, 默认是1字节。
fetch.wait.max.ms: Server在回答一个fetch request之前能block的最大时间(可能的block原因是返回数据大小还没达到fetch.min.bytes规定);
rebalance.backoff.ms: 当rebalance发生时,两个相邻retry操作之间需要间隔的时间。
refresh.leader.backoff.ms: 如果一个Consumer发现一个partition暂时没有leader, 那么Consumer会继续等待的最大时间窗口(这段时间内会refresh partition leader);
auto.offset.reset: 当发现offset超出合理范围(out of range)时,应该设成的大小(默认是设成offsetRequest中指定的值):
smallest: 自动把该consumer的offset设为最小的offset;
largest: 自动把该consumer的offset设为最大的offset;

 anything else: throw exception to the consumer;


consumer.timeout.ms: 如果在该规定时间内没有消息可供消费,则向Consumer抛出timeout exception;
该参数默认为-1, 即不指定Consumer timeout;
client.id: 区分不同consumer的ID,默认是group.id

先从一个消费者的demo开始:

public class ConsumerDemo {
      private final ConsumerConnector consumer;
      private final String topic;
      private ExecutorService executor;
      public ConsumerDemo(String a_zookeeper, String a_groupId, String a_topic) {
            consumer = Consumer.createJavaConsumerConnector(createConsumerConfig(a_zookeeper,a_groupId));
            this.topic = a_topic;
            public void shutdown() {
                  if (consumer != null)
                  consumer.shutdown();
                  if (executor != null)
                  executor.shutdown();
                  public void run(int numThreads) {
                        Map<String, Integer> topicCountMap = new HashMap<String, Integer>();
                        topicCountMap.put(topic, new Integer(numThreads));
                        Map<String, List<KafkaStream<byte[], byte[]>>> consumerMap = consumer
                        .createMessageStreams(topicCountMap);
                        List<KafkaStream<byte[], byte[]>> streams = consumerMap.get(topic);
                        // now launch all the threads
                        executor = Executors.newFixedThreadPool(numThreads);
                        // now create an object to consume the messages
                        int threadNumber = 0;
                        for (final KafkaStream stream : streams) {
                              executor.submit(new ConsumerMsgTask(stream, threadNumber));
                              threadNumber++;
                              private static ConsumerConfig createConsumerConfig(String a_zookeeper,
                              String a_groupId) {
                                    Properties props = new Properties();
                                    props.put("zookeeper.connect", a_zookeeper);
                                    props.put("group.id", a_groupId);
                                    props.put("zookeeper.session.timeout.ms", "400");
                                    props.put("zookeeper.sync.time.ms", "200");
                                    props.put("auto.commit.interval.ms", "1000");
                                    return new ConsumerConfig(props);
                                    public static void main(String[] arg) {
                                    String[] args = { "172.168.63.221:2188", "group-1", "page_visits", "12" };
                                          String zooKeeper = args[0];
                                          String groupId = args[1];
                                          String topic = args[2];
                                          int threads = Integer.parseInt(args[3]);
                                          ConsumerDemo demo = new ConsumerDemo(zooKeeper, groupId, topic);
                                          demo.run(threads);
                                          try {
                                                Thread.sleep(10000);
                                          } catch (InterruptedException ie) {
                                          demo.shutdown();


   上面的例子是用java编写的消费者的例子,也是官网提供的例子,那么我们的源码分析就从下面这一行开始:

Map<String, List<KafkaStream<byte[], byte[]>>> consumerMap = consumer
.createMessageStreams(topicCountMap);


 从createMessagesStreams方法进入后直接到kafka.javaapi.consumer.ZookeeperConsumerConnector类。

private[kafka] class ZookeeperConsumerConnector(val config: ConsumerConfig,
val enableFetcher: Boolean) // for testing only
extends ConsumerConnector {
      //初始化伴生对象
      private val underlying = new kafka.consumer.ZookeeperConsumerConnector(config, enableFetcher)
      private val messageStreamCreated = new AtomicBoolean(false)
      def this(config: ConsumerConfig) = this(config, true)
      // for java client
      def createMessageStreams[K,V](
      topicCountMap: java.util.Map[String,java.lang.Integer],
      keyDecoder: Decoder[K],
      valueDecoder: Decoder[V])
      : java.util.Map[String,java.util.List[KafkaStream[K,V]]] = {
            if (messageStreamCreated.getAndSet(true))
            throw new MessageStreamsExistException(this.getClass.getSimpleName +
            " can create message streams at most once",null)
            val scalaTopicCountMap: Map[String, Int] = {
                  import JavaConversions._
                  Map.empty[String, Int] ++ (topicCountMap.asInstanceOf[java.util.Map[String, Int]]: mutable.Map[String, Int])
                  val scalaReturn = underlying.consume(scalaTopicCountMap, keyDecoder, valueDecoder)
                  val ret = new java.util.HashMap[String,java.util.List[KafkaStream[K,V]]]
                  for ((topic, streams) <- scalaReturn) {
                        var javaStreamList = new java.util.ArrayList[KafkaStream[K,V]]
                        for (stream <- streams)
                        javaStreamList.add(stream)
                        ret.put(topic, javaStreamList)
                        ret
                  }


这个类是整体Consumer的核心类,首先要初始化ZookeeperConsumerConnector的伴生对象(关于伴生对象请大家查看scala语法,实际就是一个静态对象,每一个class都要有一个伴生对象,像我们的静态方法都要定义在这里面),在createMessageStreams中,topicCountMap主要是消费线程数,这个参数和partition的数量有直接有关系。

  通过val scalaReturn = underlying.consume(scalaTopicCountMap, keyDecoder, valueDecoder)这行代码,将进入到伴生对象中,直接可以跟踪消费的内部逻辑。

def consume[K, V](topicCountMap: scala.collection.Map[String,Int], keyDecoder: Decoder[K], valueDecoder: Decoder[V])
: Map[String,List[KafkaStream[K,V]]] = {
      debug("entering consume ")
      if (topicCountMap == null)
      throw new RuntimeException("topicCountMap is null")
      //封装成一个TopicCount对象,参数分别是消费者的ids字符串和线程数map
      val topicCount = TopicCount.constructTopicCount(consumerIdString, topicCountMap)
      //解析出每个topic对应多少个消费者线程,topicThreadsIds是一个map结构
      val topicThreadIds = topicCount.getConsumerThreadIdsPerTopic
      //针对每一个消费者线程创建一个BlockingQueue队列,队列中存储的是FetchedDataChunk数据块,每一个数据块中包括多条记录。
      val queuesAndStreams = topicThreadIds.values.map(threadIdSet =>
      threadIdSet.map(_ => {
            val queue = new LinkedBlockingQueue[FetchedDataChunk](config.queuedMaxMessages)
            val stream = new KafkaStream[K,V](
            queue, config.consumerTimeoutMs, keyDecoder, valueDecoder, config.clientId)
            (queue, stream)
            ).flatten.toList
            val dirs = new ZKGroupDirs(config.groupId)
            //将consumer的topic信息注册到zookeeper中,格式如下:
            //Consumer id registry:/consumers/[group_id]/ids[consumer_id] -> topic1,...topicN
            registerConsumerInZK(dirs, consumerIdString, topicCount)
            reinitializeConsumer(topicCount, queuesAndStreams)
            loadBalancerListener.kafkaMessageAndMetadataStreams.asInstanceOf[Map[String, List[KafkaStream[K,V]]]]
      }


 结合代码中的注释请看下面的图:

   

说明: 

创建consumer thread

consumer thread数量与BlockingQueue一一对应。

a.当consumer thread count=1时

此时有一个blockingQueue1,三个fetch thread线程,该topic分布在几个node上就有几个fetch thread,每个fetch thread会于kafka broker建立一个连接。3个fetch thread线程去拉取消息数据,最终放到blockingQueue1中,等待consumer thread来消费。

接着看上面代码中的这个方法:

registerConsumerInZK(dirs, consumerIdString, topicCount)


这个方法是将consumer的topic信息注册到zookeeper中,格式如下:

Consumer id registry:
/consumers/[group_id]/ids[consumer_id] -> topic1,...topicN


进入重新初始化Consumer方法:

registerConsumerInZK(dirs, consumerIdString, topicCount)


 这个方法会建立一系列的侦听器:

1、负载平衡器侦听器:ZKRebalancerListener。

2、会话超时侦听器:ZKSessionExpireListener。

3、监控topic和partition变化侦听器:ZKTopicPartitionChangeListener。

客户端启动后会在消费者注册目录上添加子节点变化的监听ZKRebalancerListener,ZKRebalancerListener实例会在内部创建一个线程,这个线程定时检查监听的事件有没有执行(消费者发生变化),如果没有变化则wait1秒钟,当发生了变化就调用 syncedRebalance 方法,去rebalance消费者,代码如下:

private val watcherExecutorThread = new Thread(consumerIdString + "_watcher_executor") {
      override def run() {
            info("starting watcher executor thread for consumer " + consumerIdString)
            var doRebalance = false
            while (!isShuttingDown.get) {
                  try {
                        lock.lock()
                        try {
                              if (!isWatcherTriggered)
                              cond.await(1000, TimeUnit.MILLISECONDS) // wake up periodically so that it can check the shutdown flag
                        } finally {
                        doRebalance = isWatcherTriggered
                        isWatcherTriggered = false
                        lock.unlock()
                        if (doRebalance)
                        syncedRebalance
                  } catch {
                  case t: Throwable => error("error during syncedRebalance", t)
                  info("stopping watcher executor thread for consumer " + consumerIdString)
                  watcherExecutorThread.start()
                  @throws(classOf[Exception])
                  def handleChildChange(parentPath : String, curChilds : java.util.List[String]) {
                  rebalanceEventTriggered()
                  def rebalanceEventTriggered() {
                  inLock(lock) {
                  isWatcherTriggered = true
                  cond.signalAll()


 syncedRebalance方法在内部会调用def rebalance(cluster: Cluster): Boolean方法,去真正执行操作。

 在这个方法中,获取者必须停止,避免重复的数据,重新平衡尝试失败,被释放的分区被另一个consumers拥有。如果我们不首先停止获取数据,消费者将继续并发的返回数据,所以要先停止之前的获取者,再更新当前的消费者信息,重新更新启动获取者。代码如下:

private def rebalance(cluster: Cluster): Boolean = {
      val myTopicThreadIdsMap = TopicCount.constructTopicCount(
      group, consumerIdString, zkClient, config.excludeInternalTopics).getConsumerThreadIdsPerTopic
      val brokers = getAllBrokersInCluster(zkClient)
      if (brokers.size == 0) {
            // This can happen in a rare case when there are no brokers available in the cluster when the consumer is started.
            // We log an warning and register for child changes on brokers/id so that rebalance can be triggered when the brokers
            // are up.
            warn("no brokers found when trying to rebalance.")
            zkClient.subscribeChildChanges(ZkUtils.BrokerIdsPath, loadBalancerListener)
            true
            else {
                  /**
                  * fetchers must be stopped to avoid data duplication, since if the current
                  * rebalancing attempt fails, the partitions that are released could be owned by another consumer.
                  * But if we don't stop the fetchers first, this consumer would continue returning data for released
                  * partitions in parallel. So, not stopping the fetchers leads to duplicate data.
                  //在这行要先停止之前的获取者线程,再更新启动当前最新消费者的。
                  closeFetchers(cluster, kafkaMessageAndMetadataStreams, myTopicThreadIdsMap)
                  releasePartitionOwnership(topicRegistry)
                  val assignmentContext = new AssignmentContext(group, consumerIdString, config.excludeInternalTopics, zkClient)
                  val partitionOwnershipDecision = partitionAssignor.assign(assignmentContext)
                  val currentTopicRegistry = new Pool[String, Pool[Int, PartitionTopicInfo]](
                  valueFactory = Some((topic: String) => new Pool[Int, PartitionTopicInfo]))
                  // fetch current offsets for all topic-partitions
                  val topicPartitions = partitionOwnershipDecision.keySet.toSeq
                  val offsetFetchResponseOpt = fetchOffsets(topicPartitions)
                  if (isShuttingDown.get || !offsetFetchResponseOpt.isDefined)
                  false
                  else {
                        val offsetFetchResponse = offsetFetchResponseOpt.get
                        topicPartitions.foreach(topicAndPartition => {
                              val (topic, partition) = topicAndPartition.asTuple
                              val offset = offsetFetchResponse.requestInfo(topicAndPartition).offset
                              val threadId = partitionOwnershipDecision(topicAndPartition)
                              addPartitionTopicInfo(currentTopicRegistry, partition, topic, offset, threadId)
                              /**
                              * move the partition ownership here, since that can be used to indicate a truly successful rebalancing attempt
                              * A rebalancing attempt is completed successfully only after the fetchers have been started correctly
                              if(reflectPartitionOwnershipDecision(partitionOwnershipDecision)) {
                                    allTopicsOwnedPartitionsCount = partitionOwnershipDecision.size
                              partitionOwnershipDecision.view.groupBy { case(topicPartition, consumerThreadId) => topicPartition.topic }
                                    .foreach { case (topic, partitionThreadPairs) =>
                                          newGauge("OwnedPartitionsCount",
                                          new Gauge[Int] {
                                                def value() = partitionThreadPairs.size
                                                ownedPartitionsCountMetricTags(topic))
                                                topicRegistry = currentTopicRegistry
                                                updateFetcher(cluster)
                                                true
                                          } else {
                                          false


   上面代码的流程图如下:

   

 我们要了解Rebalance如何动作,看下updateFetcher怎么实现的。

private def updateFetcher(cluster: Cluster) {
// 遍历topicRegistry中保存的当前消费者的分区信息,修改Fetcher的partitions信息
var allPartitionInfos : List[PartitionTopicInfo] = Nil
for (partitionInfos <- topicRegistry.values)
for (partition <- partitionInfos.values)
allPartitionInfos ::= partition
info("Consumer " + consumerIdString + " selected partitions : " +
allPartitionInfos.sortWith((s,t) => s.partition < t.partition).map(_.toString).mkString(","))
fetcher match {
case Some(f) =>
// 调用fetcher的startConnections方法,初始化Fetcher并启动它
f.startConnections(allPartitionInfos, cluster)
case None =>


 注意下面这行代码:

f.startConnections(allPartitionInfos, cluster)  在这个方法里面其实是启动了一个LeaderFinderThread线程的,这个线程主要是通过ClientUtils的io,获取最新的topic元数据,将topic:partitionLeaderId和brokerId对应起来,封装成Map结构。

for ((brokerAndFetcherId, partitionAndOffsets) <- partitionsPerFetcher) {
var fetcherThread: AbstractFetcherThread = null
fetcherThreadMap.get(brokerAndFetcherId) match {
case Some(f) => fetcherThread = f
case None =>
fetcherThread = createFetcherThread(brokerAndFetcherId.fetcherId, brokerAndFetcherId.broker)
fetcherThreadMap.put(brokerAndFetcherId, fetcherThread)
fetcherThread.start
fetcherThreadMap(brokerAndFetcherId).addPartitions(partitionAndOffsets.map { case (topicAndPartition, brokerAndInitOffset) =>
topicAndPartition -> brokerAndInitOffset.initOffset


对每个broker创建一个FetcherRunnable线程,插入到fetcherThreadMap中并启动它。这个线程负责从服务器上不断获取数据,把数据插入内部阻塞队列的操作 。

下面看一下ConsumerIterator的实现,客户端用它不断的从分区信息的内部队列中取数据。它实现了IteratorTemplate的接口,它的内部保存一个Iterator的属性current,每次调用makeNext时会检查它,如果有则从中取否则从队列中取。下面给出代码

protected def makeNext(): MessageAndMetadata[T] = {
      var currentDataChunk: FetchedDataChunk = null
      // if we don't have an iterator, get one,从内部变量中取数据
      var localCurrent = current.get()
      if(localCurrent == null || !localCurrent.hasNext) {
            // 内部变量中取不到值,检查timeout的值
            if (consumerTimeoutMs < 0)
            currentDataChunk = channel.take // 是负数(-1),则表示永不过期,如果接下来无新数据可取,客户端线程会在channel.take阻塞住
            else {
                  // 设置了过期时间,在没有新数据可用时,pool会在相应的时间返回,返回值为空,则说明没有取到新数据,抛出timeout的异常
                  currentDataChunk = channel.poll(consumerTimeoutMs, TimeUnit.MILLISECONDS)
                  if (currentDataChunk == null) {
                        // reset state to make the iterator re-iterable
                        resetState()
                        throw new ConsumerTimeoutException
                        // kafka把shutdown的命令也做为一个datachunk放到队列中,用这种方法来保证消息的顺序性
                        if(currentDataChunk eq ZookeeperConsumerConnector.shutdownCommand) {
                              debug("Received the shutdown command")
                              channel.offer(currentDataChunk)
                              return allDone
                        } else {
                        currentTopicInfo = currentDataChunk.topicInfo
                        if (currentTopicInfo.getConsumeOffset != currentDataChunk.fetchOffset) {
                        error("consumed offset: %d doesn't match fetch offset: %d for %s;\n Consumer may lose data"
                        .format(currentTopicInfo.getConsumeOffset, currentDataChunk.fetchOffset, currentTopicInfo))
                        currentTopicInfo.resetConsumeOffset(currentDataChunk.fetchOffset)
                        // 把取出chunk中的消息转化为iterator
                        localCurrent = if (enableShallowIterator) currentDataChunk.messages.shallowIterator
                        else currentDataChunk.messages.iterator
                        // 使用这个新的iterator初始化current,下次可直接从current中取数据
                        current.set(localCurrent)
                        // 取出下一条数据,并用下一条数据的offset值设置consumedOffset
                        val item = localCurrent.next()
                        consumedOffset = item.offset
                        // 解码消息,封装消息和它的topic信息到MessageAndMetadata对象,返回
                        new MessageAndMetadata(decoder.toEvent(item.message), currentTopicInfo.topic)


 下面看一下它的next方法的逻辑:

override def next(): MessageAndMetadata[T] = {
val item = super.next()
if(consumedOffset < 0)
throw new IllegalStateException("Offset returned by the message set is invalid %d".format(consumedOffset))
// 使用makeNext方法设置的consumedOffset,去修改topicInfo的消费offset
currentTopicInfo.resetConsumeOffset(consumedOffset)
val topic = currentTopicInfo.topic
trace("Setting %s consumed offset to %d".format(topic, consumedOffset))
ConsumerTopicStat.getConsumerTopicStat(topic).recordMessagesPerTopic(1)
ConsumerTopicStat.getConsumerAllTopicStat().recordMessagesPerTopic(1)
// 返回makeNext得到的item
item




赞(0)    操作        顶端 
总帖数
1
每页帖数
101/1页1
返回列表
发新帖子
请输入验证码: 点击刷新验证码
您需要登录后才可以回帖 登录 | 注册
技术讨论