• RecordAccumulator 1


    介绍

    前面讲过producer会将数据保存在RecordAccumulator中,并通过Sender发送数据。RecordAccumulator 就相当于一个队列保存着那些准备发送到server的数据。
    在producer中,有几个参数和RecordAccumulator 有关系:

    1. buffer.memory

      buffer.memory主要用来保存要发送的数据,里面的内存大部分是用来让RecordAccumulator保存数据的。

    2. compression.type
      压缩格式

    3. batch.size
      每个发送的batch大小

    4. linger.ms
      如果batch没有达到batch.size大小,但是已经等待了linger.ms长的时间,也会发送

    从上面的内容我们大体可以看出RecordAccumulator的作用:

    1. 数据读进来了,分配内存,并保存数据到一个一个的batch中,并返回是添加成功还是失败了。
    2. 找到那些满足发送条件的batches,然后由sender发送,发送的时候,如果有需要保证发送信息的前后顺序。
    3. flush数据,将所有的消息都发送出去。
    4. 强行停止,所有的batch都不发送了。
    5. 释放内存,2,3,4执行完了后,都需要将对应的batch占用的内存释放掉。

    RecordAccumulator 的数据都保存在指定大小的内存中,所以会有一个内存池来分配内存。这个变量就是private final BufferPool free;

    private final ConcurrentMap<TopicPartition, Deque<RecordBatch>> batches; 是用来保存消息队列的。里面每个TopicPartition,都会有一个Deque,保存每个RecordBatch。RecordBatch的本质就是一个ByteBuffer,它的大小就是前面介绍中提到的batch.size的大小。

    图1
    图1表示的RecordAccumulator的内存分配,大部分都是给了batches,还有一小部分给了正在飞的batch(发送到服务器,但是没有收到response)

    添加数据append

    在KafkaProducer的doSend函数中,会调用append函数将数据写入accumulator 中。

    
     private Future<RecordMetadata> doSend(ProducerRecord<K, V> record, Callback callback) {
     
        ....
        RecordAccumulator.RecordAppendResult result = accumulator.append(tp, timestamp, serializedKey, serializedValue, 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;
        ....
     }
    

    append 函数主要将消息append到TopicPartition的batch中。在append的时候,如果batch已经存在了,就直接添加到对应的batch中。如果batch不存在,那就从bufferPool中申请一个新的内存,然后写入消息。

        public RecordAppendResult append(TopicPartition tp,
                                         long timestamp,
                                         byte[] key,
                                         byte[] value,
                                         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();
            try {
                // check if we have an in-progress batch
                // 创建或者获取 tp 对应的 Deque
                Deque<RecordBatch> dq = getOrCreateDeque(tp);
                // 如果有Deque中有batch,就往这个batch中添加信息,并返回添加结果,如果没有,就返回null
                synchronized (dq) {
                    if (closed)
                        throw new IllegalStateException("Cannot send after the producer is closed.");
                    RecordAppendResult appendResult = tryAppend(timestamp, key, value, callback, dq);
                    if (appendResult != null)
                        return appendResult;
                }
    
                // we don't have an in-progress record batch try to allocate a new batch
                // 如果没有batch, 就分配一个内存出来
                int size = Math.max(this.batchSize, Records.LOG_OVERHEAD + Record.recordSize(key, value));
                log.trace("Allocating a new {} byte message buffer for topic {} partition {}", size, tp.topic(), tp.partition());
                ByteBuffer 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.");
                    //再次尝试添加,如果添加成功了,那就说明已经有另外一个线程建好了batch,这个时候就把刚分配好的内存还到bufferPool
                    RecordAppendResult appendResult = tryAppend(timestamp, key, value, callback, dq);
                    if (appendResult != null) {
                        // Somebody else found us a batch, return the one we waited for! Hopefully this doesn't happen often...
                        free.deallocate(buffer);
                        return appendResult;
                    }
                    // 开始创建 batch
                    MemoryRecords records = MemoryRecords.emptyRecords(buffer, compression, this.batchSize);
                    RecordBatch batch = new RecordBatch(tp, records, time.milliseconds());
                    //开始添加消息到batch中,如果这次添加失败了,那就说明有问题了,抛出一个异常
                    // 不过应该不会发生返回null的情况
                    FutureRecordMetadata future = Utils.notNull(batch.tryAppend(timestamp, key, value, callback, time.milliseconds()));
    
                    dq.addLast(batch);
                    // 将这个batch 标记为不完整
                    incomplete.add(batch);
                    return new RecordAppendResult(future, dq.size() > 1 || batch.records.isFull(), true);
                }
            } finally {
                appendsInProgress.decrementAndGet();
            }
        }
    

    在上面函数中有几点需要理解的地方:

    1. 分配内存这段代码并没有包含在synchronized 中,所以很可能同时会有多个线程申请内存。这个时候如果线程A申请到内存后,线程B已经创建好了,并且创建好了batch(这段代码用synchronized包含,所以同时只有一个线程进行操作)。那么线程A应该再次去尝试添加,如果添加成功了,就释放内存,将内存还给BufferPool。
    2. 为什么分配内存这段代码没有被包含在synchronized 中呢,因为BufferPool会一直等待,直到有足够的内存分配给申请的线程。如果加到synchronized中,那整个Deque都会被锁住,那其他线程就没办法访问这个Deque了。
    3. 如果数据写入到batch的频率和Sender发送的频率相等,那么每次写入batch的时候都需要申请内存,创建batch。如果数据写入到batch的频率大于Sender发送的频率,那么每次写入batch的时候就可以直接写入这个batch,直到batch满了或者等待的时间大于等于linger.ms。

    获取数据

    整个数据的获取都包含在Sender 的 run 函数中。

    1. 找到集群中那些已经准备好发送信息的节点。
    2. 获取要发送到这些节点的RecordBatchs.
    3. 找到那些已经过期的RecordBatchs。
     void run(long now) {
            获取到当前的集群信息
            Cluster cluster = metadata.fetch();
            // get the list of partitions with data ready to send
            获取当前准备发送的partitions,获取的条件如下:
            1.record set 满了
            2.record 等待的时间达到了 lingerms
            3.accumulator 的内存满了
            4.accumulator 要关闭了
            RecordAccumulator.ReadyCheckResult result = this.accumulator.ready(cluster, now);
            如果有些partition没有leader信息,更新metadata
            // if there are any partitions whose leaders are not known yet, force metadata update
            if (!result.unknownLeaderTopics.isEmpty()) {
                // The set of topics with unknown leader contains topics with leader election pending as well as
                // topics which may have expired. Add the topic again to metadata to ensure it is included
                // and request metadata update, since there are messages to send to the topic.
                for (String topic : result.unknownLeaderTopics)
                    this.metadata.add(topic);
                this.metadata.requestUpdate();
            }
            去掉那些不能发送信息的节点,能够发送的原因有:
            1.当前节点的信息是可以信赖的
            2.能够往这些节点发送信息
            // remove any nodes we aren't ready to send to
            Iterator<Node> iter = result.readyNodes.iterator();
            long notReadyTimeout = Long.MAX_VALUE;
            while (iter.hasNext()) {
                Node node = iter.next();
                if (!this.client.ready(node, now)) {
                    iter.remove();
                    notReadyTimeout = Math.min(notReadyTimeout, this.client.connectionDelay(node, now));
                }
            }
    
            获取要发送的records
            // create produce requests
            Map<Integer, List<RecordBatch>> batches = this.accumulator.drain(cluster,
                                                                             result.readyNodes,
                                                                             this.maxRequestSize,
                                                                             now);
            保证发送的顺序                                               
            if (guaranteeMessageOrder) {
                // Mute all the partitions drained
                for (List<RecordBatch> batchList : batches.values()) {
                    for (RecordBatch batch : batchList)
                        this.accumulator.mutePartition(batch.topicPartition);
                }
            }
            
            检查那些过期的records
            List<RecordBatch> expiredBatches = this.accumulator.abortExpiredBatches(this.requestTimeout, now);
            // update sensors
            for (RecordBatch expiredBatch : expiredBatches)
                this.sensors.recordErrors(expiredBatch.topicPartition.topic(), expiredBatch.recordCount);
    
            sensors.updateProduceRequestMetrics(batches);
            
            构建request并发送
            List<ClientRequest> requests = createProduceRequests(batches, now);
            // If we have any nodes that are ready to send + have sendable data, poll with 0 timeout so this can immediately
            // loop and try sending more data. Otherwise, the timeout is determined by nodes that have partitions with data
            // that isn't yet sendable (e.g. lingering, backing off). Note that this specifically does not include nodes
            // with sendable data that aren't ready to send since they would cause busy looping.
            long pollTimeout = Math.min(result.nextReadyCheckDelayMs, notReadyTimeout);
            if (result.readyNodes.size() > 0) {
                log.trace("Nodes with data ready to send: {}", result.readyNodes);
                log.trace("Created {} produce requests: {}", requests.size(), requests);
                pollTimeout = 0;
            }
            将这些requests加入channel中
            for (ClientRequest request : requests)
                client.send(request, now);
    
            // if some partitions are already ready to be sent, the select time would be 0;
            // otherwise if some partition already has some data accumulated but not ready yet,
            // the select time will be the time difference between now and its linger expiry time;
            // otherwise the select time will be the time difference between now and the metadata expiry time;
            真正的发送消息
            this.client.poll(pollTimeout, now);
        }
    

    在发送消息之前,produer需要直到那些节点是可以发送消息的,而这个就是通过 public ReadyCheckResult ready(Cluster cluster, long nowMs) 来获得的。

    mute

    在这里需要了解RecordAccumulator 的一个成员变量private final Set<TopicPartition> muted;。这个set里面保存了所有已经发送batch到server中,但是没有收到ack的TopicPartition,俗称inflight。等到接收到server的reponse后,会将对应的TopicPartition从set中去掉。这样子就可以保证每个TopicPartition的发送顺序。

    举例子,假如topic1要发送A,B,C三个batch到server。如果直接将A,B,C按照顺序发送过去,server的收到的顺序可能是C,B,A,这样子落到log中的顺序就变掉了。如果使用mute,发送A,mute里面就包含了topic1, 这个时候,Sender就不会从topic1所在的Deque中取batch了,直到produer收到了batch A 对应的response,然后从mute中去掉topic1。然后发送B...这样子就保证了服务器接收的顺序和producer发送的消息是一样的。

    ready

    在发送消息之前,需要确定一些信息:

    1. 哪些TopicPartition所对应的Node节点是可以发送信息的。
    2. 下次检查节点是否ready的时间。
    3. 哪些TopicPartition对应的leader找不到。

    这些都是在ready函数中实现的,返回的结果封装在ReadyCheckResult中。
    实际上,在发送过程中,可以向一个节点发送消息的时候需要满足下面的条件:

    1. 这个节点中至少有一个partition是可以正常发送的(没有处在backing off状态),并且这个 partition 没有 muted。
    2. batch 没有满,但是已经等了lingerMs 长的时间。
    3. accumulator 满了。
    4. accumulator 关闭了。
        public ReadyCheckResult ready(Cluster cluster, long nowMs) {
            Set<Node> readyNodes = new HashSet<>();
            long nextReadyCheckDelayMs = Long.MAX_VALUE;
            Set<String> unknownLeaderTopics = new HashSet<>();
    
            boolean exhausted = this.free.queued() > 0;
    
            for (Map.Entry<TopicPartition, Deque<RecordBatch>> entry : this.batches.entrySet()) {
                TopicPartition part = entry.getKey();
                Deque<RecordBatch> deque = entry.getValue();
                /*
                 * 遍历batches中每个tp
                 * 获取tp对应的leader
                 */
                Node leader = cluster.leaderFor(part);
    
                synchronized (deque) {
                    // 如果 leader 为 null ,并且deque 不为空,说明要发送消息却找不到cluster中接收消息的节点
                    // 就添加到 unknownLeaderTopics
                    if (leader == null && !deque.isEmpty()) {
                        // This is a partition for which leader is not known, but messages are available to send.
                        // Note that entries are currently not removed from batches when deque is empty.
                        unknownLeaderTopics.add(part.topic());
                    // 如果leader 没有ready, 并且 part 没有在飞
                    } else if (!readyNodes.contains(leader) && !muted.contains(part)) {
                        RecordBatch batch = deque.peekFirst();
                        if (batch != null) {
                            // 如果这个 batch 重试了, 就看看这个 batch 上次发送的时间 + retryBackoffMs 的时间长度 比当前时间要晚
                            boolean backingOff = batch.attempts > 0 && batch.lastAttemptMs + retryBackoffMs > nowMs;
                            // 等待的时间
                            long waitedTimeMs = nowMs - batch.lastAttemptMs;
                            // 等待时间
                            long timeToWaitMs = backingOff ? retryBackoffMs : lingerMs;
                            // 剩余的时间
                            long timeLeftMs = Math.max(timeToWaitMs - waitedTimeMs, 0);
                            // batch 满了
                            boolean full = deque.size() > 1 || batch.records.isFull();
                            // batch 过期了,它等待的时间已经超过了 timeToWaitMs
                            boolean expired = waitedTimeMs >= timeToWaitMs;
                            boolean sendable = full || expired || exhausted || closed || flushInProgress();
    
                            if (sendable && !backingOff) {
                                readyNodes.add(leader);
                            } else {
                                // Note that this results in a conservative estimate since an un-sendable partition may have
                                // a leader that will later be found to have sendable data. However, this is good enough
                                // since we'll just wake up and then sleep again for the remaining time.
                                nextReadyCheckDelayMs = Math.min(timeLeftMs, nextReadyCheckDelayMs);
                            }
                        }
                    }
                }
            }
    
            return new ReadyCheckResult(readyNodes, nextReadyCheckDelayMs, unknownLeaderTopics);
        }
    

    drain

    知道了要往那些Node 发送数据,就需要从accumulator中获取要发送的数据,要获取的数据的大小为max.request.size, 它是由几个不同的partition的batch组成的。这些batch可以被发送的的条件是:

    1. batch对应的tp没有数据在飞(已经发送出去了,但是没有收到response)。
    2. batch处在retry状态,并且已经等待了backoff长的时间。

    通过drain 函数,就可以得到这次request要发送batches了。因为drain是多线程并发的,所以在从Deque中取batch的时候,需要synchronized(deque)。

        public Map<Integer, List<RecordBatch>> drain(Cluster cluster,
                                                     Set<Node> nodes,
                                                     int maxSize,
                                                     long now) {
            if (nodes.isEmpty())
                return Collections.emptyMap();
    
            Map<Integer, List<RecordBatch>> batches = new HashMap<>();
            for (Node node : nodes) {
                int size = 0;
                //获取node 中对应的partition
                List<PartitionInfo> parts = cluster.partitionsForNode(node.id());
                List<RecordBatch> ready = new ArrayList<>();
                /* to make starvation less likely this loop doesn't start at 0 */
                // 避免每次都从一个相同的partition开始,别的partition没机会发送数据
                int start = drainIndex = drainIndex % parts.size();
                do {
                    PartitionInfo part = parts.get(drainIndex);
                    TopicPartition tp = new TopicPartition(part.topic(), part.partition());
                    // Only proceed if the partition has no in-flight batches.
                    if (!muted.contains(tp)) {
                        Deque<RecordBatch> deque = getDeque(new TopicPartition(part.topic(), part.partition()));
                        if (deque != null) {
                        // 注意synchronized
                            synchronized (deque) {
                                RecordBatch first = deque.peekFirst();
                                if (first != null) {
                                    // 查看当前batch 是不是在retry,并且没有达到需要等待的 backoff时间,如果不是的话,就加入
                                    boolean backoff = first.attempts > 0 && first.lastAttemptMs + retryBackoffMs > now;
                                    // Only drain the batch if it is not during backoff period.
                                    if (!backoff) {
                                        // 如果batch的大小大于maxSize 并且 ready 里面有东西,就停止
                                        // 这时候有一种情况就是batch的大小大于maxSize但是ready 里面没有内容就把这个batch加入ready中
                                        if (size + first.records.sizeInBytes() > maxSize && !ready.isEmpty()) {
                                            // there is a rare case that a single batch size is larger than the request size due
                                            // to compression; in this case we will still eventually send this batch in a single
                                            // request
                                            break;
                                        } else {
                                            //添加到ready, 注意要close batch.records
                                            RecordBatch batch = deque.pollFirst();
                                            batch.records.close();
                                            size += batch.records.sizeInBytes();
                                            ready.add(batch);
                                            batch.drainedMs = now;
                                        }
                                    }
                                }
                            }
                        }
                    }
                    this.drainIndex = (this.drainIndex + 1) % parts.size();
                } while (start != drainIndex);
                batches.put(node.id(), ready);
            }
            return batches;
        }
    

    flush

    在发送消息的时候,如果想要将所有的数据都发送出去,就需要调用kafkaproducer的flush函数。调用flush后,会将所有的batch都发送出去(不严谨)。

        public void flush() {
            log.trace("Flushing accumulated records in producer.");
            this.accumulator.beginFlush();
            this.sender.wakeup();
            try {
                this.accumulator.awaitFlushCompletion();
            } catch (InterruptedException e) {
                throw new InterruptException("Flush interrupted.", e);
            }
        }
    
    

    上面是flush函数的实现,首先开始flush,然后通知sender 发送消息,最后等待所有消息发送完毕。

    这里面涉及到 RecordAccumulator 的一个成员变量flushesInProgress,它是一个AtomicInteger。当它大于0的时候,所有的batch都会被发送出去。
    那么beginFlush就是将flushesInProgress++。

        public void beginFlush() {
            this.flushesInProgress.getAndIncrement();
        }
    

    在ready函数中,判断是否可以发送batch的条件如下:

        public ReadyCheckResult ready(Cluster cluster, long nowMs) {
            ....
            for (Map.Entry<TopicPartition, Deque<RecordBatch>> entry : this.batches.entrySet()) {
                            //判断条件
                            boolean sendable = full || expired || exhausted || closed || flushInProgress();
    
                            if (sendable && !backingOff) {
                                readyNodes.add(leader);
                            } else {
                            ....
                            }
            ....
        }
        
        boolean flushInProgress() {
            return flushesInProgress.get() > 0;
        }
        
    

    在append 数据的时候,如果batch是新建的,就会将这个batch加入到incomplete 的Set中,直到收到了服务器的response,才会将这个batch从 incomplete 去掉。而awaitFlushCompletion就是等待incomplete 为空后,就结束了。这样子accumulator中所有的数据都会被发送出去。

        public void awaitFlushCompletion() throws InterruptedException {
            try {
                for (RecordBatch batch : this.incomplete.all())
                    batch.produceFuture.await();
            } finally {
                this.flushesInProgress.decrementAndGet();
            }
        }
    

    abort

    当然还有Sender要强制关闭的时候,这个时候就需要将accumulator中所有的batch占用的内存释放掉,然后close掉就Ok了。

  • 相关阅读:
    MySql
    Zookeeper
    Kafka
    Mybatis
    Spring/Spring MVC
    Spring Boot/Spring Cloud
    网络
    设计模式
    Strassen algorithm(O(n^lg7))
    dynamic programming:find max subarray
  • 原文地址:https://www.cnblogs.com/SpeakSoftlyLove/p/6769550.html
Copyright © 2020-2023  润新知