Procuder相对consumer来说比较简单，根据topic从zk拿broker列表，注意这里只拿master类型的broker，slave型的broker和master拥有同样的broker id，主要为了HA用。roubd-robin取一个partition，发送消息。
1.MetaMessageSessionFactory初始化zookeeper连接，创建MessageProducer，默认SimpleMessageProducer
2.订阅topic，是一个异步操作
public void publishTopic(final String topic, final Object ref) {
if (this.topicConnectionListeners.get(topic) != null) {
this.addRef(topic, ref);
return;
}
final FutureTask<BrokerConnectionListener> task =
new FutureTask<BrokerConnectionListener>(new Callable<BrokerConnectionListener>() {
@Override
public BrokerConnectionListener call() throws Exception {
final BrokerConnectionListener listener = new BrokerConnectionListener(topic);
if (ProducerZooKeeper.this.zkClient != null) {
ProducerZooKeeper.this.publishTopicInternal(topic, listener);
}
listener.references.add(ref);
return listener;
}
});
final FutureTask<BrokerConnectionListener> existsTask = this.topicConnectionListeners.putIfAbsent(topic, task);
if (existsTask == null) {
task.run();
}
else {
this.addRef(topic, ref);
}
}
订阅过程
private void publishTopicInternal(final String topic, final BrokerConnectionListener listener) throws Exception,
NotifyRemotingException, InterruptedException {
///meta/brokers/topics-pub/的topic节点，
final String partitionPath = this.metaZookeeper.brokerTopicsPubPath + "/" + topic;
ZkUtils.makeSurePersistentPathExists(ProducerZooKeeper.this.zkClient, partitionPath);
//BrokerConnectionListener处理broker变化的情况
ProducerZooKeeper.this.zkClient.subscribeChildChanges(partitionPath, listener);
// 第一次要同步等待就绪
listener.syncedUpdateBrokersInfo();
}
同步broker信息过程
void syncedUpdateBrokersInfo() throws NotifyRemotingException, InterruptedException {
this.lock.lock();
try {
//取该topic下所有master节点，节点名称有后缀'-m'，key是brokerId，value是broker地址，从节点目录/meta/brokers/ids/0/master下读取
final Map<Integer, String> newBrokerStringMap =
ProducerZooKeeper.this.metaZookeeper.getMasterBrokersByTopic(this.topic);
final List<String> topics = new ArrayList<String>(1);
topics.add(this.topic);
//返回master的topic到partition映射的map，key是topic，value是对应的master中配置的partition集合
final Map<String, List<Partition>> newTopicPartitionMap =
ProducerZooKeeper.this.metaZookeeper.getPartitionsForTopicsFromMaster(topics);
log.warn("Begin receiving broker changes for topic " + this.topic + ",broker ids:"
+ newTopicPartitionMap);
// Connect to new brokers
//新broker创建连接
for (final Map.Entry<Integer, String> newEntry : newBrokerStringMap.entrySet()) {
final Integer newBrokerId = newEntry.getKey();
final String newBrokerString = newEntry.getValue();
// 新的有，旧的没有，创建
if (!this.brokersInfo.oldBrokerStringMap.containsKey(newBrokerId)) {
ProducerZooKeeper.this.remotingClient.connect(newBrokerString, this);
ProducerZooKeeper.this.remotingClient.awaitReadyInterrupt(newBrokerString);
log.warn("Connect to " + newBrokerString);
}
}
// Close removed brokers.
//没用的broker删除连接
for (final Map.Entry<Integer, String> oldEntry : this.brokersInfo.oldBrokerStringMap.entrySet()) {
final Integer oldBrokerId = oldEntry.getKey();
final String oldBrokerString = oldEntry.getValue();
final String newBrokerString = newBrokerStringMap.get(oldBrokerId);
// 新旧都有
if (newBrokerStringMap.containsKey(oldBrokerId)) {
// 判断内容是否变化
if (!newBrokerString.equals(oldBrokerString)) {
log.warn("Close " + oldBrokerString + ",connect to " + newBrokerString);
ProducerZooKeeper.this.remotingClient.connect(newBrokerString, this);
ProducerZooKeeper.this.remotingClient.awaitReadyInterrupt(newBrokerString);
ProducerZooKeeper.this.remotingClient.close(oldBrokerString, this, false);
}
else {
// ignore
}
}
else {
// 新的没有，旧的有，关闭
ProducerZooKeeper.this.remotingClient.close(oldBrokerString, this, false);
log.warn("Close " + oldBrokerString);
}
}
// Set the new brokers info.
this.brokersInfo = new BrokersInfo(newBrokerStringMap, newTopicPartitionMap);
log.warn("End receiving broker changes for topic " + this.topic);
}
finally {
this.lock.unlock();
}
}
}
3. 发送消息
private SendResult send0(final Message message, final byte[] encodedData, final long timeout, final TimeUnit unit)
throws InterruptedException, MetaClientException {
try {
final String topic = message.getTopic();
Partition partition = null;
String serverUrl = null;
// 如果在事务内，则使用上一次发送消息时选择的broker
if (this.isInTransaction()) {
final LastSentInfo info = this.lastSentInfo.get();
if (info != null) {
serverUrl = info.serverUrl;
// 选择该broker内的某个分区
partition =
this.producerZooKeeper.selectPartition(topic, message, this.partitionSelector, serverUrl);
if (partition == null) {
// 没有可用分区，抛出异常
throw new MetaClientException("There is no partitions in `" + serverUrl
+ "` to send message with topic `" + topic + "` in a transaction");
}
}
}
//默认采用round-robin策略发送，partition里取一个
if (partition == null) {
partition = this.selectPartition(message);
}
if (partition == null) {
throw new MetaClientException("There is no aviable partition for topic " + topic
+ ",maybe you don't publish it at first?");
}
if (serverUrl == null) {
serverUrl = this.producerZooKeeper.selectBroker(topic, partition);
}
if (serverUrl == null) {
throw new MetaClientException("There is no aviable server right now for topic " + topic
+ " and partition " + partition + ",maybe you don't publish it at first?");
}
if (this.isInTransaction() && this.lastSentInfo.get() == null) {
// 第一次发送，需要启动事务
this.beforeSendMessageFirstTime(serverUrl);
}
final int flag = MessageFlagUtils.getFlag(message);
//Command拼装
final PutCommand putCommand =
new PutCommand(topic, partition.getPartition(), encodedData, flag, CheckSum.crc32(encodedData),
this.getTransactionId(), OpaqueGenerator.getNextOpaque());
//发送
final BooleanCommand resp = this.invokeToGroup(serverUrl, partition, putCommand, message, timeout, unit);
//返回结果
return this.genSendResult(message, partition, serverUrl, resp);
}
catch (final TimeoutException e) {
throw new MetaOpeartionTimeoutException("Send message timeout in "
+ TimeUnit.MILLISECONDS.convert(timeout, unit) + " mills");
}
catch (final InterruptedException e) {
throw e;
}
catch (final MetaClientException e) {
throw e;
}
catch (final Exception e) {
throw new MetaClientException("send message failed", e);
}
}