RocketMQ中PullConsumer的啟動源碼分析

通過DefaultMQPullConsumer作為預設實現，這裡的啟動過程和Producer很相似，但相比複雜一些【RocketMQ中Producer的啟動源碼分析】 DefaultMQPullConsumer的構造方法：這裡會封裝一個DefaultMQPullConsumerImpl，類似於P ...

通過DefaultMQPullConsumer作為預設實現，這裡的啟動過程和Producer很相似，但相比複雜一些

【RocketMQ中Producer的啟動源碼分析】

DefaultMQPullConsumer的構造方法：

1 public DefaultMQPullConsumer(final String consumerGroup, RPCHook rpcHook) {
2     this.consumerGroup = consumerGroup;
3     defaultMQPullConsumerImpl = new DefaultMQPullConsumerImpl(this, rpcHook);
4 }

這裡會封裝一個DefaultMQPullConsumerImpl，類似於Producer中DefaultMQProducerImpl

DefaultMQPullConsumerImpl：

 1 public class DefaultMQPullConsumerImpl implements MQConsumerInner {
 2     private final InternalLogger log = ClientLogger.getLog();
 3     private final DefaultMQPullConsumer defaultMQPullConsumer;
 4     private final long consumerStartTimestamp = System.currentTimeMillis();
 5     private final RPCHook rpcHook;
 6     private final ArrayList<ConsumeMessageHook> consumeMessageHookList = new ArrayList<ConsumeMessageHook>();
 7     private final ArrayList<FilterMessageHook> filterMessageHookList = new ArrayList<FilterMessageHook>();
 8     private volatile ServiceState serviceState = ServiceState.CREATE_JUST;
 9     private MQClientInstance mQClientFactory;
10     private PullAPIWrapper pullAPIWrapper;
11     private OffsetStore offsetStore;
12     private RebalanceImpl rebalanceImpl = new RebalancePullImpl(this);
13 
14     public DefaultMQPullConsumerImpl(final DefaultMQPullConsumer defaultMQPullConsumer, final RPCHook rpcHook) {
15         this.defaultMQPullConsumer = defaultMQPullConsumer;
16         this.rpcHook = rpcHook;
17     }
18     ......
19 }

如上會封裝這些東西，在後面遇到了再詳細介紹

而DefaultMQPullConsumer的start方法，其實際上調用的是DefaultMQPullConsumerImpl的start方法

DefaultMQPullConsumerImpl的start方法：

 1 public synchronized void start() throws MQClientException {
 2     switch (this.serviceState) {
 3         case CREATE_JUST:
 4             this.serviceState = ServiceState.START_FAILED;
 5 
 6             this.checkConfig();
 7 
 8             this.copySubscription();
 9 
10             if (this.defaultMQPullConsumer.getMessageModel() == MessageModel.CLUSTERING) {
11                 this.defaultMQPullConsumer.changeInstanceNameToPID();
12             }
13 
14             this.mQClientFactory = MQClientManager.getInstance().getAndCreateMQClientInstance(this.defaultMQPullConsumer, this.rpcHook);
15 
16             this.rebalanceImpl.setConsumerGroup(this.defaultMQPullConsumer.getConsumerGroup());
17             this.rebalanceImpl.setMessageModel(this.defaultMQPullConsumer.getMessageModel());
18             this.rebalanceImpl.setAllocateMessageQueueStrategy(this.defaultMQPullConsumer.getAllocateMessageQueueStrategy());
19             this.rebalanceImpl.setmQClientFactory(this.mQClientFactory);
20 
21             this.pullAPIWrapper = new PullAPIWrapper(
22                 mQClientFactory,
23                 this.defaultMQPullConsumer.getConsumerGroup(), isUnitMode());
24             this.pullAPIWrapper.registerFilterMessageHook(filterMessageHookList);
25 
26             if (this.defaultMQPullConsumer.getOffsetStore() != null) {
27                 this.offsetStore = this.defaultMQPullConsumer.getOffsetStore();
28             } else {
29                 switch (this.defaultMQPullConsumer.getMessageModel()) {
30                     case BROADCASTING:
31                         this.offsetStore = new LocalFileOffsetStore(this.mQClientFactory, this.defaultMQPullConsumer.getConsumerGroup());
32                         break;
33                     case CLUSTERING:
34                         this.offsetStore = new RemoteBrokerOffsetStore(this.mQClientFactory, this.defaultMQPullConsumer.getConsumerGroup());
35                         break;
36                     default:
37                         break;
38                 }
39                 this.defaultMQPullConsumer.setOffsetStore(this.offsetStore);
40             }
41 
42             this.offsetStore.load();
43 
44             boolean registerOK = mQClientFactory.registerConsumer(this.defaultMQPullConsumer.getConsumerGroup(), this);
45             if (!registerOK) {
46                 this.serviceState = ServiceState.CREATE_JUST;
47 
48                 throw new MQClientException("The consumer group[" + this.defaultMQPullConsumer.getConsumerGroup()
49                     + "] has been created before, specify another name please." + FAQUrl.suggestTodo(FAQUrl.GROUP_NAME_DUPLICATE_URL),
50                     null);
51             }
52 
53             mQClientFactory.start();
54             log.info("the consumer [{}] start OK", this.defaultMQPullConsumer.getConsumerGroup());
55             this.serviceState = ServiceState.RUNNING;
56             break;
57         case RUNNING:
58         case START_FAILED:
59         case SHUTDOWN_ALREADY:
60             throw new MQClientException("The PullConsumer service state not OK, maybe started once, "
61                 + this.serviceState
62                 + FAQUrl.suggestTodo(FAQUrl.CLIENT_SERVICE_NOT_OK),
63                 null);
64         default:
65             break;
66     }
67 
68 }

首先checkConfig方法會對配置做檢查

接著copySubscription方法：

 1 private void copySubscription() throws MQClientException {
 2     try {
 3         Set<String> registerTopics = this.defaultMQPullConsumer.getRegisterTopics();
 4         if (registerTopics != null) {
 5             for (final String topic : registerTopics) {
 6                 SubscriptionData subscriptionData = FilterAPI.buildSubscriptionData(this.defaultMQPullConsumer.getConsumerGroup(),
 7                     topic, SubscriptionData.SUB_ALL);
 8                 this.rebalanceImpl.getSubscriptionInner().put(topic, subscriptionData);
 9             }
10         }
11     } catch (Exception e) {
12         throw new MQClientException("subscription exception", e);
13     }
14 }

這裡的registerTopics是由用戶調用setRegisterTopics方法註冊進來的Topic集合
在這裡會將集合中的Topic包裝成SubscriptionData保存在rebalanceImpl中

SubscriptionData：

 1 public class SubscriptionData implements Comparable<SubscriptionData> {
 2     public final static String SUB_ALL = "*";
 3     private boolean classFilterMode = false;
 4     private String topic;
 5     private String subString;
 6     private Set<String> tagsSet = new HashSet<String>();
 7     private Set<Integer> codeSet = new HashSet<Integer>();
 8     private long subVersion = System.currentTimeMillis();
 9     private String expressionType = ExpressionType.TAG;
10     ......
11 }

RebalanceImpl：

 1 public abstract class RebalanceImpl {
 2     protected final ConcurrentMap<MessageQueue, ProcessQueue> processQueueTable = new ConcurrentHashMap<MessageQueue, ProcessQueue>(64);
 3     protected final ConcurrentMap<String/* topic */, Set<MessageQueue>> topicSubscribeInfoTable =
 4         new ConcurrentHashMap<String, Set<MessageQueue>>();
 5     protected final ConcurrentMap<String /* topic */, SubscriptionData> subscriptionInner =
 6         new ConcurrentHashMap<String, SubscriptionData>();
 7     protected String consumerGroup;
 8     protected MessageModel messageModel;
 9     protected AllocateMessageQueueStrategy allocateMessageQueueStrategy;
10     protected MQClientInstance mQClientFactory;
11     ......
12 }

回到start方法，接著和Producer中一樣通過MQClientManager獲取一個MQClientInstance
然後會完成對rebalanceImpl屬性的填充

接著會實例化一個PullAPIWrapper，同時向其註冊過濾器的鉤子，這個對象在之後分析消息拉取時詳細介紹

接下來會根據消息的模式，決定使用不同方式的OffsetStore

 1 public enum MessageModel {
 2     /**
 3      * broadcast
 4      */
 5     BROADCASTING("BROADCASTING"),
 6     /**
 7      * clustering
 8      */
 9     CLUSTERING("CLUSTERING");
10     ......
11 }

分別是廣播模式和集群模式
廣播模式（BROADCASTING）：同一個ConsumerGroup里的每個Consumer都能消費到所訂閱Topic的全部消息，也就是一個消息會被多次分發，被多個Consumer消費
集群模式（CLUSTERING）：同一個ConsumerGroup里的每個Consumer只消費所訂閱消息的一部分內容，同一個ConsumerGroup里所有的Consumer消費的內容合起來才是所訂閱Topic內容的整體

採用廣播模式，消費者的消費進度offset會被保存在本地；而採用集群模式，消費者的消費進度offset會被保存在遠端（broker）上
故廣播模式使用LocalFileOffsetStore，集群模式使用RemoteBrokerOffsetStore

在採用廣播模式，即LocalFileOffsetStore，調用load方法會對其配置文件offsets.json進行載入，而RemoteBrokerOffsetStore時沒意義的非同步操作
LocalFileOffsetStore的load方法：

 1 public void load() throws MQClientException {
 2     OffsetSerializeWrapper offsetSerializeWrapper = this.readLocalOffset();
 3     if (offsetSerializeWrapper != null && offsetSerializeWrapper.getOffsetTable() != null) {
 4         offsetTable.putAll(offsetSerializeWrapper.getOffsetTable());
 5 
 6         for (MessageQueue mq : offsetSerializeWrapper.getOffsetTable().keySet()) {
 7             AtomicLong offset = offsetSerializeWrapper.getOffsetTable().get(mq);
 8             log.info("load consumer's offset, {} {} {}",
 9                 this.groupName,
10                 mq,
11                 offset.get());
12         }
13     }
14 }

readLocalOffset方法會將offsets.json文件中的json字元串轉換成OffsetSerializeWrapper對象封裝

 1 public class OffsetSerializeWrapper extends RemotingSerializable {
 2     private ConcurrentMap<MessageQueue, AtomicLong> offsetTable =
 3         new ConcurrentHashMap<MessageQueue, AtomicLong>();
 4 
 5     public ConcurrentMap<MessageQueue, AtomicLong> getOffsetTable() {
 6         return offsetTable;
 7     }
 8 
 9     public void setOffsetTable(ConcurrentMap<MessageQueue, AtomicLong> offsetTable) {
10         this.offsetTable = offsetTable;
11     }
12 }

從這裡就可里大致理解json文件中的內容，其中AtomicLong就對應MessageQueue下具體的Offset
之後在load方法中，會將該map保存在LocalFileOffsetStore中的offsetTable中

接著會調用mQClientFactory的start方法，這個方法在【RocketMQ中Producer的啟動源碼分析】中進行過分析

 1 public void start() throws MQClientException {
 2     synchronized (this) {
 3         switch (this.serviceState) {
 4             case CREATE_JUST:
 5                 this.serviceState = ServiceState.START_FAILED;
 6                 // If not specified,looking address from name server
 7                 if (null == this.clientConfig.getNamesrvAddr()) {
 8                     this.mQClientAPIImpl.fetchNameServerAddr();
 9                 }
10                 // Start request-response channel
11                 this.mQClientAPIImpl.start();
12                 // Start various schedule tasks
13                 this.startScheduledTask();
14                 // Start pull service
15                 this.pullMessageService.start();
16                 // Start rebalance service
17                 this.rebalanceService.start();
18                 // Start push service
19                 this.defaultMQProducer.getDefaultMQProducerImpl().start(false);
20                 log.info("the client factory [{}] start OK", this.clientId);
21                 this.serviceState = ServiceState.RUNNING;
22                 break;
23             case RUNNING:
24                 break;
25             case SHUTDOWN_ALREADY:
26                 break;
27             case START_FAILED:
28                 throw new MQClientException("The Factory object[" + this.getClientId() + "] has been created before, and failed.", null);
29             default:
30                 break;
31         }
32     }
33 }

首先若是沒有設置NameServer的地址，會調用fetchNameServerAddr方法進行自動定址，詳見Producer的啟動

之後mQClientAPIImpl的start方法會完成對Netty客戶端的綁定操作，詳見Producer的啟動

startScheduledTask方法則會設置五個定時任務：
①若是名稱服務地址namesrvAddr不存在，則調用前面的fetchNameServerAddr方法，定時更新名稱服務
②定時更新Topic所對應的路由信息
③定時清除離線的Broker，以及向當前線上的Broker發送心跳包
（以上詳見Producer的啟動）

④定時持久化消費者隊列的消費進度
DefaultMQPullConsumerImpl中的實現：

 1 public void persistConsumerOffset() {
 2     try {
 3         this.makeSureStateOK();
 4         Set<MessageQueue> mqs = new HashSet<MessageQueue>();
 5         Set<MessageQueue> allocateMq = this.rebalanceImpl.getProcessQueueTable().keySet();
 6         mqs.addAll(allocateMq);
 7         this.offsetStore.persistAll(mqs);
 8     } catch (Exception e) {
 9         log.error("group: " + this.defaultMQPullConsumer.getConsumerGroup() + " persistConsumerOffset exception", e);
10     }
11 }

首先從rebalanceImpl中取出所有處理的消費隊列MessageQueue集合
然後調用offsetStore的persistAll方法進一步處理該集合

由於廣播模式和集群模式，所以這裡有兩種實現：
廣播模式LocalFileOffsetStore的persistAll方法：

 1 public void persistAll(Set<MessageQueue> mqs) {
 2     if (null == mqs || mqs.isEmpty())
 3         return;
 4 
 5     OffsetSerializeWrapper offsetSerializeWrapper = new OffsetSerializeWrapper();
 6     for (Map.Entry<MessageQueue, AtomicLong> entry : this.offsetTable.entrySet()) {
 7         if (mqs.contains(entry.getKey())) {
 8             AtomicLong offset = entry.getValue();
 9             offsetSerializeWrapper.getOffsetTable().put(entry.getKey(), offset);
10         }
11     }
12 
13     String jsonString = offsetSerializeWrapper.toJson(true);
14     if (jsonString != null) {
15         try {
16             MixAll.string2File(jsonString, this.storePath);
17         } catch (IOException e) {
18             log.error("persistAll consumer offset Exception, " + this.storePath, e);
19         }
20     }
21 }

這裡和之前的load方法相反，會將MessageQueue對應的offset信息替換掉原來的json文件中的內容
這樣就完成了廣播模式下定時持久化消費者隊列的消費進度

集群模式RemoteBrokerOffsetStore的persistAll方法的實現：

 1 public void persistAll(Set<MessageQueue> mqs) {
 2     if (null == mqs || mqs.isEmpty())
 3         return;
 4 
 5     final HashSet<MessageQueue> unusedMQ = new HashSet<MessageQueue>();
 6     if (!mqs.isEmpty()) {
 7         for (Map.Entry<MessageQueue, AtomicLong> entry : this.offsetTable.entrySet()) {
 8             MessageQueue mq = entry.getKey();
 9             AtomicLong offset = entry.getValue();
10             if (offset != null) {
11                 if (mqs.contains(mq)) {
12                     try {
13                         this.updateConsumeOffsetToBroker(mq, offset.get());
14                         log.info("[persistAll] Group: {} ClientId: {} updateConsumeOffsetToBroker {} {}",
15                             this.groupName,
16                             this.mQClientFactory.getClientId(),
17                             mq,
18                             offset.get());
19                     } catch (Exception e) {
20                         log.error("updateConsumeOffsetToBroker exception, " + mq.toString(), e);
21                     }
22                 } else {
23                     unusedMQ.add(mq);
24                 }
25             }
26         }
27     }
28 
29     if (!unusedMQ.isEmpty()) {
30         for (MessageQueue mq : unusedMQ) {
31             this.offsetTable.remove(mq);
32             log.info("remove unused mq, {}, {}", mq, this.groupName);
33         }
34     }
35 }

和上面類似，遍歷offsetTable中的內容，只不過不是保存在了本地，而是通過updateConsumeOffsetToBroker向Broker發送
updateConsumeOffsetToBroker方法：

 1 private void updateConsumeOffsetToBroker(MessageQueue mq, long offset) throws RemotingException,
 2     MQBrokerException, InterruptedException, MQClientException {
 3     updateConsumeOffsetToBroker(mq, offset, true);
 4 }
 5 
 6 public void updateConsumeOffsetToBroker(MessageQueue mq, long offset, boolean isOneway) throws RemotingException,
 7     MQBrokerException, InterruptedException, MQClientException {
 8     FindBrokerResult findBrokerResult = this.mQClientFactory.findBrokerAddressInAdmin(mq.getBrokerName());
 9     if (null == findBrokerResult) {
10 
11         this.mQClientFactory.updateTopicRouteInfoFromNameServer(mq.getTopic());
12         findBrokerResult = this.mQClientFactory.findBrokerAddressInAdmin(mq.getBrokerName());
13     }
14 
15     if (findBrokerResult != null) {
16         UpdateConsumerOffsetRequestHeader requestHeader = new UpdateConsumerOffsetRequestHeader();
17         requestHeader.setTopic(mq.getTopic());
18         requestHeader.setConsumerGroup(this.groupName);
19         requestHeader.setQueueId(mq.getQueueId());
20         requestHeader.setCommitOffset(offset);
21 
22         if (isOneway) {
23             this.mQClientFactory.getMQClientAPIImpl().updateConsumerOffsetOneway(
24                 findBrokerResult.getBrokerAddr(), requestHeader, 1000 * 5);
25         } else {
26             this.mQClientFactory.getMQClientAPIImpl().updateConsumerOffset(
27                 findBrokerResult.getBrokerAddr(), requestHeader, 1000 * 5);
28         }
29     } else {
30         throw new MQClientException("The broker[" + mq.getBrokerName() + "] not exist", null);
31     }
32 }

首先根據BrokerName查找Broker的路由信息：

 1 public FindBrokerResult findBrokerAddressInAdmin(final String brokerName) {
 2     String brokerAddr = null;
 3     boolean slave = false;
 4     boolean found = false;
 5 
 6     HashMap<Long/* brokerId */, String/* address */> map = this.brokerAddrTable.get(brokerName);
 7     if (map != null && !map.isEmpty()) {
 8         for (Map.Entry<Long, String> entry : map.entrySet()) {
 9             Long id = entry.getKey();
10             brokerAddr = entry.getValue();
11             if (brokerAddr != null) {
12                 found = true;
13                 if (MixAll.MASTER_ID == id) {
14                     slave = false;
15                 } else {
16                     slave = true;
17                 }
18                 break;
19 
20             }
21         } // end of for
22     }
23 
24     if (found) {
25         return new FindBrokerResult(brokerAddr, slave, findBrokerVersion(brokerName, brokerAddr));
26     }
27 
28     return null;
29 }

brokerAddrTable中的borker的路由信息會由 ②定時更新Topic所對應的路由信息，來完成更新，在brokerAddrTable中只要找的一個Broker的信息後，將其封裝為FindBrokerResult返回

若是沒有找到會執行updateTopicRouteInfoFromNameServer方法，也就是執行了一次定時任務中的方法，立即更新一次，再通過findBrokerAddressInAdmin方法，重新查找

找到之後，實例化一個請求頭 UpdateConsumerOffsetRequestHeader，將相應信息封裝，由於使用的是Oneway模式，所以這裡採用updateConsumerOffsetOneway方法，通過Netty向Broker發送

 1 public void updateConsumerOffsetOneway(
 2     final String addr,
 3     final UpdateConsumerOffsetRequestHeader requestHeader,
 4     final long timeoutMillis
 5 ) throws RemotingConnectException, RemotingTooMuchRequestException, RemotingTimeoutException, RemotingSendRequestException,
 6     InterruptedException {
 7     RemotingCommand request = RemotingCommand.createRequestCommand(RequestCode.UPDATE_CONSUMER_OFFSET, requestHeader);
 8 
 9     this.remotingClient.invokeOneway(MixAll.brokerVIPChannel(this.clientConfig.isVipChannelEnabled(), addr), request, timeoutMillis);
10 }

其實這裡就非常簡單地調用了invokeOneway方法，完成向Broker的消息單向發送

【RocketMQ中Producer消息的發送源碼分析】

非OneWay則採用同步發送
這樣，在集群模式下，消費進度也就交給了Broker管理，之後的負載均衡以此為基礎

⑤定時調整消費者端的線程池的大小
這裡針對的是PushConsumer，後續博客再介紹

對於PullConsumer來說rebalanceService服務的開啟才是最重要的

RebalanceService：

 1 public void run() {
 2     log.info(this.getServiceName() + " service started");
 3 
 4     while (!this.isStopped()) {
 5         this.waitForRunning(waitInterval);
 6         this.mqClientFactory.doRebalance();
 7     }
 8 
 9     log.info(this.getServiceName() + " service end");
10 }

這裡的waitForRunning和Broker的刷盤以及主從複製類似，會進行超時阻塞（預設20s），也可以通過Broker發送的NOTIFY_CONSUMER_IDS_CHANGED請求將其喚醒，之後會調用doRebalance方法

RebalanceImpl的doRebalance方法：

 1 public void doRebalance(final boolean isOrder) {
 2    Map<String, SubscriptionData> subTable = this.getSubscriptionInner();
 3     if (subTable != null) {
 4         for (final Map.Entry<String, SubscriptionData> entry : subTable.entrySet()) {
 5             final String topic = entry.getKey();
 6             try {
 7                 this.rebalanceByTopic(topic, isOrder);
 8             } catch (Throwable e) {
 9                 if (!topic.startsWith(MixAll.RETRY_GROUP_TOPIC_PREFIX)) {
10                     log.warn("rebalanceByTopic Exception", e);
11                 }
12             }
13         }
14     }
15 
16     this.truncateMessageQueueNotMyTopic();
17 }

這裡就會取得copySubscription方法中說過的訂閱Topic集合，這個集合會在②中的定時任務會通過NameServer來進行更新

通過rebalanceByTopic方法，處理訂閱的Topic：

 1 private void rebalanceByTopic(final String topic, final boolean isOrder) {
 2     switch (messageModel) {
 3         case BROADCASTING: {
 4             Set<MessageQueue> mqSet = this.topicSubscribeInfoTable.get(topic);
 5             if (mqSet != null) {
 6                 boolean changed = this.updateProcessQueueTableInRebalance(topic, mqSet, isOrder);
 7                 if (changed) {
 8                     this.messageQueueChanged(topic, mqSet, mqSet);
 9                     log.info("messageQueueChanged {} {} {} {}",
10                         consumerGroup,
11                         topic,
12                         mqSet,
13                         mqSet);
14                 }
15             } else {
16                 log.warn("doRebalance, {}, but the topic[{}] not exist.", consumerGroup, topic);
17             }
18             break;
19         }
20         case CLUSTERING: {
21             Set<MessageQueue> mqSet = this.topicSubscribeInfoTable.get(topic);