摘要: 原创出处 http://www.iocoder.cn/RocketMQ/high-availability/ 「芋道源码」欢迎转载，保留摘要，谢谢！
本文主要基于 RocketMQ 4.0.x 正式版
1. 概述
本文主要解析 Namesrv、Broker 如何实现高可用，Producer、Consumer 怎么与它们通信保证高可用。
2. Namesrv 高可用
启动多个 Namesrv 实现高可用。
相较于 Zookeeper、Consul、Etcd 等，Namesrv 是一个超轻量级的注册中心，提供命名服务。
2.1 Broker 注册到 Namesrv
- 📌 多个
Namesrv之间，没有任何关系（不存在类似Zookeeper的Leader/Follower等角色），不进行通信与数据同步。通过Broker循环注册多个Namesrv。
1: // ⬇️⬇️⬇️【BrokerOuterAPI.java】
2: public RegisterBrokerResult registerBrokerAll(
3: final String clusterName,
4: final String brokerAddr,
5: final String brokerName,
6: final long brokerId,
7: final String haServerAddr,
8: final TopicConfigSerializeWrapper topicConfigWrapper,
9: final List<String> filterServerList,
10: final boolean oneway,
11: final int timeoutMills) {
12: RegisterBrokerResult registerBrokerResult = null;
13:
14: List<String> nameServerAddressList = this.remotingClient.getNameServerAddressList();
15: if (nameServerAddressList != null) {
16: for (String namesrvAddr : nameServerAddressList) { // 循环多个 Namesrv
17: try {
18: RegisterBrokerResult result = this.registerBroker(namesrvAddr, clusterName, brokerAddr, brokerName, brokerId,
19: haServerAddr, topicConfigWrapper, filterServerList, oneway, timeoutMills);
20: if (result != null) {
21: registerBrokerResult = result;
22: }
23:
24: log.info("register broker to name server {} OK", namesrvAddr);
25: } catch (Exception e) {
26: log.warn("registerBroker Exception, {}", namesrvAddr, e);
27: }
28: }
29: }
30:
31: return registerBrokerResult;
32: }
2.2 Producer、Consumer 访问 Namesrv
- 📌
Producer、Consumer从Namesrv列表选择一个可连接的进行通信。
1: // ⬇️⬇️⬇️【NettyRemotingClient.java】
2: private Channel getAndCreateNameserverChannel() throws InterruptedException {
3: // 返回已选择、可连接Namesrv
4: String addr = this.namesrvAddrChoosed.get();
5: if (addr != null) {
6: ChannelWrapper cw = this.channelTables.get(addr);
7: if (cw != null && cw.isOK()) {
8: return cw.getChannel();
9: }
10: }
11: //
12: final List<String> addrList = this.namesrvAddrList.get();
13: if (this.lockNamesrvChannel.tryLock(LOCK_TIMEOUT_MILLIS, TimeUnit.MILLISECONDS)) {
14: try {
15: // 返回已选择、可连接的Namesrv
16: addr = this.namesrvAddrChoosed.get();
17: if (addr != null) {
18: ChannelWrapper cw = this.channelTables.get(addr);
19: if (cw != null && cw.isOK()) {
20: return cw.getChannel();
21: }
22: }
23: // 从【Namesrv列表】中选择一个连接的返回
24: if (addrList != null && !addrList.isEmpty()) {
25: for (int i = 0; i < addrList.size(); i++) {
26: int index = this.namesrvIndex.incrementAndGet();
27: index = Math.abs(index);
28: index = index % addrList.size();
29: String newAddr = addrList.get(index);
30:
31: this.namesrvAddrChoosed.set(newAddr);
32: Channel channelNew = this.createChannel(newAddr);
33: if (channelNew != null)
34: return channelNew;
35: }
36: }
37: } catch (Exception e) {
38: log.error("getAndCreateNameserverChannel: create name server channel exception", e);
39: } finally {
40: this.lockNamesrvChannel.unlock();
41: }
42: } else {
43: log.warn("getAndCreateNameserverChannel: try to lock name server, but timeout, {}ms", LOCK_TIMEOUT_MILLIS);
44: }
45:
46: return null;
47: }
3. Broker 高可用
启动多个 Broker分组 形成 集群 实现高可用。
Broker分组 = Master节点x1 + Slave节点xN。
类似 MySQL，Master节点 提供读写服务，Slave节点 只提供读服务。
3.2 Broker 主从
- 每个分组，
Master节点 不断发送新的CommitLog给Slave节点。Slave节点 不断上报本地的CommitLog已经同步到的位置给Master节点。 Broker分组与Broker分组之间没有任何关系，不进行通信与数据同步。- 消费进度 目前不支持
Master/Slave同步。org.apache.rocketmq.broker.slave.SlaveSynchronize类，Slave节点会从Master节点拉取消费进度、Topic 配置等等。
集群内，Master节点 有两种类型：Master_SYNC、Master_ASYNC：前者在 Producer 发送消息时，等待 Slave节点 存储完毕后再返回发送结果，而后者不需要等待。
3.1.1 配置
目前官方提供三套配置：
- 2m-2s-async
| brokerClusterName | brokerName | brokerRole | brokerId |
|---|---|---|---|
| DefaultCluster | broker-a | ASYNC_MASTER | 0 |
| DefaultCluster | broker-a | SLAVE | 1 |
| DefaultCluster | broker-b | ASYNC_MASTER | 0 |
| DefaultCluster | broker-b | SLAVE | 1 |
- 2m-2s-sync
| brokerClusterName | brokerName | brokerRole | brokerId |
|---|---|---|---|
| DefaultCluster | broker-a | SYNC_MASTER | 0 |
| DefaultCluster | broker-a | SLAVE | 1 |
| DefaultCluster | broker-b | SYNC_MASTER | 0 |
| DefaultCluster | broker-b | SLAVE | 1 |
- 2m-noslave
| brokerClusterName | brokerName | brokerRole | brokerId |
|---|---|---|---|
| DefaultCluster | broker-a | ASYNC_MASTER | 0 |
| DefaultCluster | broker-b | ASYNC_MASTER | 0 |
3.1.2 组件
再看具体实现代码之前，我们来看看 Master/Slave节点 包含的组件：
Master节点AcceptSocketService：接收Slave节点 连接。HAConnectionReadSocketService：读来自Slave节点 的数据。WriteSocketService：写到往Slave节点 的数据。
Slave节点HAClient：对Master节点 连接、读写数据。
3.1.3 通信协议
Master节点 与 Slave节点 通信协议很简单，只有如下两条。
| 对象 | 用途 | 第几位 | 字段 | 数据类型 | 字节数 | 说明 |
|---|---|---|---|---|---|---|
| Slave=>Master | 上报CommitLog已经同步到的物理位置 | |||||
| 0 | maxPhyOffset | Long | 8 | CommitLog最大物理位置 | ||
| Master=>Slave | 传输新的 CommitLog 数据 |
|||||
| 0 | fromPhyOffset | Long | 8 | CommitLog开始物理位置 | ||
| 1 | size | Int | 4 | 传输CommitLog数据长度 | ||
| 2 | body | Bytes | size | 传输CommitLog数据 |
3.1.4 Slave
Slave主循环，实现了不断不断不断从Master传输CommitLog数据，上传Master自己本地的CommitLog已经同步物理位置。
1: // ⬇️⬇️⬇️【HAClient.java】
2: public void run() {
3: log.info(this.getServiceName() + " service started");
4:
5: while (!this.isStopped()) {
6: try {
7: if (this.connectMaster()) {
8: // 若到满足上报间隔，上报到Master进度
9: if (this.isTimeToReportOffset()) {
10: boolean result = this.reportSlaveMaxOffset(this.currentReportedOffset);
11: if (!result) {
12: this.closeMaster();
13: }
14: }
15:
16: this.selector.select(1000);
17:
18: // 处理读取事件
19: boolean ok = this.processReadEvent();
20: if (!ok) {
21: this.closeMaster();
22: }
23:
24: // 若进度有变化，上报到Master进度
25: if (!reportSlaveMaxOffsetPlus()) {
26: continue;
27: }
28:
29: // Master过久未返回数据，关闭连接
30: long interval = HAService.this.getDefaultMessageStore().getSystemClock().now() - this.lastWriteTimestamp;
31: if (interval > HAService.this.getDefaultMessageStore().getMessageStoreConfig()
32: .getHaHousekeepingInterval()) {
33: log.warn("HAClient, housekeeping, found this connection[" + this.masterAddress
34: + "] expired, " + interval);
35: this.closeMaster();
36: log.warn("HAClient, master not response some time, so close connection");
37: }
38: } else {
39: this.waitForRunning(1000 * 5);
40: }
41: } catch (Exception e) {
42: log.warn(this.getServiceName() + " service has exception. ", e);
43: this.waitForRunning(1000 * 5);
44: }
45: }
46:
47: log.info(this.getServiceName() + " service end");
48: }
- 第 8 至 14 行 ：固定间隔（默认5s）向
Master上报Slave本地CommitLog已经同步到的物理位置。该操作还有心跳的作用。 - 第 16 至 22 行 ：处理
Master传输Slave的CommitLog数据。
- 我们来看看
#dispatchReadRequest(...)与#reportSlaveMaxOffset(...)是怎么实现的。
1: // 【HAClient.java】
2: /**
3: * 读取Master传输的CommitLog数据，并返回是异常
4: * 如果读取到数据，写入CommitLog
5: * 异常原因：
6: * 1. Master传输来的数据offset 不等于 Slave的CommitLog数据最大offset
7: * 2. 上报到Master进度失败
8: *
9: * @return 是否异常
10: */
11: private boolean dispatchReadRequest() {
12: final int msgHeaderSize = 8 + 4; // phyoffset + size
13: int readSocketPos = this.byteBufferRead.position();
14:
15: while (true) {
16: // 读取到请求
17: int diff = this.byteBufferRead.position() - this.dispatchPostion;
18: if (diff >= msgHeaderSize) {
19: // 读取masterPhyOffset、bodySize。使用dispatchPostion的原因是：处理数据“粘包”导致数据读取不完整。
20: long masterPhyOffset = this.byteBufferRead.getLong(this.dispatchPostion);
21: int bodySize = this.byteBufferRead.getInt(this.dispatchPostion + 8);
22: // 校验 Master传输来的数据offset 是否和 Slave的CommitLog数据最大offset 是否相同。
23: long slavePhyOffset = HAService.this.defaultMessageStore.getMaxPhyOffset();
24: if (slavePhyOffset != 0) {
25: if (slavePhyOffset != masterPhyOffset) {
26: log.error("master pushed offset not equal the max phy offset in slave, SLAVE: "
27: + slavePhyOffset + " MASTER: " + masterPhyOffset);
28: return false;
29: }
30: }
31: // 读取到消息
32: if (diff >= (msgHeaderSize + bodySize)) {
33: // 写入CommitLog
34: byte[] bodyData = new byte[bodySize];
35: this.byteBufferRead.position(this.dispatchPostion + msgHeaderSize);
36: this.byteBufferRead.get(bodyData);
37: HAService.this.defaultMessageStore.appendToCommitLog(masterPhyOffset, bodyData);
38: // 设置处理到的位置
39: this.byteBufferRead.position(readSocketPos);
40: this.dispatchPostion += msgHeaderSize + bodySize;
41: // 上报到Master进度
42: if (!reportSlaveMaxOffsetPlus()) {
43: return false;
44: }
45: // 继续循环
46: continue;
47: }
48: }
49:
50: // 空间写满，重新分配空间
51: if (!this.byteBufferRead.hasRemaining()) {
52: this.reallocateByteBuffer();
53: }
54:
55: break;
56: }
57:
58: return true;
59: }
60:
61: /**
62: * 上报进度
63: *
64: * @param maxOffset 进度
65: * @return 是否上报成功
66: */
67: private boolean reportSlaveMaxOffset(final long maxOffset) {
68: this.reportOffset.position(0);
69: this.reportOffset.limit(8);
70: this.reportOffset.putLong(maxOffset);
71: this.reportOffset.position(0);
72: this.reportOffset.limit(8);
73:
74: for (int i = 0; i < 3 && this.reportOffset.hasRemaining(); i++) {
75: try {
76: this.socketChannel.write(this.reportOffset);
77: } catch (IOException e) {
78: log.error(this.getServiceName()
79: + "reportSlaveMaxOffset this.socketChannel.write exception", e);
80: return false;
81: }
82: }
83:
84: return !this.reportOffset.hasRemaining();
85: }
3.1.5 Master
ReadSocketService逻辑同HAClient#processReadEvent(...)基本相同，我们直接看代码。
1: // ⬇️⬇️⬇️【ReadSocketService.java】
2: private boolean processReadEvent() {
3: int readSizeZeroTimes = 0;
4:
5: // 清空byteBufferRead
6: if (!this.byteBufferRead.hasRemaining()) {
7: this.byteBufferRead.flip();
8: this.processPostion = 0;
9: }
10:
11: while (this.byteBufferRead.hasRemaining()) {
12: try {
13: int readSize = this.socketChannel.read(this.byteBufferRead);
14: if (readSize > 0) {
15: readSizeZeroTimes = 0;
16:
17: // 设置最后读取时间
18: this.lastReadTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now();
19:
20: if ((this.byteBufferRead.position() - this.processPostion) >= 8) {
21: // 读取Slave 请求来的CommitLog的最大位置
22: int pos = this.byteBufferRead.position() - (this.byteBufferRead.position() % 8);
23: long readOffset = this.byteBufferRead.getLong(pos - 8);
24: this.processPostion = pos;
25:
26: // 设置Slave CommitLog的最大位置
27: HAConnection.this.slaveAckOffset = readOffset;
28:
29: // 设置Slave 第一次请求的位置
30: if (HAConnection.this.slaveRequestOffset < 0) {
31: HAConnection.this.slaveRequestOffset = readOffset;
32: log.info("slave[" + HAConnection.this.clientAddr + "] request offset " + readOffset);
33: }
34:
35: // 通知目前Slave进度。主要用于Master节点为同步类型的。
36: HAConnection.this.haService.notifyTransferSome(HAConnection.this.slaveAckOffset);
37: }
38: } else if (readSize == 0) {
39: if (++readSizeZeroTimes >= 3) {
40: break;
41: }
42: } else {
43: log.error("read socket[" + HAConnection.this.clientAddr + "] < 0");
44: return false;
45: }
46: } catch (IOException e) {
47: log.error("processReadEvent exception", e);
48: return false;
49: }
50: }
51:
52: return true;
53: }
WriteSocketService计算Slave开始同步的位置后，不断向Slave传输新的CommitLog数据。
1: // ⬇️⬇️⬇️【WriteSocketService.java】
2: @Override
3: public void run() {
4: HAConnection.log.info(this.getServiceName() + " service started");
5:
6: while (!this.isStopped()) {
7: try {
8: this.selector.select(1000);
9:
10: // 未获得Slave读取进度请求，sleep等待。
11: if (-1 == HAConnection.this.slaveRequestOffset) {
12: Thread.sleep(10);
13: continue;
14: }
15:
16: // 计算初始化nextTransferFromWhere
17: if (-1 == this.nextTransferFromWhere) {
18: if (0 == HAConnection.this.slaveRequestOffset) {
19: long masterOffset = HAConnection.this.haService.getDefaultMessageStore().getCommitLog().getMaxOffset();
20: masterOffset = masterOffset - (masterOffset % HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getMapedFileSizeCommitLog());
21: if (masterOffset < 0) {
22: masterOffset = 0;
23: }
24:
25: this.nextTransferFromWhere = masterOffset;
26: } else {
27: this.nextTransferFromWhere = HAConnection.this.slaveRequestOffset;
28: }
29:
30: log.info("master transfer data from " + this.nextTransferFromWhere + " to slave[" + HAConnection.this.clientAddr
31: + "], and slave request " + HAConnection.this.slaveRequestOffset);
32: }
33:
34: if (this.lastWriteOver) {
35: long interval = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now() - this.lastWriteTimestamp;
36: if (interval > HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaSendHeartbeatInterval()) { // 心跳
37:
38: // Build Header
39: this.byteBufferHeader.position(0);
40: this.byteBufferHeader.limit(headerSize);
41: this.byteBufferHeader.putLong(this.nextTransferFromWhere);
42: this.byteBufferHeader.putInt(0);
43: this.byteBufferHeader.flip();
44:
45: this.lastWriteOver = this.transferData();
46: if (!this.lastWriteOver)
47: continue;
48: }
49: } else { // 未传输完成，继续传输
50: this.lastWriteOver = this.transferData();
51: if (!this.lastWriteOver)
52: continue;
53: }
54:
55: // 选择新的CommitLog数据进行传输
56: SelectMappedBufferResult selectResult =
57: HAConnection.this.haService.getDefaultMessageStore().getCommitLogData(this.nextTransferFromWhere);
58: if (selectResult != null) {
59: int size = selectResult.getSize();
60: if (size > HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaTransferBatchSize()) {
61: size = HAConnection.this.haService.getDefaultMessageStore().getMessageStoreConfig().getHaTransferBatchSize();
62: }
63:
64: long thisOffset = this.nextTransferFromWhere;
65: this.nextTransferFromWhere += size;
66:
67: selectResult.getByteBuffer().limit(size);
68: this.selectMappedBufferResult = selectResult;
69:
70: // Build Header
71: this.byteBufferHeader.position(0);
72: this.byteBufferHeader.limit(headerSize);
73: this.byteBufferHeader.putLong(thisOffset);
74: this.byteBufferHeader.putInt(size);
75: this.byteBufferHeader.flip();
76:
77: this.lastWriteOver = this.transferData();
78: } else { // 没新的消息，挂起等待
79: HAConnection.this.haService.getWaitNotifyObject().allWaitForRunning(100);
80: }
81: } catch (Exception e) {
82:
83: HAConnection.log.error(this.getServiceName() + " service has exception.", e);
84: break;
85: }
86: }
87:
88: // 断开连接 & 暂停写线程 & 暂停读线程 & 释放CommitLog
89: if (this.selectMappedBufferResult != null) {
90: this.selectMappedBufferResult.release();
91: }
92:
93: this.makeStop();
94:
95: readSocketService.makeStop();
96:
97: haService.removeConnection(HAConnection.this);
98:
99: SelectionKey sk = this.socketChannel.keyFor(this.selector);
100: if (sk != null) {
101: sk.cancel();
102: }
103:
104: try {
105: this.selector.close();
106: this.socketChannel.close();
107: } catch (IOException e) {
108: HAConnection.log.error("", e);
109: }
110:
111: HAConnection.log.info(this.getServiceName() + " service end");
112: }
113:
114: /**
115: * 传输数据
116: */
117: private boolean transferData() throws Exception {
118: int writeSizeZeroTimes = 0;
119: // Write Header
120: while (this.byteBufferHeader.hasRemaining()) {
121: int writeSize = this.socketChannel.write(this.byteBufferHeader);
122: if (writeSize > 0) {
123: writeSizeZeroTimes = 0;
124: this.lastWriteTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now();
125: } else if (writeSize == 0) {
126: if (++writeSizeZeroTimes >= 3) {
127: break;
128: }
129: } else {
130: throw new Exception("ha master write header error < 0");
131: }
132: }
133:
134: if (null == this.selectMappedBufferResult) {
135: return !this.byteBufferHeader.hasRemaining();
136: }
137:
138: writeSizeZeroTimes = 0;
139:
140: // Write Body
141: if (!this.byteBufferHeader.hasRemaining()) {
142: while (this.selectMappedBufferResult.getByteBuffer().hasRemaining()) {
143: int writeSize = this.socketChannel.write(this.selectMappedBufferResult.getByteBuffer());
144: if (writeSize > 0) {
145: writeSizeZeroTimes = 0;
146: this.lastWriteTimestamp = HAConnection.this.haService.getDefaultMessageStore().getSystemClock().now();
147: } else if (writeSize == 0) {
148: if (++writeSizeZeroTimes >= 3) {
149: break;
150: }
151: } else {
152: throw new Exception("ha master write body error < 0");
153: }
154: }
155: }
156:
157: boolean result = !this.byteBufferHeader.hasRemaining() && !this.selectMappedBufferResult.getByteBuffer().hasRemaining();
158:
159: if (!this.selectMappedBufferResult.getByteBuffer().hasRemaining()) {
160: this.selectMappedBufferResult.release();
161: this.selectMappedBufferResult = null;
162: }
163:
164: return result;
165: }
3.1.6 Master_SYNC
Producer发送消息时，Master_SYNC节点 会等待Slave节点 存储完毕后再返回发送结果。
核心代码如下：
1: // ⬇️⬇️⬇️【CommitLog.java】
2: public PutMessageResult putMessage(final MessageExtBrokerInner msg) {
3: // ....省略处理发送代码
4: // Synchronous write double 如果是同步Master，同步到从节点
5: if (BrokerRole.SYNC_MASTER == this.defaultMessageStore.getMessageStoreConfig().getBrokerRole()) {
6: HAService service = this.defaultMessageStore.getHaService();
7: if (msg.isWaitStoreMsgOK()) {
8: // Determine whether to wait
9: if (service.isSlaveOK(result.getWroteOffset() + result.getWroteBytes())) {
10: if (null == request) {
11: request = new GroupCommitRequest(result.getWroteOffset() + result.getWroteBytes());
12: }
13: service.putRequest(request);
14:
15: // 唤醒WriteSocketService
16: service.getWaitNotifyObject().wakeupAll();
17:
18: boolean flushOK = request.waitForFlush(this.defaultMessageStore.getMessageStoreConfig().getSyncFlushTimeout());
19: if (!flushOK) {
20: log.error("do sync transfer other node, wait return, but failed, topic: " + msg.getTopic() + " tags: "
21: + msg.getTags() + " client address: " + msg.getBornHostString());
22: putMessageResult.setPutMessageStatus(PutMessageStatus.FLUSH_SLAVE_TIMEOUT);
23: }
24: }
25: // Slave problem
26: else {
27: // Tell the producer, slave not available
28: putMessageResult.setPutMessageStatus(PutMessageStatus.SLAVE_NOT_AVAILABLE);
29: }
30: }
31: }
32:
33: return putMessageResult;
34: }
- 第 16 行 ：唤醒
WriteSocketService。- 唤醒后，
WriteSocketService挂起等待新消息结束，Master传输Slave新的CommitLog数据。 Slave收到数据后，立即上报最新的CommitLog同步进度到Master。ReadSocketService唤醒第 18 行：request#waitForFlush(...)。
- 唤醒后，
我们来看下 GroupTransferService 的核心逻辑代码：
1: // ⬇️⬇️⬇️【GroupTransferService.java】
2: private void doWaitTransfer() {
3: synchronized (this.requestsRead) {
4: if (!this.requestsRead.isEmpty()) {
5: for (CommitLog.GroupCommitRequest req : this.requestsRead) {
6: // 等待Slave上传进度
7: boolean transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset();
8: for (int i = 0; !transferOK && i < 5; i++) {
9: this.notifyTransferObject.waitForRunning(1000); // 唤醒
10: transferOK = HAService.this.push2SlaveMaxOffset.get() >= req.getNextOffset();
11: }
12:
13: if (!transferOK) {
14: log.warn("transfer messsage to slave timeout, " + req.getNextOffset());
15: }
16:
17: // 唤醒请求，并设置是否Slave同步成功
18: req.wakeupCustomer(transferOK);
19: }
20:
21: this.requestsRead.clear();
22: }
23: }
24: }
3.2 Producer 发送消息
Producer发送消息时，会对Broker集群 的所有队列进行选择。
核心代码如下：
1: // ⬇️⬇️⬇️【DefaultMQProducerImpl.java】
2: private SendResult sendDefaultImpl(//
3: Message msg, //
4: final CommunicationMode communicationMode, //
5: final SendCallback sendCallback, //
6: final long timeout//
7: ) throws MQClientException, RemotingException, MQBrokerException, InterruptedException {
8: // .... 省略：处理【校验逻辑】
9: // 获取 Topic路由信息
10: TopicPublishInfo topicPublishInfo = this.tryToFindTopicPublishInfo(msg.getTopic());
11: if (topicPublishInfo != null && topicPublishInfo.ok()) {
12: MessageQueue mq = null; // 最后选择消息要发送到的队列
13: Exception exception = null;
14: SendResult sendResult = null; // 最后一次发送结果
15: int timesTotal = communicationMode == CommunicationMode.SYNC ? 1 + this.defaultMQProducer.getRetryTimesWhenSendFailed() : 1; // 同步多次调用
16: int times = 0; // 第几次发送
17: String[] brokersSent = new String[timesTotal]; // 存储每次发送消息选择的broker名
18: // 循环调用发送消息，直到成功
19: for (; times < timesTotal; times++) {
20: String lastBrokerName = null == mq ? null : mq.getBrokerName();
21: MessageQueue tmpmq = this.selectOneMessageQueue(topicPublishInfo, lastBrokerName); // 选择消息要发送到的队列
22: if (tmpmq != null) {
23: mq = tmpmq;
24: brokersSent[times] = mq.getBrokerName();
25: try {
26: beginTimestampPrev = System.currentTimeMillis();
27: // 调用发送消息核心方法
28: sendResult = this.sendKernelImpl(msg, mq, communicationMode, sendCallback, topicPublishInfo, timeout);
29: endTimestamp = System.currentTimeMillis();
30: // 更新Broker可用性信息
31: this.updateFaultItem(mq.getBrokerName(), endTimestamp - beginTimestampPrev, false);
32: // .... 省略：处理【发送返回结果】
33: }
34: } catch (e) { // .... 省略：处理【异常】
35:
36: }
37: } else {
38: break;
39: }
40: }
41: // .... 省略：处理【发送返回结果】
42: }
43: // .... 省略：处理【找不到消息路由】
44: }
如下是调试 #sendDefaultImpl(...) 时 TopicPublishInfo 的结果，Producer 获得到了 broker-a,broker-b 两个 Broker分组 的消息队列： 
3.3 Consumer 消费消息
Consumer消费消息时，会对Broker集群 的所有队列进行选择。
4. 总结
666. 彩蛋
