关于高并发下kafka producer send异步发送耗时问题的分析

最近开发网关服务的过程当中，需要用到kafka转发消息与保存日志，在进行压测的过程中由于是多线程并发操作kafka producer 进行异步send，发现send耗时有时会达到几十毫秒的阻塞，很大程度上上影响了并发的性能，而在后续的测试中发现单线程发送反而比多线程发送效率高出几倍。所以就对kafka API send 的源码进行了一下跟踪和分析，在此总结记录一下。

首先看springboot下 kafka producer 的使用

在config中进行配置，向IOC容器中注入DefaultKafkaProducerFactory生产者工厂的实例

    @Bean
    public ProducerFactory<Object, Object> producerFactory() {
        return new DefaultKafkaProducerFactory<>(producerConfigs());
    }

创建producer

this.producer = producerFactory.createProducer();

大家都知道springboot下IOC容器管理的实例默认都是单例模式；而DefaultKafkaProducerFactory本身也是一个单例工厂

    @Override
    public Producer<K, V> createProducer() {
        if (this.transactionIdPrefix != null) {
            return createTransactionalProducer();
        }
        if (this.producer == null) {
            synchronized (this) {
                if (this.producer == null) {
                    this.producer = new CloseSafeProducer<K, V>(createKafkaProducer());
                }
            }
        }
        return this.producer;
    }

我们创建的producer也是个单例。

接下来就是具体的发送，用过kafka的小伙伴都知道producer.send是个异步操作，会返回一个Future<RecordMetadata> 类型的结果。那么为什么单线程和多线程send效率会较大的差距呢，我们进入KafkaProducer内部看下producer.send的具体源码实现来找下答案

private Future<RecordMetadata> doSend(ProducerRecord<K, V> record, Callback callback) {
        TopicPartition tp = null;
        try {
            //保证主题的元数据可用
            ClusterAndWaitTime clusterAndWaitTime = waitOnMetadata(record.topic(), record.partition(), maxBlockTimeMs);
            long remainingWaitMs = Math.max(0, maxBlockTimeMs - clusterAndWaitTime.waitedOnMetadataMs);
            Cluster cluster = clusterAndWaitTime.cluster;
            byte[] serializedKey;
            try {
                //序列化key
                serializedKey = keySerializer.serialize(record.topic(), record.headers(), record.key());
            } catch (ClassCastException cce) {
                throw new SerializationException("Can't convert key of class " + record.key().getClass().getName() +
                        " to class " + producerConfig.getClass(ProducerConfig.KEY_SERIALIZER_CLASS_CONFIG).getName() +
                        " specified in key.serializer", cce);
            }
            byte[] serializedValue;
            try {
                //序列化Value
                serializedValue = valueSerializer.serialize(record.topic(), record.headers(), record.value());
            } catch (ClassCastException cce) {
                throw new SerializationException("Can't convert value of class " + record.value().getClass().getName() +
                        " to class " + producerConfig.getClass(ProducerConfig.VALUE_SERIALIZER_CLASS_CONFIG).getName() +
                        " specified in value.serializer", cce);
            }
            //计算出具体的partition
            int partition = partition(record, serializedKey, serializedValue, cluster);
            tp = new TopicPartition(record.topic(), partition);

            setReadOnly(record.headers());
            Header[] headers = record.headers().toArray();

            int serializedSize = AbstractRecords.estimateSizeInBytesUpperBound(apiVersions.maxUsableProduceMagic(),
                    compressionType, serializedKey, serializedValue, headers);
            ensureValidRecordSize(serializedSize);
            long timestamp = record.timestamp() == null ? time.milliseconds() : record.timestamp();
            log.trace("Sending record {} with callback {} to topic {} partition {}", record, callback, record.topic(), partition);
            // producer callback will make sure to call both 'callback' and interceptor callback
            Callback interceptCallback = new InterceptorCallback<>(callback, this.interceptors, tp);

            if (transactionManager != null && transactionManager.isTransactional())
                transactionManager.maybeAddPartitionToTransaction(tp);
            //向队列容器中添加数据
            RecordAccumulator.RecordAppendResult result = accumulator.append(tp, timestamp, serializedKey,
                    serializedValue, headers, interceptCallback, remainingWaitMs);
            if (result.batchIsFull || result.newBatchCreated) {
                log.trace("Waking up the sender since topic {} partition {} is either full or getting a new batch", record.topic(), partition);
                this.sender.wakeup();
            }
            return result.future;
            // handling exceptions and record the errors;
            // for API exceptions return them in the future,
            // for other exceptions throw directly
        } catch (ApiException e) {
            log.debug("Exception occurred during message send:", e);
            if (callback != null)
                callback.onCompletion(null, e);
            this.errors.record();
            this.interceptors.onSendError(record, tp, e);
            return new FutureFailure(e);
        } catch (InterruptedException e) {
            this.errors.record();
            this.interceptors.onSendError(record, tp, e);
            throw new InterruptException(e);
        } catch (BufferExhaustedException e) {
            this.errors.record();
            this.metrics.sensor("buffer-exhausted-records").record();
            this.interceptors.onSendError(record, tp, e);
            throw e;
        } catch (KafkaException e) {
            this.errors.record();
            this.interceptors.onSendError(record, tp, e);
            throw e;
        } catch (Exception e) {
            // we notify interceptor about all exceptions, since onSend is called before anything else in this method
            this.interceptors.onSendError(record, tp, e);
            throw e;
        }
    }

这里除了前面做的一些序列化操作和判断，最关键的就是向队列容器中执行添加数据操作

RecordAccumulator.RecordAppendResult result = accumulator.append(tp, timestamp, serializedKey,
                    serializedValue, headers, interceptCallback, remainingWaitMs);

accumulator是RecordAccumulator这个类的一个实例，RecordAccumulator类是一个队列容器类；它的内部维护了一个ConcurrentMap，每一个TopicPartition都对应一个专属的消息队列。

private final ConcurrentMap<TopicPartition, Deque<ProducerBatch>> batches;

我们进入accumulator.append内部看下具体的实现

public RecordAppendResult append(TopicPartition tp,
                                     long timestamp,
                                     byte[] key,
                                     byte[] value,
                                     Header[] headers,
                                     Callback callback,
                                     long maxTimeToBlock) throws InterruptedException {
        // We keep track of the number of appending thread to make sure we do not miss batches in
        // abortIncompleteBatches().
        appendsInProgress.incrementAndGet();
        ByteBuffer buffer = null;
        if (headers == null) headers = Record.EMPTY_HEADERS;
        try {
            //根据TopicPartition拿到对应的批处理队列
            Deque<ProducerBatch> dq = getOrCreateDeque(tp);
            //同步队列，保证线程安全
            synchronized (dq) {
                if (closed)
                    throw new IllegalStateException("Cannot send after the producer is closed.");
                //把序列化后的数据放入队列，并返回结果
                RecordAppendResult appendResult = tryAppend(timestamp, key, value, headers, callback, dq);
                if (appendResult != null)
                    return appendResult;
            }

            // we don't have an in-progress record batch try to allocate a new batch
            byte maxUsableMagic = apiVersions.maxUsableProduceMagic();
            int size = Math.max(this.batchSize, AbstractRecords.estimateSizeInBytesUpperBound(maxUsableMagic, compression, key, value, headers));
            log.trace("Allocating a new {} byte message buffer for topic {} partition {}", size, tp.topic(), tp.partition());
            buffer = free.allocate(size, maxTimeToBlock);
            synchronized (dq) {
                // Need to check if producer is closed again after grabbing the dequeue lock.
                if (closed)
                    throw new IllegalStateException("Cannot send after the producer is closed.");

                RecordAppendResult appendResult = tryAppend(timestamp, key, value, headers, callback, dq);
                if (appendResult != null) {
                    // Somebody else found us a batch, return the one we waited for! Hopefully this doesn't happen often...
                    return appendResult;
                }

                MemoryRecordsBuilder recordsBuilder = recordsBuilder(buffer, maxUsableMagic);
                ProducerBatch batch = new ProducerBatch(tp, recordsBuilder, time.milliseconds());
                FutureRecordMetadata future = Utils.notNull(batch.tryAppend(timestamp, key, value, headers, callback, time.milliseconds()));

                dq.addLast(batch);
                incomplete.add(batch);

                // Don't deallocate this buffer in the finally block as it's being used in the record batch
                buffer = null;

                return new RecordAppendResult(future, dq.size() > 1 || batch.isFull(), true);
            }
        } finally {
            if (buffer != null)
                free.deallocate(buffer);
            appendsInProgress.decrementAndGet();
        }
    }

在getOrCreateDeque中我们根据TopicPartition从ConcurrentMap获取对应队列，没有的话就初始化一个。

    private Deque<ProducerBatch> getOrCreateDeque(TopicPartition tp) {
        Deque<ProducerBatch> d = this.batches.get(tp);
        if (d != null)
            return d;
        d = new ArrayDeque<>();
        Deque<ProducerBatch> previous = this.batches.putIfAbsent(tp, d);
        if (previous == null)
            return d;
        else
            return previous;
    }

更关键的是为了保证并发时的线程安全，执行 RecordAppendResult appendResult = tryAppend(timestamp, key, value, headers, callback, dq)时，Deque<ProducerBatch>必然需要同步处理。

synchronized (dq) {
                if (closed)
                    throw new IllegalStateException("Cannot send after the producer is closed.");
                RecordAppendResult appendResult = tryAppend(timestamp, key, value, headers, callback, dq);
                if (appendResult != null)
                    return appendResult;
            }

在这里我们可以看出，多线程高并发情况下，针对dq的操作会存在比较大的资源竞争，虽然是基于内存的操作，每个线程持有锁的时间极短，但相比单线程情况，高并发情况下线程开辟较多，锁竞争和cpu上下文切换都比较频繁，会造成一定的性能损耗，产生阻塞耗时。

分析到这里你就会发现，其实KafkaProducer这个异步发送是建立在生产者和消费者模式上的，send的真正操作并不是直接异步发送，而是把数据放在一个中间队列中。那么既然有生产者在往内存队列中放入数据，那么必然会有一个专有的线程负责把这些数据真正发送出去。我们通过监控jvm线程信息可以看到，KafkaProducer创建后确实会启动一个守护线程用于消息的发送。

OK，我们再回到 KafkaProducer中，会看到里面有这样两个对象，Sender就是kafka发送数据的后台线程

    private final Sender sender;
    private final Thread ioThread;

在KafkaProducer的构造函数中会启动Sender线程

            this.sender = new Sender(logContext,
                    client,
                    this.metadata,
                    this.accumulator,
                    maxInflightRequests == 1,
                    config.getInt(ProducerConfig.MAX_REQUEST_SIZE_CONFIG),
                    acks,
                    retries,
                    metricsRegistry.senderMetrics,
                    Time.SYSTEM,
                    this.requestTimeoutMs,
                    config.getLong(ProducerConfig.RETRY_BACKOFF_MS_CONFIG),
                    this.transactionManager,
                    apiVersions);
            String ioThreadName = NETWORK_THREAD_PREFIX + " | " + clientId;
            this.ioThread = new KafkaThread(ioThreadName, this.sender, true);
            this.ioThread.start();

进入Sender内部可以看到这个线程的作用就是一直轮询发送数据。

    public void run() {
        log.debug("Starting Kafka producer I/O thread.");

        // main loop, runs until close is called
        while (running) {
            try {
                run(time.milliseconds());
            } catch (Exception e) {
                log.error("Uncaught error in kafka producer I/O thread: ", e);
            }
        }

        log.debug("Beginning shutdown of Kafka producer I/O thread, sending remaining records.");

        // okay we stopped accepting requests but there may still be
        // requests in the accumulator or waiting for acknowledgment,
        // wait until these are completed.
        while (!forceClose && (this.accumulator.hasUndrained() || this.client.inFlightRequestCount() > 0)) {
            try {
                run(time.milliseconds());
            } catch (Exception e) {
                log.error("Uncaught error in kafka producer I/O thread: ", e);
            }
        }
        if (forceClose) {
            // We need to fail all the incomplete batches and wake up the threads waiting on
            // the futures.
            log.debug("Aborting incomplete batches due to forced shutdown");
            this.accumulator.abortIncompleteBatches();
        }
        try {
            this.client.close();
        } catch (Exception e) {
            log.error("Failed to close network client", e);
        }

        log.debug("Shutdown of Kafka producer I/O thread has completed.");
    }

    /**
     * Run a single iteration of sending
     *
     * @param now The current POSIX time in milliseconds
     */
    void run(long now) {
        if (transactionManager != null) {
            try {
                if (transactionManager.shouldResetProducerStateAfterResolvingSequences())
                    // Check if the previous run expired batches which requires a reset of the producer state.
                    transactionManager.resetProducerId();

                if (!transactionManager.isTransactional()) {
                    // this is an idempotent producer, so make sure we have a producer id
                    maybeWaitForProducerId();
                } else if (transactionManager.hasUnresolvedSequences() && !transactionManager.hasFatalError()) {
                    transactionManager.transitionToFatalError(new KafkaException("The client hasn't received acknowledgment for " +
                            "some previously sent messages and can no longer retry them. It isn't safe to continue."));
                } else if (transactionManager.hasInFlightTransactionalRequest() || maybeSendTransactionalRequest(now)) {
                    // as long as there are outstanding transactional requests, we simply wait for them to return
                    client.poll(retryBackoffMs, now);
                    return;
                }

                // do not continue sending if the transaction manager is in a failed state or if there
                // is no producer id (for the idempotent case).
                if (transactionManager.hasFatalError() || !transactionManager.hasProducerId()) {
                    RuntimeException lastError = transactionManager.lastError();
                    if (lastError != null)
                        maybeAbortBatches(lastError);
                    client.poll(retryBackoffMs, now);
                    return;
                } else if (transactionManager.hasAbortableError()) {
                    accumulator.abortUndrainedBatches(transactionManager.lastError());
                }
            } catch (AuthenticationException e) {
                // This is already logged as error, but propagated here to perform any clean ups.
                log.trace("Authentication exception while processing transactional request: {}", e);
                transactionManager.authenticationFailed(e);
            }
        }

        long pollTimeout = sendProducerData(now);
        client.poll(pollTimeout, now);
    }

通过上面的分析我们可以看出producer.send操作本身其实是个基于内存的存储操作，耗时几乎可以忽略不计，但由于高并发情况下，线程同步会有一定的性能损耗，当然这个损耗在一般的应用场景下几乎是可以忽略不计的，但如果是数据量比较大，高并发的场景下会比较明显。

针对上面的问题分析，这里说下我个人的一些总结：

1、首先避免多线程操作producer发送数据，你可以采用生产者消费者模式把producer.send从你的多线程操作中解耦出来，维护一个你要发送的消息队列，单独开辟一个线程操作；

2、可能有的小伙伴会问，那么多创建几个producer的实例或者维护一个producer池可以吗，我原本也是这个想法，只是在测试中发现效果也不是很理想，我估计是由于创建producer实例过多，导致线程数量也跟着增加，本身的业务线程再加上kafka的线程，线程上下文切换比较频繁，CPU资源压力比较大，效率也不如单线程操作；

3、这个问题其实真是针对API操作来讲的，send操作并不是真正的数据发送，真正的数据发送由守护线程进行；按照kafka本身的设计思想，如果操作本身就成为了你性能的瓶颈，你应该考虑的是集群部署，负载均衡；

4、无锁才是真正的高性能；

关注微信公众号，查看更多技术文章。