一、问题复现
1、描述
两个一样的Consumer Group的Consumer订阅同一个Topic,但是是不同的tag,Consumer1订阅Topic的tag1,Consumer2订阅Topic的tag2,然后分别启动。这时候往Topic的tag1里发送10条数据,Topic的tag2里发送10条。目测应该是Consumer1和Consumer2分别收到对应的10条消息。结果却是只有Consumer2收到了消息,而且只收到了4-6条消息,不固定。
2、代码
2.1、Consumer
public class Consumer { public static void main(String[] args) throws MQClientException { DefaultMQPushConsumer consumer = new DefaultMQPushConsumer("test-consumer"); consumer.setNamesrvAddr("124.57.180.156:9876"); consumer.subscribe("TopicTest2","tag1"); consumer.registerMessageListener(new MessageListenerConcurrently() { public ConsumeConcurrentlyStatus consumeMessage( List<MessageExt> msgs, ConsumeConcurrentlyContext context) { MessageExt msg = msgs.get(0); System.out.println(msg.getTags()); return ConsumeConcurrentlyStatus.CONSUME_SUCCESS; } }); consumer.start(); System.out.println("ConsumerStarted."); } }
启动这个订阅了TopicTest2的tag1标签的Consumer,然后将tag1改为tag2再次启动Consumer。这就相当于启动了两个Consumer进程,一个订阅了TopicTest2的tag1标签,另一个订阅了TopicTest2的tag2标签。
2.2、Producer
public class Producer { public static void main(String[] args) throws MQClientException { final DefaultMQProducer producer = new DefaultMQProducer("test-producer"); producer.setNamesrvAddr("124.57.180.156:9876"); producer.start(); for (int i = 0; i < 10; i++){ try { Message msg = new Message("TopicTest2", "tag1", ("Hello tag1 - "+i).getBytes()); SendResult sendResult = producer.send(msg); System.out.println(sendResult); }catch(Exception e) { e.printStackTrace(); } } } }
启动Producer,往TopicTest2的tag1里发10条消息。再次将tag1改为tag2,然后再次启动Producer进行发送,这样就是TopicTest2的tag1下有10条消息,TopicTest2的tag2下也有10条消息。
3、结果
Consumer和Producer都启动后,发现如下:
- Producer发送了20条消息正常。
- Consumer1没有消费到tag1下的数据
- Consumer2消费了一半(不一定是几条,有时候5条,有时候6条的)消息。
二、问题答案
- 首先这是Broker决定的,而不是Consumer端决定的
之前看过一篇文章写的有理有据,写的是Consumer端,还贴出了debug的源码,说后者覆盖了前者,但是我想说:你启动了两个独立的Consumer,那是两个独立的进程,根本不存在覆盖不覆盖的问题,那就是独立的。JVM就一个。又不是共享的JVM,何来覆盖?
- Consumer端发心跳给Broker,Broker收到后存到consumerTable里(就是个Map),key是GroupName,value是ConsumerGroupInfo。
- ConsumerGroupInfo里面是包含topic等信息的,但是问题就出在上一步骤,key是groupName,你同GroupName的话Broker心跳最后收到的Consumer会覆盖前者的。相当于如下代码:
map.put(groupName, ConsumerGroupInfo);
这样同key,肯定产生了覆盖。所以Consumer1不会收到任何消息,但是Consumer2为什么只收到了一半(不固定)消息呢?
那是因为:你是集群模式消费,它会负载均衡分配到各个节点去消费,所以一半消息(不固定个数)跑到了Consumer1上,结果Consumer1订阅的是tag1,所以不会任何输出。
如果换成BROADCASTING,那绝逼后者会收到全部消息,而不是一半,因为广播是广播全部Consumer。
三、源码验证
1、调用链
# 核心在于如下这个方法 org.apache.rocketmq.broker.client.ConsumerManager#registerConsumer() # 关键调用链如下 # 入口是Broker启动的时候 org.apache.rocketmq.broker.BrokerStartup#start() org.apache.rocketmq.broker.BrokerController#start() org.apache.rocketmq.remoting.netty.NettyRemotingServer#start() org.apache.rocketmq.remoting.netty.NettyRemotingServer#prepareSharableHandlers() org.apache.rocketmq.remoting.netty.NettyRemotingServer.NettyServerHandler#channelRead0() org.apache.rocketmq.remoting.netty.NettyRemotingAbstract#processMessageReceived() org.apache.rocketmq.remoting.netty.NettyRemotingAbstract#processRequestCommand() org.apache.rocketmq.broker.processor.ClientManageProcessor#processRequest() org.apache.rocketmq.broker.processor.ClientManageProcessor#heartBeat() org.apache.rocketmq.broker.client.ConsumerManager#registerConsumer()
2、源码
2.1、registerConsumer
/** * Consumer信息 */ public class ConsumerGroupInfo { // 组名 private final String groupName; // topic信息,比如topic、tag等 private final ConcurrentMap<String/* Topic */, SubscriptionData> subscriptionTable = new ConcurrentHashMap<String, SubscriptionData>(); // 客户端信息,比如clientId等 private final ConcurrentMap<Channel, ClientChannelInfo> channelInfoTable = new ConcurrentHashMap<Channel, ClientChannelInfo>(16); // PULL/PUSH private volatile ConsumeType consumeType; // 消费模式:BROADCASTING/CLUSTERING private volatile MessageModel messageModel; // 消费到哪了 private volatile ConsumeFromWhere consumeFromWhere; } /** * 通过心跳将Consumer信息注册到Broker端。 */ public boolean registerConsumer(final String group, final ClientChannelInfo clientChannelInfo, ConsumeType consumeType, MessageModel messageModel, ConsumeFromWhere consumeFromWhere, final Set<SubscriptionData> subList, boolean isNotifyConsumerIdsChangedEnable) { // consumerTable:维护所有的Consumer ConsumerGroupInfo consumerGroupInfo = this.consumerTable.get(group); // 如果没有Consumer,则put到map里 if (null == consumerGroupInfo) { ConsumerGroupInfo tmp = new ConsumerGroupInfo(group, consumeType, messageModel, consumeFromWhere); // put到map里 ConsumerGroupInfo prev = this.consumerTable.putIfAbsent(group, tmp); consumerGroupInfo = prev != null ? prev : tmp; } // 更新Consumer信息,客户端信息 boolean r1 = consumerGroupInfo.updateChannel(clientChannelInfo, consumeType, messageModel, consumeFromWhere); // 更新订阅Topic信息 boolean r2 = consumerGroupInfo.updateSubscription(subList); if (r1 || r2) { if (isNotifyConsumerIdsChangedEnable) { this.consumerIdsChangeListener.handle(ConsumerGroupEvent.CHANGE, group, consumerGroupInfo.getAllChannel()); } } this.consumerIdsChangeListener.handle(ConsumerGroupEvent.REGISTER, group, subList); return r1 || r2; }
从这一步可以看出消费者信息是以groupName为key,ConsumerGroupInfo为value存到map(consumerTable)里的,那很明显了,后者肯定会覆盖前者的,因为key是一样的。而后者的tag是tag2,那肯定覆盖了前者的tag1,这部分是存到ConsumerGroupInfo的subscriptionTable里面的
private final ConcurrentMap<String/* Topic */, SubscriptionData> subscriptionTable = new ConcurrentHashMap<String, SubscriptionData>();
SubscriptionData包含了topic等信息
public class SubscriptionData implements Comparable<SubscriptionData> { // topic private String topic; private String subString; // tags private Set<String> tagsSet = new HashSet<String>(); private Set<Integer> codeSet = new HashSet<Integer>(); }
2.2、两个问题
1.topic、tag等信息是怎么覆盖的?
boolean r1 = consumerGroupInfo.updateChannel(clientChannelInfo, consumeType, messageModel,consumeFromWhere);
/** * 其实很简单,就是以topic为key,SubscriptionData为value。而SubscriptionData里包含了tags信息,所以直接覆盖掉 */ public boolean updateSubscription(final Set<SubscriptionData> subList) { for (SubscriptionData sub : subList) { SubscriptionData old = this.subscriptionTable.get(sub.getTopic()); if (old == null) { SubscriptionData prev = this.subscriptionTable.putIfAbsent(sub.getTopic(), sub); } else if (sub.getSubVersion() > old.getSubVersion()) { this.subscriptionTable.put(sub.getTopic(), sub); } } }
等等,这里好像有新发现ConsumerGroupInfo#subscriptionTable
// {@link org.apache.rocketmq.broker.client.ConsumerGroupInfo#subscriptionTable} private final ConcurrentMap<String/* Topic */, SubscriptionData> subscriptionTable = new ConcurrentHashMap<String, SubscriptionData>();
可以有意外收获就是topic作为map的key,那岂不是一个Consumer可以订阅多个Topic?是的,通过这段源码可以发现是没毛病的,我也测试过。
2.这么看的话Consumer端只会存在一个进程,因为同组,注册进去就覆盖了呀?
大哥,注意ConsumerGroupInfo里的channelInfoTable
// 客户端信息,比如clientId等 private final ConcurrentMap<Channel, ClientChannelInfo> channelInfoTable = new ConcurrentHashMap<Channel, ClientChannelInfo>(16);
ClientChannelInfo是包含clientId等信息的,代表一个Consumer。注册方法是:
boolean r2 = consumerGroupInfo.updateSubscription(subList);
/** * 下面是删减后的代码,其实就是以Channel作为key,每个Consumer的Channel是不一样的。所以能存多个Consumer客户端 */ public boolean updateChannel(final ClientChannelInfo infoNew, ConsumeType consumeType, MessageModel messageModel, ConsumeFromWhere consumeFromWhere) { ClientChannelInfo infoOld = this.channelInfoTable.get(infoNew.getChannel()); if (null == infoOld) { ClientChannelInfo prev = this.channelInfoTable.put(infoNew.getChannel(), infoNew); } }
END