RocketMQ中如何實現(xiàn)并行模式

這篇文章主要介紹了RocketMQ中如何實現(xiàn)并行模式，具有一定借鑒價值，感興趣的朋友可以參考下，希望大家閱讀完這篇文章之后大有收獲，下面讓小編帶著大家一起了解一下。

DefaultMQPushConsumerImpl.pullMessage中的PullCallback在接收到拉取的message之后，會調(diào)用ConsumeMessageService.submitConsumeRequest方法將消息“推”給listener來執(zhí)行業(yè)務處理。RocketMQ支持并行和順序兩種消費模式，本文主要講解并行模式ConsumeMessageConcurrentlyService的實現(xiàn)。該類包含一下關(guān)鍵屬性：

public class ConsumeMessageConcurrentlyService implements ConsumeMessageService {
    // ...
    private final MessageListenerConcurrently messageListener; // 業(yè)務處理回掉
    private final BlockingQueue<Runnable> consumeRequestQueue;  // consumerExecutor的并發(fā)隊列
    private final ThreadPoolExecutor consumeExecutor;
    private final String consumerGroup;

    // ...

    public ConsumeMessageConcurrentlyService(DefaultMQPushConsumerImpl defaultMQPushConsumerImpl,
        MessageListenerConcurrently messageListener) {
        this.defaultMQPushConsumerImpl = defaultMQPushConsumerImpl;
        this.messageListener = messageListener;

        this.defaultMQPushConsumer = this.defaultMQPushConsumerImpl.getDefaultMQPushConsumer();
        this.consumerGroup = this.defaultMQPushConsumer.getConsumerGroup();
        this.consumeRequestQueue = new LinkedBlockingQueue<Runnable>();

        this.consumeExecutor = new ThreadPoolExecutor(
            this.defaultMQPushConsumer.getConsumeThreadMin(),
            this.defaultMQPushConsumer.getConsumeThreadMax(),
            1000 * 60,
            TimeUnit.MILLISECONDS,
            this.consumeRequestQueue,
            new ThreadFactoryImpl("ConsumeMessageThread_"));

        this.scheduledExecutorService = Executors.newSingleThreadScheduledExecutor(new ThreadFactoryImpl("ConsumeMessageScheduledThread_"));
        this.cleanExpireMsgExecutors = Executors.newSingleThreadScheduledExecutor(new ThreadFactoryImpl("CleanExpireMsgScheduledThread_"));
    }

    //...
}

ConsumeMessageConcurrentlyService.submitConsumeRequest會將拉取到的message列表按配置的分批策略做分割，提交執(zhí)行：

    public void submitConsumeRequest(
        final List<MessageExt> msgs,
        final ProcessQueue processQueue,
        final MessageQueue messageQueue,
        final boolean dispatchToConsume) {
        final int consumeBatchSize = this.defaultMQPushConsumer.getConsumeMessageBatchMaxSize(); // 配置的批次大小
        if (msgs.size() <= consumeBatchSize) {
            // 拉取的消息列表長度小于配置的批次大小，一次性提交處理
            ConsumeRequest consumeRequest = new ConsumeRequest(msgs, processQueue, messageQueue);
            try {
                this.consumeExecutor.submit(consumeRequest);
            } catch (RejectedExecutionException e) {
                this.submitConsumeRequestLater(consumeRequest);
            }
        } else {
            // 否則，將message分割后提交
            for (int total = 0; total < msgs.size(); ) {
                List<MessageExt> msgThis = new ArrayList<MessageExt>(consumeBatchSize);
                for (int i = 0; i < consumeBatchSize; i++, total++) {
                    if (total < msgs.size()) {
                        msgThis.add(msgs.get(total));
                    } else {
                        break;
                    }
                }

                ConsumeRequest consumeRequest = new ConsumeRequest(msgThis, processQueue, messageQueue);
                try {
                    this.consumeExecutor.submit(consumeRequest);
                } catch (RejectedExecutionException e) {
                    for (; total < msgs.size(); total++) {
                        msgThis.add(msgs.get(total));
                    }

                    this.submitConsumeRequestLater(consumeRequest);
                }
            }
        }
    }

實際處理邏輯在ConsumeRequest.run方法中

    // ConsumeRequest    
    public void run() {
            if (this.processQueue.isDropped()) {
                log.info("the message queue not be able to consume, because it's dropped. group={} {}", ConsumeMessageConcurrentlyService.this.consumerGroup, this.messageQueue);
                return;
            }

            MessageListenerConcurrently listener = ConsumeMessageConcurrentlyService.this.messageListener;
            ConsumeConcurrentlyContext context = new ConsumeConcurrentlyContext(messageQueue);
            ConsumeConcurrentlyStatus status = null;
            defaultMQPushConsumerImpl.resetRetryAndNamespace(msgs, defaultMQPushConsumer.getConsumerGroup());

            ConsumeMessageContext consumeMessageContext = null;
            // ....

            long beginTimestamp = System.currentTimeMillis();
            boolean hasException = false;
            ConsumeReturnType returnType = ConsumeReturnType.SUCCESS;
            try {
                if (msgs != null && !msgs.isEmpty()) {
                    for (MessageExt msg : msgs) {
                        MessageAccessor.setConsumeStartTimeStamp(msg, String.valueOf(System.currentTimeMillis()));
                    }
                }
                // 1. 調(diào)用listener
                status = listener.consumeMessage(Collections.unmodifiableList(msgs), context);
            } catch (Throwable e) {
                log.warn("consumeMessage exception: {} Group: {} Msgs: {} MQ: {}",
                    RemotingHelper.exceptionSimpleDesc(e),
                    ConsumeMessageConcurrentlyService.this.consumerGroup,
                    msgs,
                    messageQueue);
                hasException = true;
            }
            long consumeRT = System.currentTimeMillis() - beginTimestamp;
            // ...

            if (!processQueue.isDropped()) {
                // 2. 處理業(yè)務調(diào)用結(jié)果
                ConsumeMessageConcurrentlyService.this.processConsumeResult(status, context, this);
            } else {
                // ...
            }
        }
    }

processConsumeResult將處理消費結(jié)果：

    public void processConsumeResult(
        final ConsumeConcurrentlyStatus status,
        final ConsumeConcurrentlyContext context,
        final ConsumeRequest consumeRequest
    ) {
        int ackIndex = context.getAckIndex(); // 消費成功的index

        if (consumeRequest.getMsgs().isEmpty())
            return;

        // ...

        // 1. 消費失敗msg處理
        switch (this.defaultMQPushConsumer.getMessageModel()) {
            case BROADCASTING:
                for (int i = ackIndex + 1; i < consumeRequest.getMsgs().size(); i++) {
                    // 廣播模式，僅打印消費失敗的msg
                    MessageExt msg = consumeRequest.getMsgs().get(i);
                    log.warn("BROADCASTING, the message consume failed, drop it, {}", msg.toString());
                }
                break;
            case CLUSTERING:
                // 集群模式，首先嘗試將消費失敗的msg發(fā)回到broker，若失敗則在本地嘗試reconsume
                List<MessageExt> msgBackFailed = new ArrayList<MessageExt>(consumeRequest.getMsgs().size());
                for (int i = ackIndex + 1; i < consumeRequest.getMsgs().size(); i++) {
                    MessageExt msg = consumeRequest.getMsgs().get(i);
                    boolean result = this.sendMessageBack(msg, context);
                    if (!result) {
                        msg.setReconsumeTimes(msg.getReconsumeTimes() + 1);
                        msgBackFailed.add(msg);
                    }
                }

                if (!msgBackFailed.isEmpty()) {
                    consumeRequest.getMsgs().removeAll(msgBackFailed);

                    this.submitConsumeRequestLater(msgBackFailed, consumeRequest.getProcessQueue(), consumeRequest.getMessageQueue());
                }
                break;
            default:
                break;
        }

        // 2. 更新offset store
        long offset = consumeRequest.getProcessQueue().removeMessage(consumeRequest.getMsgs());
        if (offset >= 0 && !consumeRequest.getProcessQueue().isDropped()) {
            this.defaultMQPushConsumerImpl.getOffsetStore().updateOffset(consumeRequest.getMessageQueue(), offset, true);
        }
    }

感謝你能夠認真閱讀完這篇文章，希望小編分享的“RocketMQ中如何實現(xiàn)并行模式”這篇文章對大家有幫助，同時也希望大家多多支持億速云，關(guān)注億速云行業(yè)資訊頻道，更多相關(guān)知識等著你來學習!