在【RocketMQ】消息的拉取一文中可知，消费者在启动的时候，会创建消息拉取API对象PullAPIWrapper，调用pullKernelImpl方法向Broker发送拉取消息的请求，那么在主从模式下消费者是如何选择向哪个Broker发送拉取请求的？
进入pullKernelImpl方法中，可以看到会调用recalculatePullFromWhichNode方法选择一个Broker：
public class PullAPIWrapper {
public PullResult pullKernelImpl(
final MessageQueue mq,
final String subExpression,
final String expressionType,
final long subVersion,
final long offset,
final int maxNums,
final int sysFlag,
final long commitOffset,
final long brokerSuspendMaxTimeMillis,
final long timeoutMillis,
final CommunicationMode communicationMode,
final PullCallback pullCallback
) throws MQClientException, RemotingException, MQBrokerException, InterruptedException {
// 调用recalculatePullFromWhichNode方法获取Broker ID，再调用findBrokerAddressInSubscribe根据ID获取Broker的相关信息
FindBrokerResult findBrokerResult =
this.mQClientFactory.findBrokerAddressInSubscribe(mq.getBrokerName(),
this.recalculatePullFromWhichNode(mq), false);
// ...
if (findBrokerResult != null) {
// ...
// 获取Broker地址
String brokerAddr = findBrokerResult.getBrokerAddr();
if (PullSysFlag.hasClassFilterFlag(sysFlagInner)) {
brokerAddr = computePullFromWhichFilterServer(mq.getTopic(), brokerAddr);
}
// 发送消息拉取请求
PullResult pullResult = this.mQClientFactory.getMQClientAPIImpl().pullMessage(
brokerAddr,
requestHeader,
timeoutMillis,
communicationMode,
pullCallback);
return pullResult;
}
}
}
在recalculatePullFromWhichNode方法中，会从pullFromWhichNodeTable中根据消息队列获取一个建议的Broker ID，如果获取为空就返回Master节点的Broker ID，ROCKETMQ中Master角色的Broker ID为0，既然从pullFromWhichNodeTable中可以知道从哪个Broker拉取数据，那么pullFromWhichNodeTable中的数据又是从哪里来的？
public class PullAPIWrapper {
// KEY为消息队列，VALUE为建议的Broker ID
private ConcurrentMap<MessageQueue, AtomicLong/* brokerId */> pullFromWhichNodeTable =
new ConcurrentHashMap<MessageQueue, AtomicLong>(32);
public long recalculatePullFromWhichNode(final MessageQueue mq) {
if (this.isConnectBrokerByUser()) {
return this.defaultBrokerId;
}
// 从pullFromWhichNodeTable中获取建议的broker ID
AtomicLong suggest = this.pullFromWhichNodeTable.get(mq);
if (suggest != null) {
return suggest.get();
}
// 返回Master Broker ID
return MixAll.MASTER_ID;
}
}
通过调用关系可知，在updatePullFromWhichNode方法中更新了pullFromWhichNodeTable的值，而updatePullFromWhichNode方法又是被processPullResult方法调用的，消费者向Broker发送拉取消息请求后，Broker对拉取请求进行处理时会设置一个broker ID（后面会讲到），建议下次从这个Broker拉取消息，消费者对拉取请求返回的响应数据进行处理时会调用processPullResult方法，在这里将建议的BrokerID取出，调用updatePullFromWhichNode方法将其加入到了pullFromWhichNodeTable中：
public class PullAPIWrapper {
private ConcurrentMap<MessageQueue, AtomicLong/* brokerId */> pullFromWhichNodeTable =
new ConcurrentHashMap<MessageQueue, AtomicLong>(32);
public PullResult processPullResult(final MessageQueue mq, final PullResult pullResult,
final SubscriptionData subscriptionData) {
PullResultExt pullResultExt = (PullResultExt) pullResult;
// 将拉取消息请求返回的建议Broker ID，加入到pullFromWhichNodeTable中
this.updatePullFromWhichNode(mq, pullResultExt.getSuggestWhichBrokerId());
// ...
}
public void updatePullFromWhichNode(final MessageQueue mq, final long brokerId) {
AtomicLong suggest = this.pullFromWhichNodeTable.get(mq);
if (null == suggest) {
// 向pullFromWhichNodeTable中添加数据
this.pullFromWhichNodeTable.put(mq, new AtomicLong(brokerId));
} else {
suggest.set(brokerId);
}
}
}
接下来去看下是根据什么条件决定选择哪个Broker的。
Broker在处理消费者拉取请求时，会调用PullMessageProcessor的processRequest方法，首先会调用MessageStore的getMessage方法获取消息内容，在返回的结果GetMessageResult中设置了一个是否建议从Slave节点拉取的属性(这个值的设置稍后再说)，会根据是否建议从slave节点进行以下处理：
subscriptionGroupConfig订阅分组配置的getWhichBrokerWhenConsumeSlowly方法获取从节点将ID设置到响应中，否则下次依旧建议从主节点拉取消息，将MASTER节点的ID设置到响应中；订阅分组配置
mqadmin命令的-i参数可以指定从哪个Broker消费消息（subscriptionGroupConfig的getBrokerId返回的值），-w参数可以指定建议从slave节点消费的时候，从哪个slave消费（subscriptionGroupConfig的getWhichBrokerWhenConsumeSlowly方法返回的值）：
usage: mqadmin updateSubGroup [-a <arg>] [-b <arg>] [-c <arg>] [-d <arg>] -g <arg> [-h] [-i <arg>] [-m <arg>]
[-n <arg>] [-q <arg>] [-r <arg>] [-s <arg>] [-w <arg>]
-i,--brokerId <arg> consumer from which broker id
-w,--whichBrokerWhenConsumeSlowly <arg> which broker id when consume slowly
public class PullMessageProcessor extends AsyncNettyRequestProcessor implements NettyRequestProcessor {
private RemotingCommand processRequest(final Channel channel, RemotingCommand request, boolean brokerAllowSuspend)
throws RemotingCommandException {
// ...
// 根据拉取偏移量获取消息
final GetMessageResult getMessageResult =
this.brokerController.getMessageStore().getMessage(requestHeader.getConsumerGroup(), requestHeader.getTopic(),
requestHeader.getQueueId(), requestHeader.getQueueOffset(), requestHeader.getMaxMsgNums(), messageFilter);
if (getMessageResult != null) {
response.setRemark(getMessageResult.getStatus().name());
responseHeader.setNextBeginOffset(getMessageResult.getNextBeginOffset());
responseHeader.setMinOffset(getMessageResult.getMinOffset());
responseHeader.setMaxOffset(getMessageResult.getMaxOffset());
// 是否建议从从节点拉取消息
if (getMessageResult.isSuggestPullingFromSlave()) {
// 选择一个从节点
responseHeader.setSuggestWhichBrokerId(subscriptionGroupConfig.getWhichBrokerWhenConsumeSlowly());
} else {
responseHeader.setSuggestWhichBrokerId(MixAll.MASTER_ID);
}
// 判断Broker的角色
switch (this.brokerController.getMessageStoreConfig().getBrokerRole()) {
case ASYNC_MASTER:
case SYNC_MASTER:
break;
case SLAVE:
// 如果不允许从从节点读取数据，设置为MasterID
if (!this.brokerController.getBrokerConfig().isSlaveReadEnable()) {
response.setCode(ResponseCode.PULL_RETRY_IMMEDIATELY);
responseHeader.setSuggestWhichBrokerId(MixAll.MASTER_ID);
}
break;
}
// 如果开启了允许从从节点读取数据
if (this.brokerController.getBrokerConfig().isSlaveReadEnable()) {
// 如果建议从从节点拉消息
if (getMessageResult.isSuggestPullingFromSlave()) {
// 获取从节点
responseHeader.setSuggestWhichBrokerId(subscriptionGroupConfig.getWhichBrokerWhenConsumeSlowly());
}
else {
// 获取指定的broker
responseHeader.setSuggestWhichBrokerId(subscriptionGroupConfig.getBrokerId());
}
} else {
// 使用Master节点
responseHeader.setSuggestWhichBrokerId(MixAll.MASTER_ID);
}
} else {
response.setCode(ResponseCode.SYSTEM_ERROR);
response.setRemark("store getMessage return null");
}
}
}
DefaultMessageStore的getMessage方法中用于获取消息内容，并会根据消费者的拉取进度判断是否建议下次从Slave节点拉取消息，判断过程如下：
public class DefaultMessageStore implements MessageStore {
public GetMessageResult getMessage(final String group, final String topic, final int queueId, final long offset,
final int maxMsgNums,
final MessageFilter messageFilter) {
// ...
// 当前CommitLog的最大偏移量
final long maxOffsetPy = this.commitLog.getMaxOffset();
ConsumeQueue consumeQueue = findConsumeQueue(topic, queueId);
if (consumeQueue != null) {
minOffset = consumeQueue.getMinOffsetInQueue();
maxOffset = consumeQueue.getMaxOffsetInQueue();
if (maxOffset == 0) {
// ...
} else {
// 根据消费进度获取消息队列
SelectMappedBufferResult bufferConsumeQueue = consumeQueue.getIndexBuffer(offset);
if (bufferConsumeQueue != null) {
try {
// ...
// CommitLog最大偏移量减去本次拉取消息的最大物理偏移量
long diff = maxOffsetPy - maxPhyOffsetPulling;
// 计算消息在PageCache中的总大小（总物理内存 * 消息存储在内存中的阀值/100）
long memory = (long) (StoreUtil.TOTAL_PHYSICAL_MEMORY_SIZE
* (this.messageStoreConfig.getAccessMessageInMemoryMaxRatio() / 100.0));
// 是否建议下次去从节点拉取消息
getResult.setSuggestPullingFromSlave(diff > memory);
} finally {
bufferConsumeQueue.release();
}
} else {
// ...
}
}
} else {
status = GetMessageStatus.NO_MATCHED_LOGIC_QUEUE;
nextBeginOffset = nextOffsetCorrection(offset, 0);
}
// ...
return getResult;
}
}
总结
消费者在启动后需要向Broker发送拉取消息的请求，Broker收到请求后会根据消息的拉取进度，返回一个建议的BrokerID，并设置到响应中返回，消费者处理响应时将建议的BrokerID放入pullFromWhichNodeTable，下次拉去消息的时候从pullFromWhichNodeTable中取出，并向其发送请求拉取消息。
上面讲解了主从模式下如何选择从哪个Broker拉取消息，接下来看下消费进度的持久化，因为广播模式下消费进度保存在每个消费者端，集群模式下消费进度保存在Broker端，所以接下来以集群模式为例。
在【RocketMQ】消息的拉取一文中可知，集群模式下主要是通过RemoteBrokerOffsetStore进行消费进度管理的，在持久化方法persistAll中会调用updateConsumeOffsetToBroker更新Broker端的消费进度：
public class RemoteBrokerOffsetStore implements OffsetStore {
@Override
public void persistAll(Set<MessageQueue> mqs) {
if (null == mqs || mqs.isEmpty())
return;
final HashSet<MessageQueue> unusedMQ = new HashSet<MessageQueue>();
for (Map.Entry<MessageQueue, AtomicLong> entry : this.offsetTable.entrySet()) {
MessageQueue mq = entry.getKey();
AtomicLong offset = entry.getValue();
if (offset != null) {
if (mqs.contains(mq)) {
try {
// 向Broker发送请求更新消费进度
this.updateConsumeOffsetToBroker(mq, offset.get());
log.info("[persistAll] Group: {} ClientId: {} updateConsumeOffsetToBroker {} {}",
this.groupName,
this.mQClientFactory.getClientId(),
mq,
offset.get());
} catch (Exception e) {
log.error("updateConsumeOffsetToBroker exception, " + mq.toString(), e);
}
} else {
unusedMQ.add(mq);
}
}
}
// ...
}
}
由于updateConsumeOffsetToBroker方法中先调用了findBrokerAddressInSubscribe方法获取Broker的信息，所以这里先看findBrokerAddressInSubscribe方法是如何选择Broker的，它需要传入三个参数，分别为：Broker名称、Broker ID、是否只查找参数中传入的那个BrokerID，方法的处理逻辑如下：
brokerAddrTable中根据Broker的名称获取所有的Broker集合（主从模式下他们的Broker名称一致，但是ID不一致），KEY为BrokerID，VALUE为Broker的地址；public class MQClientInstance {
public FindBrokerResult findBrokerAddressInSubscribe(
final String brokerName, // Broker名称
final long brokerId, // Broker ID
final boolean onlyThisBroker // 是否只查找参数中传入的那个BrokerID
) {
String brokerAddr = null;
boolean slave = false;
boolean found = false;
// 获取所有的Broker ID
HashMap<Long/* brokerId */, String/* address */> map = this.brokerAddrTable.get(brokerName);
if (map != null && !map.isEmpty()) {
brokerAddr = map.get(brokerId);
// 是否是从节点
slave = brokerId != MixAll.MASTER_ID;
// 地址是否为空
found = brokerAddr != null;
// 如果地址为空并且是从节点
if (!found && slave) {
// 获取下一个Broker
brokerAddr = map.get(brokerId + 1);
found = brokerAddr != null;
}
// 如果地址为空
if (!found && !onlyThisBroker) {
// 获取集合中的第一个节点
Entry<Long, String> entry = map.entrySet().iterator().next();
// 获取地址
brokerAddr = entry.getValue();
// 是否是从节点
slave = entry.getKey() != MixAll.MASTER_ID;
// 置为true
found = true;
}
}
if (found) {
// 返回数据
return new FindBrokerResult(brokerAddr, slave, findBrokerVersion(brokerName, brokerAddr));
}
return null;
}
}
回到updateConsumeOffsetToBroker方法，先看第一次调用findBrokerAddressInSubscribe方法获取Broker信息，传入的三个参数分别为：Broker名称、Master节点的ID、true，根据上面讲解的findBrokerAddressInSubscribe方法里面的查找逻辑，如果查找到Master节点的信息，就正常返回，如果此时Master宕机未能正常查找到，由于传入的Master节点的ID并且onlyThisBroker置为true，所以会查找失败返回NULL。
如果第一次调用为空，会进行第二次调用，与第一次调用不同的地方是第三个参数置为了false，也就是说不是必须选择参数中指定的那个Broker,此时依旧优先查找Master节点，如果Master节点未查找到，由于onlyThisBroker置为了false，会迭代集合选择第一个节点返回，此时返回的有可能是从节点。
总结：消费者会优先选择向主节点发送请求进行消费进度保存，假如主节点宕机等原因未能获取到主节点的信息，会迭代集合选择第一个节点返回，所以消费者也可以向从节点发送请求进行进度保存，待主节点恢复后，依旧优先选择主节点。
public class RemoteBrokerOffsetStore implements OffsetStore {
private void updateConsumeOffsetToBroker(MessageQueue mq, long offset) throws RemotingException,
MQBrokerException, InterruptedException, MQClientException {
// 更新消费进度
updateConsumeOffsetToBroker(mq, offset, true);
}
@Override
public void updateConsumeOffsetToBroker(MessageQueue mq, long offset, boolean isOneway) throws RemotingException,
MQBrokerException, InterruptedException, MQClientException {
// 第一次调用findBrokerAddressInSubscribe方法获取Broker信息，三个参数分别为：Broker名称、Master节点的ID、true
FindBrokerResult findBrokerResult = this.mQClientFactory.findBrokerAddressInSubscribe(mq.getBrokerName(), MixAll.MASTER_ID, true);
// 如果获取为空，进行第二次调用
if (null == findBrokerResult) {
// 三个参数分别为：Broker名称、Master节点的ID、false
this.mQClientFactory.updateTopicRouteInfoFromNameServer(mq.getTopic());
findBrokerResult = this.mQClientFactory.findBrokerAddressInSubscribe(mq.getBrokerName(), MixAll.MASTER_ID, false);
}
if (findBrokerResult != null) {
// 设置请求头
UpdateConsumerOffsetRequestHeader requestHeader = new UpdateConsumerOffsetRequestHeader();
requestHeader.setTopic(mq.getTopic());
requestHeader.setConsumerGroup(this.groupName);
requestHeader.setQueueId(mq.getQueueId());
requestHeader.setCommitOffset(offset);
// 发送保存消费进度的请求
if (isOneway) {
this.mQClientFactory.getMQClientAPIImpl().updateConsumerOffsetOneway(
findBrokerResult.getBrokerAddr(), requestHeader, 1000 * 5);
} else {
this.mQClientFactory.getMQClientAPIImpl().updateConsumerOffset(
findBrokerResult.getBrokerAddr(), requestHeader, 1000 * 5);
}
} else {
throw new MQClientException("The broker[" + mq.getBrokerName() + "] not exist", null);
}
}
}
BrokerController在构造函数中，实例化了SlaveSynchronize，并在start方法中调用了handleSlaveSynchronize方法处理从节点的数据同步，
如果当前的Broker是从节点，会注册定时任务，定时调用SlaveSynchronize的syncAll方法进行数据同步：
public class BrokerController {
private final SlaveSynchronize slaveSynchronize;
public BrokerController(
final BrokerConfig brokerConfig,
final NettyServerConfig nettyServerConfig,
final NettyClientConfig nettyClientConfig,
final MessageStoreConfig messageStoreConfig
) {
// ...
this.slaveSynchronize = new SlaveSynchronize(this);
//...
}
public void start() throws Exception {
if (!messageStoreConfig.isEnableDLegerCommitLog()) {
startProcessorByHa(messageStoreConfig.getBrokerRole());
// 处理从节点的同步
handleSlaveSynchronize(messageStoreConfig.getBrokerRole());
this.registerBrokerAll(true, false, true);
}
}
private void handleSlaveSynchronize(BrokerRole role) {
// 如果是SLAVE节点
if (role == BrokerRole.SLAVE) {
if (null != slaveSyncFuture) {
slaveSyncFuture.cancel(false);
}
this.slaveSynchronize.setMasterAddr(null);
// 设置定时任务，定时进行数据同步
slaveSyncFuture = this.scheduledExecutorService.scheduleAtFixedRate(new Runnable() {
@Override
public void run() {
try {
// 同步数据
BrokerController.this.slaveSynchronize.syncAll();
}
catch (Throwable e) {
log.error("ScheduledTask SlaveSynchronize syncAll error.", e);
}
}
}, 1000 * 3, 1000 * 10, TimeUnit.MILLISECONDS);
} else {
//handle the slave synchronise
if (null != slaveSyncFuture) {
slaveSyncFuture.cancel(false);
}
this.slaveSynchronize.setMasterAddr(null);
}
}
}
在SlaveSynchronize的syncAll方法中，又调用了syncConsumerOffset方法同步消费进度：
public class SlaveSynchronize {
public void syncAll() {
this.syncTopicConfig();
// 同步消费进度
this.syncConsumerOffset();
this.syncDelayOffset();
this.syncSubscriptionGroupConfig();
}
private void syncConsumerOffset() {
String masterAddrBak = this.masterAddr;
if (masterAddrBak != null && !masterAddrBak.equals(brokerController.getBrokerAddr())) {
try {
// 向主节点发送请求获取消费进度信息
ConsumerOffsetSerializeWrapper offsetWrapper =
this.brokerController.getBrokerOuterAPI().getAllConsumerOffset(masterAddrBak);
// 设置数据
this.brokerController.getConsumerOffsetManager().getOffsetTable()
.putAll(offsetWrapper.getOffsetTable());
// 将获取到的消费进度数据进行持久化
this.brokerController.getConsumerOffsetManager().persist();
log.info("Update slave consumer offset from master, {}", masterAddrBak);
} catch (Exception e) {
log.error("SyncConsumerOffset Exception, {}", masterAddrBak, e);
}
}
}
}
参考
丁威、周继锋《RocketMQ技术内幕》
RocketMQ 主从同步若干问题答疑
RocketMq 订阅分组创建和删除
RocketMQ版本：4.9.3