该文所涉及的 RocketMQ 源码版本为 4.9.3。

RockerMQ Nameserver 如何与 Broker 进行通信的?

nameserver 每隔 10s 扫描一次 Broker,移除处于未激活状态的 Broker

核心代码:

this.scheduledExecutorService.scheduleAtFixedRate(NamesrvController.this.routeInfoManager::scanNotActiveBroker, 5, 10, TimeUnit.*SECONDS*);

  1. public int scanNotActiveBroker() {
  2. int removeCount = 0;
  3. Iterator<Entry<String, BrokerLiveInfo>> it = this.brokerLiveTable.entrySet().iterator();
  4. while (it.hasNext()) {
  5. Entry<String, BrokerLiveInfo> next = it.next();
  6. long last = next.getValue().getLastUpdateTimestamp();
  7. if ((last +BROKER_CHANNEL_EXPIRED_TIME) < System.currentTimeMillis()) {
  8. RemotingUtil.closeChannel(next.getValue().getChannel());
  9. it.remove();
  10. log.warn("The broker channel expired, {} {}ms", next.getKey(),BROKER_CHANNEL_EXPIRED_TIME);
  11. this.onChannelDestroy(next.getKey(), next.getValue().getChannel());
  12. removeCount++;
  13. }
  14. }
  15. return removeCount;
  16. }

broker 每隔 30 秒会向集群中所有的 NameServer 发送心跳包

核心代码:

  1. this.scheduledExecutorService.scheduleAtFixedRate(new Runnable() {
  2. @Override
  3. public void run() {
  4. try {
  5. BrokerController.this.registerBrokerAll(true, false, brokerConfig.isForceRegister());
  6. } catch (Throwable e) {
  7. log.error("registerBrokerAll Exception", e);
  8. }
  9. }
  10. }, 1000 * 10, Math.max(10000, Math.min(brokerConfig.getRegisterNameServerPeriod(), 60000)), TimeUnit.MILLISECONDS);
  1. public synchronized void registerBrokerAll(final boolean checkOrderConfig, boolean oneway, boolean forceRegister) {
  2. TopicConfigSerializeWrapper topicConfigWrapper = this.getTopicConfigManager().buildTopicConfigSerializeWrapper();
  3. if (!PermName.isWriteable(this.getBrokerConfig().getBrokerPermission())
  4. || !PermName.isReadable(this.getBrokerConfig().getBrokerPermission())) {
  5. ConcurrentHashMap<String, TopicConfig> topicConfigTable = new ConcurrentHashMap<String, TopicConfig>();
  6. for (TopicConfig topicConfig : topicConfigWrapper.getTopicConfigTable().values()) {
  7. TopicConfig tmp =
  8. new TopicConfig(topicConfig.getTopicName(), topicConfig.getReadQueueNums(), topicConfig.getWriteQueueNums(),
  9. this.brokerConfig.getBrokerPermission());
  10. topicConfigTable.put(topicConfig.getTopicName(), tmp);
  11. }
  12. topicConfigWrapper.setTopicConfigTable(topicConfigTable);
  13. }
  14. if (forceRegister || needRegister(this.brokerConfig.getBrokerClusterName(),
  15. this.getBrokerAddr(),
  16. this.brokerConfig.getBrokerName(),
  17. this.brokerConfig.getBrokerId(),
  18. this.brokerConfig.getRegisterBrokerTimeoutMills())) {
  19. doRegisterBrokerAll(checkOrderConfig, oneway, topicConfigWrapper);
  20. }
  21. }

org.apache.rocketmq.namesrv.processor.DefaultRequestProcessor 是网络处理器解析请求类型,如果请求类型为*RequestCode.REGISTER_BROKER,则请求最终转发到 org.apache.rocketmq.namesrv.routeinfo.RouteInfoManager#registerBroker*

代码太多,文字来描述一下:

第一步:路由注册需要加写锁,防止并发修改 RouteInfoManager 中的路由表。首先判断 Broker 所属集群是否存在,如果不存在,则创建集群,然后将 broker 名加入集群。

第二步:维护 BrokerData 信息,首先从 brokerAddrTable 中根据 broker 名尝试获取 Broker 信息,如果不存在,则新建 BrokerData 放入 brokerAddrTable,registerFirst 设置为 true;如果存在,直接替换原先的 Broker 信息,registerFirst 设置为 false,表示非第一次注册

第三步:如果 Broker 为主节点,并且 Broker 的 topic 配置信息发生变化或者是初次注册,则需要创建或者更新 topic 的路由元数据,并填充 topicQueueTable

根据 topicConfig 创建 QueueData 数据结构然后更新 topicQueueTable

  1. private void createAndUpdateQueueData(final String brokerName, final TopicConfig topicConfig) {
  2. QueueData queueData = new QueueData();
  3. queueData.setBrokerName(brokerName);
  4. queueData.setWriteQueueNums(topicConfig.getWriteQueueNums());
  5. queueData.setReadQueueNums(topicConfig.getReadQueueNums());
  6. queueData.setPerm(topicConfig.getPerm());
  7. queueData.setTopicSysFlag(topicConfig.getTopicSysFlag());
  8. Map<String, QueueData> queueDataMap = this.topicQueueTable.get(topicConfig.getTopicName());
  9. if (null == queueDataMap) {
  10. queueDataMap = new HashMap<>();
  11. queueDataMap.put(queueData.getBrokerName(), queueData);
  12. this.topicQueueTable.put(topicConfig.getTopicName(), queueDataMap);
  13. log.info("new topic registered, {} {}", topicConfig.getTopicName(), queueData);
  14. } else {
  15. QueueData old = queueDataMap.put(queueData.getBrokerName(), queueData);
  16. if (old != null && !old.equals(queueData)) {
  17. log.info("topic changed, {} OLD: {} NEW: {}", topicConfig.getTopicName(), old,
  18. queueData);
  19. }
  20. }
  21. }

第四步:更新 BrokerLiveInfo,存储状态正常的 Broker 信息表,BrokerLiveInfo 是执行路由删除操作的重要依据。

  1. BrokerLiveInfo prevBrokerLiveInfo = this.brokerLiveTable.put(brokerAddr,
  2. new BrokerLiveInfo(
  3. System.currentTimeMillis(),
  4. topicConfigWrapper.getDataVersion(),
  5. channel,
  6. haServerAddr));

第五步:注册 Broker 的过滤器 Server 地址列表,一个 Broker 上会关联多个 FilterServer 消息过滤服务器。如果此 Broker 为从节点,则需要查找该 Broker 的主节点信息,并更新对应的 masterAddr 属性

  1. if (MixAll.MASTER_ID!= brokerId) {
  2. String masterAddr = brokerData.getBrokerAddrs().get(MixAll.MASTER_ID);
  3. if (masterAddr != null) {
  4. BrokerLiveInfo brokerLiveInfo = this.brokerLiveTable.get(masterAddr);
  5. if (brokerLiveInfo != null) {
  6. result.setHaServerAddr(brokerLiveInfo.getHaServerAddr());
  7. result.setMasterAddr(masterAddr);
  8. }
  9. }
  10. }

总结:

NameServer 与 Broker 保持长连接,Broker 的状态信息存储在 BrokerLiveTable 中,NameServer 每收到一个心跳包,将更新 brokerLiveTable 中关于 broker 的状态信息以及路由表(topicQueueTable、brokerAddrTable、brokerLiveTable、filterServerTable)。更新上述路由表使用了锁粒度较少的读写锁,允许多个消息发送者并发读操作,保证消息发送时的高并发,同一时刻 NameServer 只处理一个 Broker 心跳包,多个心跳包请求串行执行。

NameServer 如何剔除失效的 Broker?

1、NameServer 每隔十秒注册一次 brokerLiveTable 状态表,如果 BrokerLive 的 lastUpdateTimestamp

时间戳距当前时间超过 120 秒,则认为 Broker 失效,移除该 Broker,关闭与 broker 的连接,同时更新 topicQueueTable、brokerAddrTable、brokerLiveTable、filterServerTable。

2、如果 broker 在正常关闭的情况下,会发送 unRegisterBroker 指令。

不管是哪一种方式触发的路由删除,处理逻辑是一样的

第一步:申请写锁,移除 brokerLiveTable、filterServerTable 中 Broker 相关的信息

  1. this.lock.writeLock().lockInterruptibly();
  2. BrokerLiveInfo brokerLiveInfo = this.brokerLiveTable.remove(brokerAddr);
  3. log.info("unregisterBroker, remove from brokerLiveTable {}, {}",brokerLiveInfo != null ? "OK" : "Failed",brokerAddr);
  4. this.filterServerTable.remove(brokerAddr);

第二步:维护 brokerAddrTable,找到具体的 broker,将其从 brokerData 中移除,如果移除之后不再包含其他 broker,则在 brokerAddrtable 移除该 brokerName 对应的数据

  1. BrokerData brokerData = this.brokerAddrTable.get(brokerName);
  2. if (null != brokerData) {
  3. String addr = brokerData.getBrokerAddrs().remove(brokerId);
  4. log.info("unregisterBroker, remove addr from brokerAddrTable {}, {}",addr != null ? "OK" : "Failed",brokerAddr);
  5. if (brokerData.getBrokerAddrs().isEmpty()) {
  6. this.brokerAddrTable.remove(brokerName);
  7. log.info("unregisterBroker, remove name from brokerAddrTable OK, {}",brokerName);
  8. removeBrokerName = true;
  9. }
  10. }

第三步:根据 brokerName 从 clusterAddrTable 中找到 Broker 并将其中集群中移除,如果移除后集群中不包含任何 Broker,则将该集群从 clusterAddrTable 中移除

  1. if (removeBrokerName) {
  2. Set<String> nameSet = this.clusterAddrTable.get(clusterName);
  3. if (nameSet != null) {
  4. boolean removed = nameSet.remove(brokerName);
  5. log.info("unregisterBroker, remove name from clusterAddrTable {}, {}",removed ? "OK" : "Failed",brokerName);
  6. if (nameSet.isEmpty()) {
  7. this.clusterAddrTable.remove(clusterName);
  8. log.info("unregisterBroker, remove cluster from clusterAddrTable {}",clusterName);
  9. }
  10. }
  11. }

第四步: 根据 brokerName,遍历所有主题的队列,如果队列中包含当前 broker 的队列,则移除,如果 topic 中包含待移除的 Broker 的队列,从路由表中删除该 topic

  1. this.topicQueueTable.forEach((topic, queueDataMap) -> {
  2. QueueData old = queueDataMap.remove(brokerName);
  3. if (old != null) {
  4. log.info("removeTopicByBrokerName, remove one broker's topic {} {}", topic, old);
  5. }
  6. if (queueDataMap.size() == 0) {
  7. noBrokerRegisterTopic.add(topic);
  8. log.info("removeTopicByBrokerName, remove the topic all queue {}", topic);
  9. }
  10. });

第五步:

释放锁,完成路由删除

  1. finally {
  2. this.lock.writeLock().unlock();
  3. }