这篇文章，我们聊一聊 RocketMQ 的消息轨迹设计思路。

查询消息轨迹可作为生产环境中排查问题强有力的数据支持，也是研发同学解决线上问题的重要武器之一。

1 基础概念

消息轨迹是指一条消息从生产者发送到 Broker , 再到消费者消费，整个过程中的各个相关节点的时间、状态等数据汇聚而成的完整链路信息。

聊聊 RocketMQ 消息轨迹-LMLPHP

当我们需要查询消息轨迹时，需要明白一点：消息轨迹数据是存储在 Broker 服务端，我们需要定义一个主题，在生产者，消费者端定义轨迹钩子。

2 开启轨迹

2.1 修改 Broker 配置文件

# 开启消息轨迹
traceTopicEnable=true

2.2 生产者配置

public DefaultMQProducer(final String producerGroup, boolean enableMsgTrace) 

public DefaultMQProducer(final String producerGroup, boolean enableMsgTrace, final String customizedTraceTopic)

在生产者的构造函数里，有两个核心参数：

enableMsgTrace：是否开启消息轨迹
customizedTraceTopic：记录消息轨迹的 Topic , 默认是： RMQ_SYS_TRACE_TOPIC 。

执行如下的生产者代码：

public class Producer {
    public static final String PRODUCER_GROUP = "mytestGroup";
    public static final String DEFAULT_NAMESRVADDR = "127.0.0.1:9876";
    public static final String TOPIC = "example";
    public static final String TAG = "TagA";

    public static void main(String[] args) throws MQClientException, InterruptedException {
        DefaultMQProducer producer = new DefaultMQProducer(PRODUCER_GROUP, true);
        producer.setNamesrvAddr(DEFAULT_NAMESRVADDR);
        producer.start();
        try {
            String key = UUID.randomUUID().toString();
            System.out.println(key);
            Message msg = new Message(
                    TOPIC,
                    TAG,
                    key,
                    ("Hello RocketMQ ").getBytes(RemotingHelper.DEFAULT_CHARSET));
            SendResult sendResult = producer.send(msg);
            System.out.printf("%s%n", sendResult);
        } catch (Exception e) {
            e.printStackTrace();
        }
        // 这里休眠十秒，是为了异步发送轨迹消息成功。
        Thread.sleep(10000);
        producer.shutdown();
    }
}

执行成功之后，我们从控制台查看轨迹信息。

聊聊 RocketMQ 消息轨迹-LMLPHP

从图中可以看到，消息轨迹中存储了消息的 存储时间 、 存储服务器IP 、发送耗时 。

2.3 消费者配置

和生产者类似，消费者的构造函数可以传递轨迹参数：

public DefaultMQPushConsumer(final String consumerGroup, boolean enableMsgTrace);

public DefaultMQPushConsumer(final String consumerGroup, boolean enableMsgTrace, final String customizedTraceTopic);

执行如下的消费者代码：

public class Consumer {
    public static final String CONSUMER_GROUP = "exampleGruop";
    public static final String DEFAULT_NAMESRVADDR = "127.0.0.1:9876";
    public static final String TOPIC = "example";

    public static void main(String[] args) throws InterruptedException, MQClientException {
        DefaultMQPushConsumer consumer = new DefaultMQPushConsumer(CONSUMER_GROUP , true);
        consumer.setNamesrvAddr(DEFAULT_NAMESRVADDR);
        consumer.setConsumeFromWhere(ConsumeFromWhere.CONSUME_FROM_FIRST_OFFSET);
        consumer.subscribe(TOPIC, "*");
        consumer.registerMessageListener((MessageListenerConcurrently) (msg, context) -> {
            System.out.printf("%s Receive New Messages: %s %n", Thread.currentThread().getName(), msg);
            return ConsumeConcurrentlyStatus.CONSUME_SUCCESS;
        });
        consumer.start();
        System.out.printf("Consumer Started.%n");
    }
}

聊聊 RocketMQ 消息轨迹-LMLPHP

3 实现原理

轨迹的实现原理主要是在生产者发送、消费者消费时添加相关的钩子。因此，我们只需要了解钩子的实现逻辑即可。

下面的代码是 DefaultMQProducer 的构造函数。

public DefaultMQProducer(final String namespace, final String producerGroup, RPCHook rpcHook,
    boolean enableMsgTrace, final String customizedTraceTopic) {
    this.namespace = namespace;
    this.producerGroup = producerGroup;
    defaultMQProducerImpl = new DefaultMQProducerImpl(this, rpcHook);
    // if client open the message trace feature
    if (enableMsgTrace) {
        try {
            //异步轨迹分发器
            AsyncTraceDispatcher dispatcher = new AsyncTraceDispatcher(producerGroup, TraceDispatcher.Type.PRODUCE, customizedTraceTopic, rpcHook);
            dispatcher.setHostProducer(this.defaultMQProducerImpl);
            traceDispatcher = dispatcher;
            // 发送消息时添加执行钩子
            this.defaultMQProducerImpl.registerSendMessageHook(
                new SendMessageTraceHookImpl(traceDispatcher));
            // 结束事务时添加执行钩子
            this.defaultMQProducerImpl.registerEndTransactionHook(
                new EndTransactionTraceHookImpl(traceDispatcher));
        } catch (Throwable e) {
            log.error("system mqtrace hook init failed ,maybe can't send msg trace data");
        }
    }
}

当是否开启轨迹开关打开时，创建异步轨迹分发器 AsyncTraceDispatcher ，然后给默认的生产者实现类在发送消息的钩子 SendMessageTraceHookImpl。

//发送消息时添加执行钩子
this.defaultMQProducerImpl.registerSendMessageHook(new SendMessageTraceHookImpl(traceDispatcher));

我们把生产者发送消息的流程简化如下代码：

//DefaultMQProducerImpl#sendKernelImpl
this.executeSendMessageHookBefore(context);
// 发生消息
this.mQClientFactory.getMQClientAPIImpl().sendMessage(....)
// 生产者发送消息后会执行
this.executeSendMessageHookAfter(context);

进入SendMessageTraceHookImpl 类，该类主要有两个方法 sendMessageBefore 和 sendMessageAfter 。

1、sendMessageBefore 方法

public void sendMessageBefore(SendMessageContext context) {
    //if it is message trace data,then it doesn't recorded
    if (context == null || context.getMessage().getTopic().startsWith(((AsyncTraceDispatcher)  localDispatcher).getTraceTopicName())) {
        return;
    }
    //build the context content of TuxeTraceContext
    TraceContext tuxeContext = new TraceContext();
    tuxeContext.setTraceBeans(new ArrayList<TraceBean>(1));
    context.setMqTraceContext(tuxeContext);
    tuxeContext.setTraceType(TraceType.Pub);
    tuxeContext.setGroupName(NamespaceUtil.withoutNamespace(context.getProducerGroup()));
    //build the data bean object of message trace
    TraceBean traceBean = new TraceBean();
    traceBean.setTopic(NamespaceUtil.withoutNamespace(context.getMessage().getTopic()));
    traceBean.setTags(context.getMessage().getTags());
    traceBean.setKeys(context.getMessage().getKeys());
    traceBean.setStoreHost(context.getBrokerAddr());
    traceBean.setBodyLength(context.getMessage().getBody().length);
    traceBean.setMsgType(context.getMsgType());
    tuxeContext.getTraceBeans().add(traceBean);
}

发送消息之前，先收集消息的 topic 、tag、key 、存储 Broker 的 IP 地址、消息体的长度等基础信息，并将消息轨迹数据存储在调用上下文中。

2、sendMessageAfter 方法

public void sendMessageAfter(SendMessageContext context) {
    // ...省略部分代码 
    TraceContext tuxeContext = (TraceContext) context.getMqTraceContext();
    TraceBean traceBean = tuxeContext.getTraceBeans().get(0);
    int costTime = (int) ((System.currentTimeMillis() - tuxeContext.getTimeStamp()) / tuxeContext.getTraceBeans().size());
    tuxeContext.setCostTime(costTime);
    if (context.getSendResult().getSendStatus().equals(SendStatus.SEND_OK)) {
        tuxeContext.setSuccess(true);
    } else {
        tuxeContext.setSuccess(false);
    }
    tuxeContext.setRegionId(context.getSendResult().getRegionId());
    traceBean.setMsgId(context.getSendResult().getMsgId());
    traceBean.setOffsetMsgId(context.getSendResult().getOffsetMsgId());
    traceBean.setStoreTime(tuxeContext.getTimeStamp() + costTime / 2);
    localDispatcher.append(tuxeContext);
}

跟踪对象里会保存 costTime (消息发送时间)、success （是否发送成功）、regionId (发送到 Broker 所在的分区) 、 msgId (消息 ID，全局唯一)、offsetMsgId (消息物理偏移量) ，storeTime (存储时间 ) 。

最后将跟踪上下文添加到本地轨迹分发器：

localDispatcher.append(tuxeContext);

下面我们分析下轨迹分发器的原理：

public AsyncTraceDispatcher(String group, Type type, String traceTopicName, RPCHook rpcHook) {
    // 省略代码 ....   
    this.traceContextQueue = new ArrayBlockingQueue<TraceContext>(1024);
    this.appenderQueue = new ArrayBlockingQueue<Runnable>(queueSize);
    if (!UtilAll.isBlank(traceTopicName)) {
        this.traceTopicName = traceTopicName;
    } else {
        this.traceTopicName = TopicValidator.RMQ_SYS_TRACE_TOPIC;
    }
    this.traceExecutor = new ThreadPoolExecutor(//
            10, 
            20, 
            1000 * 60, 
            TimeUnit.MILLISECONDS, 
            this.appenderQueue, 
            new ThreadFactoryImpl("MQTraceSendThread_"));
    traceProducer = getAndCreateTraceProducer(rpcHook);
}
public void start(String nameSrvAddr, AccessChannel accessChannel) throws MQClientException {
        if (isStarted.compareAndSet(false, true)) {
            traceProducer.setNamesrvAddr(nameSrvAddr);
            traceProducer.setInstanceName(TRACE_INSTANCE_NAME + "_" + nameSrvAddr);
            traceProducer.start();
        }
        this.accessChannel = accessChannel;
        this.worker = new Thread(new AsyncRunnable(), "MQ-AsyncTraceDispatcher-Thread-" + dispatcherId);
        this.worker.setDaemon(true);
        this.worker.start();
        this.registerShutDownHook();
}

上面的代码展示了分发器的构造函数和启动方法，构造函数创建了一个发送消息的线程池 traceExecutor ，启动 start 后会启动一个 worker线程。

class AsyncRunnable implements Runnable {
    private boolean stopped;
    @Override
    public void run() {
        while (!stopped) {
            synchronized (traceContextQueue) {
                long endTime = System.currentTimeMillis() + pollingTimeMil;
                while (System.currentTimeMillis() < endTime) {
                    try {
                        TraceContext traceContext = traceContextQueue.poll(
                                endTime - System.currentTimeMillis(), TimeUnit.MILLISECONDS
                        );
                        if (traceContext != null && !traceContext.getTraceBeans().isEmpty()) {
                            // get the topic which the trace message will send to
                            String traceTopicName = this.getTraceTopicName(traceContext.getRegionId());

                            // get the traceDataSegment which will save this trace message, create if null
                            TraceDataSegment traceDataSegment = taskQueueByTopic.get(traceTopicName);
                            if (traceDataSegment == null) {
                                traceDataSegment = new TraceDataSegment(traceTopicName, traceContext.getRegionId());
                                taskQueueByTopic.put(traceTopicName, traceDataSegment);
                            }

                            // encode traceContext and save it into traceDataSegment
                            // NOTE if data size in traceDataSegment more than maxMsgSize,
                            //  a AsyncDataSendTask will be created and submitted
                            TraceTransferBean traceTransferBean = TraceDataEncoder.encoderFromContextBean(traceContext);
                            traceDataSegment.addTraceTransferBean(traceTransferBean);
                        }
                    } catch (InterruptedException ignore) {
                        log.debug("traceContextQueue#poll exception");
                    }
                }
                // NOTE send the data in traceDataSegment which the first TraceTransferBean
                //  is longer than waitTimeThreshold
                sendDataByTimeThreshold();
                if (AsyncTraceDispatcher.this.stopped) {
                    this.stopped = true;
                }
            }
        }
  }

worker 启动后，会从轨迹上下文队列 traceContextQueue 中不断的取出轨迹上下文，并将上下文转换成轨迹数据片段 TraceDataSegment 。

为了提升系统的性能，并不是每一次从队列中获取到数据就直接发送到 MQ ，而是积累到一定程度的临界点才触发这个操作，我们可以简单的理解为批量操作。

这里面有两个维度 :

轨迹数据片段的数据大小大于某个数据大小阈值。笔者认为这段 RocketMQ 4.9.4 版本代码存疑，因为最新的 5.0 版本做了优化。

if (currentMsgSize >= traceProducer.getMaxMessageSize()) {
    List<TraceTransferBean> dataToSend = new ArrayList(traceTransferBeanList);
    AsyncDataSendTask asyncDataSendTask = new AsyncDataSendTask(traceTopicName, regionId, dataToSend);
    traceExecutor.submit(asyncDataSendTask);
    this.clear();
}

当前时间 - 轨迹数据片段的首次存储时间是否大于刷新时间，也就是每500毫秒刷新一次。

private void sendDataByTimeThreshold() {
    long now = System.currentTimeMillis();
    for (TraceDataSegment taskInfo : taskQueueByTopic.values()) {
        if (now - taskInfo.firstBeanAddTime >= waitTimeThresholdMil) {
            taskInfo.sendAllData();
        }
    }
}

轨迹数据存储的格式如下：

TraceBean bean = ctx.getTraceBeans().get(0);
//append the content of context and traceBean to transferBean's TransData
case Pub: {
  sb.append(ctx.getTraceType()).append(TraceConstants.CONTENT_SPLITOR)
    .append(ctx.getTimeStamp()).append(TraceConstants.CONTENT_SPLITOR)
    .append(ctx.getRegionId()).append(TraceConstants.CONTENT_SPLITOR)
    .append(ctx.getGroupName()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getTopic()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getMsgId()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getTags()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getKeys()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getStoreHost()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getBodyLength()).append(TraceConstants.CONTENT_SPLITOR)
    .append(ctx.getCostTime()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getMsgType().ordinal()).append(TraceConstants.CONTENT_SPLITOR)
    .append(bean.getOffsetMsgId()).append(TraceConstants.CONTENT_SPLITOR)
    .append(ctx.isSuccess()).append(TraceConstants.FIELD_SPLITOR);
}
break;

下图展示了事务轨迹消息数据，每个数据字段是按照 CONTENT_SPLITOR 分隔。

聊聊 RocketMQ 消息轨迹-LMLPHP

参考资料：

勇哥编程游记

聊聊 RocketMQ 消息轨迹

1 基础概念

2 开启轨迹

2.1 修改 Broker 配置文件

2.2 生产者配置

2.3 消费者配置

3 实现原理