一、场景说明

当我们初始化一个Kafka生产者后(初始化流程可以查看《Kafka源码解析之生产者初始化流程》)，通过该生产者将封装好的消息发送出去，示例代码仍然参考example模块下的Producer.java：

public class Producer extends Thread {  public Producer(String topic, Boolean isAsync) {        Properties props = new Properties();        props.put(ProducerConfig.BOOTSTRAP_SERVERS_CONFIG, KafkaProperties.KAFKA_SERVER_URL + ":" + KafkaProperties.KAFKA_SERVER_PORT);        props.put(ProducerConfig.CLIENT_ID_CONFIG, "DemoProducer");        props.put(ProducerConfig.KEY_SERIALIZER_CLASS_CONFIG, IntegerSerializer.class.getName());        props.put(ProducerConfig.VALUE_SERIALIZER_CLASS_CONFIG, StringSerializer.class.getName());        //初始化KafkaProducer        producer = new KafkaProducer<>(props);        this.topic = topic;        this.isAsync = isAsync;    }  public void run() {        int messageNo = 1;        while (true) {            String messageStr = "Message_" + messageNo;            long startTime = System.currentTimeMillis();            if (isAsync) { // Send asynchronously                //异步发送消息                producer.send(new ProducerRecord<>(topic,                    messageNo,                    messageStr), new DemoCallBack(startTime, messageNo, messageStr));            }             ...            ++messageNo;        }    }}    

发送消息的过程中就必然需要集群的元数据，比如指定的Topic有多少分区，每个分区的Leader副本在哪个节点上等等。那么客户端是如何获取集群元数据的呢？下面通过图示+源码的方式详细分析一下这个流程。

二、获取元数据流程图

这里重点分析主线程和Sender线程的切换，集群元数据的获取是通过Sender线程完成的。

三、过程源码解析

1、KafkaProducer通过send方法最终调用了doSend方法，生产者生产的消息就是通过这个方法发送给客户端的，截取部分代码如下：

private FuturedoSend(ProducerRecord record, Callback callback) {    TopicPartition tp = null;    try {        throwIfProducerClosed();        // first make sure the metadata for the topic is available        ClusterAndWaitTime clusterAndWaitTime;        try {            //TODO 步骤一：同步等待获取元数据(在这个方法中主线程会阻塞，并唤醒sender线程)            clusterAndWaitTime = waitOnMetadata(record.topic(), record.partition(), maxBlockTimeMs);        } catch (KafkaException e) {            if (metadata.isClosed())                throw new KafkaException("Producer closed while send in progress", e);            throw e;        }    }    ...}            

doSend方法中通过调用waitOnMetadata来同步等待获取元数据，会阻塞主线程，直到获取到元数据或者超时，该方法代码如下：

private ClusterAndWaitTime waitOnMetadata(String topic, Integer partition, long maxWaitMs) throws InterruptedException {    //拉取元数据，第一次进来是拉取不到的，metadata只有配置的bootstrap.servers信息    Cluster cluster = metadata.fetch();    if (cluster.invalidTopics().contains(topic))        throw new InvalidTopicException(topic);    //将指定的topic添加到元数据    metadata.add(topic);    //获取指定topic的分区数，第一次进来是没有topic的信息的，所以也获取不到分区数，partitionsCount为null    Integer partitionsCount = cluster.partitionCountForTopic(topic);    if (partitionsCount != null && (partition == null || partition < partitionsCount))        return new ClusterAndWaitTime(cluster, 0);    long begin = time.milliseconds();    long remainingWaitMs = maxWaitMs;    long elapsed;    //循环，直到获取集群元数据，或者超时    do {        if (partition != null) {            log.trace("Requesting metadata update for partition {} of topic {}.", partition, topic);        } else {            log.trace("Requesting metadata update for topic {}.", topic);        }        metadata.add(topic);        //获取当前元数据的版本号        int version = metadata.requestUpdate();        //唤醒sender线程        sender.wakeup();        try {            //阻塞等待元数据            //退出等待有两种方式：1)达到等待的时间；2)被其他线程唤醒了            metadata.awaitUpdate(version, remainingWaitMs);        } catch (TimeoutException ex) {            // Rethrow with original maxWaitMs to prevent logging exception with remainingWaitMs            throw new TimeoutException(                    String.format("Topic %s not present in metadata after %d ms.",                            topic, maxWaitMs));        }        //如果线程被唤醒，说明已经可以获取到元数据了        cluster = metadata.fetch();        //花费的时间        elapsed = time.milliseconds() - begin;        //如果超时，抛异常        if (elapsed >= maxWaitMs) {            throw new TimeoutException(partitionsCount == null ?                    String.format("Topic %s not present in metadata after %d ms.",                            topic, maxWaitMs) :                    String.format("Partition %d of topic %s with partition count %d is not present in metadata after %d ms.",                            partition, topic, partitionsCount, maxWaitMs));        }        metadata.maybeThrowExceptionForTopic(topic);        //计算剩余时间        remainingWaitMs = maxWaitMs - elapsed;        //获取分区数        partitionsCount = cluster.partitionCountForTopic(topic);    } while (partitionsCount == null || (partition != null && partition >= partitionsCount));    return new ClusterAndWaitTime(cluster, elapsed);}

该方法中通过一个do...while循环不断尝试获取元数据，我们看几段重要的代码：

//获取当前元数据的版本号int version = metadata.requestUpdate();

a .获取当前元数据的版本号，是一个递增的值，客户端每更新一次元数据，就同时更新一次版本号；

//唤醒sender线程sender.wakeup();

b.唤醒Sender线程，集群的元数据就是通过Sender线程获取到的

//阻塞等待元数据metadata.awaitUpdate(version, remainingWaitMs);

c.阻塞主线程，等待更新元数据，结束等待的条件有两个：

• 被其它线程唤醒

• 达到等待时间

awaitUpdate方法的代码如下：

public synchronized void awaitUpdate(final int lastVersion, final long timeoutMs) throws InterruptedException {    long currentTimeMs = time.milliseconds();    //最后期限=当前时间+等待时间    long deadlineMs = currentTimeMs + timeoutMs < 0 ? Long.MAX_VALUE : currentTimeMs + timeoutMs;    time.waitObject(this, () -> {        // Throw fatal exceptions, if there are any. Recoverable topic errors will be handled by the caller.        maybeThrowFatalException();        //直到最新的版本号>给定的版本号，方法返回        return updateVersion() > lastVersion || isClosed();    }, deadlineMs);    if (isClosed())        throw new KafkaException("Requested metadata update after close");}

该方法返回的条件是：updateVersion() > lastVersion，即更新后的version大于当前元数据的version。

2、既然唤醒了Sender线程来获取元数据，那就看一下它的run方法，内部调用了runOnce方法，代码如下：

void runOnce() {    //前面关于事务的代码先不看    ...    long currentTimeMs = time.milliseconds();    long pollTimeout = sendProducerData(currentTimeMs);    //TODO 真正执行网络操作的都是NetworkClient这个组件    // 包括发送请求，接收响应，处理响应    // 就是通过这个方法拉取的元数据    client.poll(pollTimeout, currentTimeMs);}

这里client的实现类是NetworkClient，是一个执行网络操作的组件，通过它的poll方法来获取元数据：

@Overridepublic Listpoll(long timeout, long now) {    //    ensureActive();    //List 如果abortedSends不为空，说明已经连接broker并获取了响应，直接处理    //第一次进来abortedSends为空，不走这个分支    if (!abortedSends.isEmpty()) {        // If there are aborted sends because of unsupported version exceptions or disconnects,        // handle them immediately without waiting for Selector#poll.        List responses = new ArrayList<>();        handleAbortedSends(responses);        completeResponses(responses);        return responses;    }    //TODO 步骤一：封装一个拉取元数据的请求    long metadataTimeout = metadataUpdater.maybeUpdate(now);    try {        //TODO 步骤二：发送请求，进行复杂的网络操作，这里用的就是java的NIO        this.selector.poll(Utils.min(timeout, metadataTimeout, defaultRequestTimeoutMs));    } catch (IOException e) {        log.error("Unexpected error during I/O", e);    }    // process completed actions    long updatedNow = this.time.milliseconds();    List responses = new ArrayList<>();    //将请求返回的响应放到responses集合    handleCompletedSends(responses, updatedNow);    //TODO 步骤三：处理响应，响应里面就会有我们需要的元数据    handleCompletedReceives(responses, updatedNow);    handleDisconnections(responses, updatedNow);    handleConnections();    handleInitiateApiVersionRequests(updatedNow);    handleTimedOutRequests(responses, updatedNow);    completeResponses(responses);    return responses;}

这个方法主要分三步：

• 封装一个拉取元数据的请求

• 向服务端发送请求，获取响应

• 处理响应，获取响应中的集群元数据信息

看一下具体的代码：

步骤一：

//TODO 步骤一：封装一个拉取元数据的请求long metadataTimeout = metadataUpdater.maybeUpdate(now);

这里metadataUpdater的实现类是DefaultMetadataUpdater，是NetworkClient类的一个内部类，其maybeUpdate方法如下：

private long maybeUpdate(long now, Node node) {    //获取连接的NodeId    String nodeConnectionId = node.idString();    //判断网络连接是否已经建立好，如果已经建立好，执行下面的代码(第二次进来时网络已经建立好了)    //第一次进来，网络连接显然是没有建立好的    if (canSendRequest(nodeConnectionId, now)) {        Metadata.MetadataRequestAndVersion requestAndVersion = metadata.newMetadataRequestAndVersion();        this.inProgressRequestVersion = requestAndVersion.requestVersion;        //构建一个拉取目标topic元数据的请求        MetadataRequest.Builder metadataRequest = requestAndVersion.requestBuilder;        log.debug("Sending metadata request {} to node {}", metadataRequest, node);        //发送拉取元数据的请求        sendInternalMetadataRequest(metadataRequest, nodeConnectionId, now);        return defaultRequestTimeoutMs;    }    if (isAnyNodeConnecting()) {        return reconnectBackoffMs;    }    //第一次进来由于没有建立好网络连接，走的是这个分支，初始化一个连接    if (connectionStates.canConnect(nodeConnectionId, now)) {        log.debug("Initialize connection to node {} for sending metadata request", node);        //初始化一个到给定节点的网络连接，其实只绑定了OP_CONNECT事件        initiateConnect(node, now);        return reconnectBackoffMs;    }    return Long.MAX_VALUE;}

当do...while循环第一次走到这个方法时，由于没有和给定节点建立连接，所以会先初始化一个网络连接；第二次进入到这个方法时，会执行下面的代码：

if (canSendRequest(nodeConnectionId, now)) {    Metadata.MetadataRequestAndVersion requestAndVersion = metadata.newMetadataRequestAndVersion();    this.inProgressRequestVersion = requestAndVersion.requestVersion;    //构建一个拉取目标topic元数据的请求    MetadataRequest.Builder metadataRequest = requestAndVersion.requestBuilder;    log.debug("Sending metadata request {} to node {}", metadataRequest, node);    //添加拉取元数据的请求    sendInternalMetadataRequest(metadataRequest, nodeConnectionId, now);    return defaultRequestTimeoutMs;}

先构建一个拉取元数据的MetadataRequest请求，然后通过sendInternalMetadataRequest方法将这个请求转为ClientRequest请求

 void sendInternalMetadataRequest(MetadataRequest.Builder builder, String nodeConnectionId, long now) {    //创建一个拉取元数据的请求    ClientRequest clientRequest = newClientRequest(nodeConnectionId, builder, now, true);    //保存要发送的请求    doSend(clientRequest, true, now);}

然后通过doSend方法将这个请求放到inFlightRequests里面，这里面保存的是已发送但是没有返回响应的请求，默认值最多保存5个请求，然后将这个请求放到发送队列等待Selector的poll方法处理。注意这里的Selector并不是JavaNIO中的那个Selector，而是kafka自己定义的。

private void doSend(ClientRequest clientRequest, boolean isInternalRequest, long now, AbstractRequest request) {    String destination = clientRequest.destination();    RequestHeader header = clientRequest.makeHeader(request.version());    if (log.isDebugEnabled()) {        int latestClientVersion = clientRequest.apiKey().latestVersion();        if (header.apiVersion() == latestClientVersion) {            log.trace("Sending {} {} with correlation id {} to node {}", clientRequest.apiKey(), request,                    clientRequest.correlationId(), destination);        } else {            log.debug("Using older server API v{} to send {} {} with correlation id {} to node {}",                    header.apiVersion(), clientRequest.apiKey(), request, clientRequest.correlationId(), destination);        }    }    Send send = request.toSend(destination, header);    InFlightRequest inFlightRequest = new InFlightRequest(            clientRequest,            header,            isInternalRequest,            request,            send,            now);    //把这个拉取元数据的请求放到inFlightRequests里面，    // 这个里面存储的是已发送请求，但是未返回响应的请求，默认最多5个    this.inFlightRequests.add(inFlightRequest);    //把请求放到发送队列等待poll方法处理    selector.send(send);}

步骤二：

//TODO 步骤二：发送请求，进行复杂的网络操作this.selector.poll(Utils.min(timeout, metadataTimeout, defaultRequestTimeoutMs));

poll方法的部分代码如下，这里用的就是Java的NIO，其中nioSelector才是Java中的Selecotr对象：

@Overridepublic void poll(long timeout) throws IOException {    ...    /* check ready keys */    long startSelect = time.nanoseconds();    //获取已经准备好io的selectionKey(channel)个数    int numReadyKeys = select(timeout);    long endSelect = time.nanoseconds();    this.sensors.selectTime.record(endSelect - startSelect, time.milliseconds());    if (numReadyKeys > 0 || !immediatelyConnectedKeys.isEmpty() || dataInBuffers) {        //获取所有准备好的selectionKey        Set readyKeys = this.nioSelector.selectedKeys();        ...        // Poll from channels where the underlying socket has more data        //遍历selectionKey进行处理        pollSelectionKeys(readyKeys, false, endSelect);        // Clear all selected keys so that they are included in the ready count for the next select        readyKeys.clear();        //处理IO操作        pollSelectionKeys(immediatelyConnectedKeys, true, endSelect);        immediatelyConnectedKeys.clear();    } else {        madeReadProgressLastPoll = true; //no work is also "progress"    }    long endIo = time.nanoseconds();    this.sensors.ioTime.record(endIo - endSelect, time.milliseconds());    ...}

其中处理IO操作的是pollSelectionKeys方法，截取部分关键代码如下：

@Overridepublic void poll(long timeout) throws IOException {    ...    /* check ready keys */    long startSelect = time.nanoseconds();    //获取已经准备好io的selectionKey(channel)个数    int numReadyKeys = select(timeout);    long endSelect = time.nanoseconds();    this.sensors.selectTime.record(endSelect - startSelect, time.milliseconds());    if (numReadyKeys > 0 || !immediatelyConnectedKeys.isEmpty() || dataInBuffers) {        //获取所有准备好的selectionKey        Set readyKeys = this.nioSelector.selectedKeys();        // Poll from channels that have buffered data (but nothing more from the underlying socket)        if (dataInBuffers) {            keysWithBufferedRead.removeAll(readyKeys); //so no channel gets polled twice            Set toPoll = keysWithBufferedRead;            keysWithBufferedRead = new HashSet<>(); //poll() calls will repopulate if needed            pollSelectionKeys(toPoll, false, endSelect);        }        //遍历selectionKey进行处理        pollSelectionKeys(readyKeys, false, endSelect);        // Clear all selected keys so that they are included in the ready count for the next select        readyKeys.clear();        pollSelectionKeys(immediatelyConnectedKeys, true, endSelect);        immediatelyConnectedKeys.clear();    } else {        madeReadProgressLastPoll = true; //no work is also "progress"    }    ...    //将stageReceives结构中的NetworkReceive对象放到completeReceive集合中    //stageReceives：Map>    //completeReceive:List    addToCompletedReceives();}

关键方法有pollSelectionKeys和addToCompletedReceives，其中pollSelectionKeys用来接收服务端返回的响应，并将响应封装成NetworkReceive保存到数据结构中；addToCompletedReceives用来将NetworkReceive对象放到特定的集合中，最后统一进行处理。

void pollSelectionKeys(Set selectionKeys,                       boolean isImmediatelyConnected,                       long currentTimeNanos) {    for (SelectionKey key : determineHandlingOrder(selectionKeys)) {        //获取对应的Kafkachannel        KafkaChannel channel = channel(key);        long channelStartTimeNanos = recordTimePerConnection ? time.nanoseconds() : 0;        boolean sendFailed = false;        ...        try {            /* complete any connections that have finished their handshake (either normally or immediately) */            //如果key对应的是连接事件,走这个分支            if (isImmediatelyConnected || key.isConnectable()) {                /**                 * TODO 核心代码                 *  最后完成网络连接的代码，如果之前初始化的时候，没有完成网络连接，这里会完成网络连接                 */                if (channel.finishConnect()) {                    //连接成功后，把brokerId放到连接成功的集合中                    this.connected.add(channel.id());                    this.sensors.connectionCreated.record();                    SocketChannel socketChannel = (SocketChannel) key.channel();                    log.debug("Created socket with SO_RCVBUF = {}, SO_SNDBUF = {}, SO_TIMEOUT = {} to node {}",                            socketChannel.socket().getReceiveBufferSize(),                            socketChannel.socket().getSendBufferSize(),                            socketChannel.socket().getSoTimeout(),                            channel.id());                } else {                    continue;                }            }            ...            //如果是接收返回的响应，走这个方法            attemptRead(key, channel);            ...                        //如果是发送数据，走这个分支            if (channel.ready() && key.isWritable() && !channel.maybeBeginClientReauthentication(                () -> channelStartTimeNanos != 0 ? channelStartTimeNanos : currentTimeNanos)) {                Send send;                try {                    //TODO 往服务端发送消息                    //方法里面消息被发送出去，并移除OP_WRITE事件                    send = channel.write();                } catch (Exception e) {                    sendFailed = true;                    throw e;                }                if (send != null) {                    //TODO 将响应添加到completedSends                    this.completedSends.add(send);                    this.sensors.recordBytesSent(channel.id(), send.size());                }            }        ...}

这里接收返回响应的方法是attemptRead(key, channel)，具体的逻辑是：如果KafkaChannel注册的是读事件，就从channel中不断地读取数据，并将NetworkReceive对象添加到stageReceive数据结构中，这是一个Map，key是KafkaChannel，value是一个NetworkReceive队列

private void attemptRead(SelectionKey key, KafkaChannel channel) throws IOException {    //如果是读请求    if (channel.ready() && (key.isReadable() || channel.hasBytesBuffered()) && !hasStagedReceive(channel)        && !explicitlyMutedChannels.contains(channel)) {        //接收服务端的响应(本质也是一个请求)        //NetworkReceive代表的就是服务端返回来的响应        NetworkReceive networkReceive;        while ((networkReceive = channel.read()) != null) {            madeReadProgressLastPoll = true;            //不断地读取数据，将这个响应放到stagedReceive队列中            addToStagedReceives(channel, networkReceive);        }        if (channel.isMute()) {            outOfMemory = true; //channel has muted itself due to memory pressure.        } else {            madeReadProgressLastPoll = true;        }    }}

而对于addToCompletedReceives方法，就是把上面的stageReceive数据结构转为List结构。

步骤三：

//TODO 步骤三：处理响应，响应里面就会有我们需要的元数据handleCompletedReceives(responses, updatedNow);

handleCompletedReceives方法代码如下，如果是关于元数据信息的响应，则执行handleCompletedMetadataResponse方法：

private void handleCompletedReceives(List responses, long now) {    //遍历completedReceives集合中的NetworkReceive    for (NetworkReceive receive : this.selector.completedReceives()) {        //获取brokerid        String source = receive.source();        //获取指定broker最后一个没有返回响应的请求        InFlightRequest req = inFlightRequests.completeNext(source);        //解析服务端返回的响应        Struct responseStruct = parseStructMaybeUpdateThrottleTimeMetrics(receive.payload(), req.header,            throttleTimeSensor, now);        ...        //TODO 如果是关于元数据信息的响应        if (req.isInternalRequest && body instanceof MetadataResponse)            metadataUpdater.handleCompletedMetadataResponse(req.header, now, (MetadataResponse) body);        else if (req.isInternalRequest && body instanceof ApiVersionsResponse)            handleApiVersionsResponse(responses, req, now, (ApiVersionsResponse) body);        else            responses.add(req.completed(body, now));    }}

截取部分handleCompletedMetadataResponse方法，其主要作用就是更新元数据：

public void handleCompletedMetadataResponse(RequestHeader requestHeader, long now, MetadataResponse response) {    ...    if (response.brokers().isEmpty()) {        log.trace("Ignoring empty metadata response with correlation id {}.", requestHeader.correlationId());        this.metadata.failedUpdate(now, null);    //如果响应中有broker信息，则更新元数据    } else {        //TODO 更新元数据，注意这里调用的是ProducerMetadata的update方法        // 里面通过notifyALL()方法来唤醒前面等待的线程        this.metadata.update(inProgressRequestVersion, response, now);    }    inProgressRequestVersion = null;}

注意Metadata的update方法：

public synchronized void update(int requestVersion, MetadataResponse response, long now) {    Objects.requireNonNull(response, "Metadata response cannot be null");    if (isClosed())        throw new IllegalStateException("Update requested after metadata close");    if (requestVersion == this.requestVersion)        this.needUpdate = false;    else        requestUpdate();    this.lastRefreshMs = now;    this.lastSuccessfulRefreshMs = now;    //更新元数据信息时，会将version值+1    this.updateVersion += 1;    String previousClusterId = cache.cluster().clusterResource().clusterId();    this.cache = handleMetadataResponse(response, topic -> retainTopic(topic.topic(), topic.isInternal(), now));    //获取响应中的cluster集群元数据    Cluster cluster = cache.cluster();    maybeSetMetadataError(cluster);    this.lastSeenLeaderEpochs.keySet().removeIf(tp -> !retainTopic(tp.topic(), false, now));    String newClusterId = cache.cluster().clusterResource().clusterId();    if (!Objects.equals(previousClusterId, newClusterId)) {        log.info("Cluster ID: {}", newClusterId);    }    //更新所有监听集群元数据的对象的元数据信息    clusterResourceListeners.onUpdate(cache.cluster().clusterResource());    log.debug("Updated cluster metadata updateVersion {} to {}", this.updateVersion, this.cache);}

关键步骤：this.updateVersion += 1;更新了元数据的version，此时更新后的version > 未更新前的verison，前面阻塞等待元数据的方法就会返回，从而继续执行KafkaProducer.waitOnMetadata方法后面的逻辑：

//如果线程被唤醒，说明已经可以获取到元数据了cluster = metadata.fetch();...//重新获取分区数partitionsCount = cluster.partitionCountForTopic(topic);

当重新获取的分区数不为null时，退出do...while循环。

至此，客户端就获取了集群的元数据信息，继续执行KafkaProducer.doSend方法后面的逻辑，继续向服务端发送数据。

总结：

• kafka获取集群元数据是通过Sender线程完成的

• 在获取集群元数据的过程中，主线程会阻塞，直到拿到元数据或者等待超时

• NetworkClient是Kafka进行网络操作的组件，拉取集群元数据的过程中进行了封装请求，发送请求和处理响应

• 元数据的版本号在拉取集群元数据的过程中起到了至关重要的作用

• Kafka网络通信采用了JavaNIO