SpringData-Redis发布订阅自动重连分析
RedisMessageListenerContainer
配置
@Bean
@Autowired
RedisMessageListenerContainer redisContainer(JedisConnectionFactory redisConnectionFactory, RedisMessageListener a) {
RedisMessageListenerContainer container
= new RedisMessageListenerContainer();
container.setConnectionFactory(redisConnectionFactory);
List<Topic> topics = Lists.newArrayList(new ChannelTopic(
CHANNEL),
new ChannelTopic(CHANNEL)
);
container.addMessageListener(new MessageListenerAdapter(a), topics);
return container;
}
启动分析
添加频道监听
//RedisMessageListenerContainer.java
public void addMessageListener(MessageListener listener, Collection<? extends Topic> topics) {
addListener(listener, topics);
lazyListen();
}
这个AddListener会 对Topic做一些记录,patternMapping, channelMapping,去重等等,然后最关键的一步:
//RedisMessageListenerContainer.java
//addListener
// check the current listening state
if (listening) {
subscriptionTask.subscribeChannel(channels.toArray(new byte[channels.size()][]));
subscriptionTask.subscribePattern(patterns.toArray(new byte[patterns.size()][]));
}
//RedisMessageListenerContainer.java
void subscribeChannel(byte[]... channels) {
if (channels != null && channels.length > 0) {
if (connection != null) {
synchronized (localMonitor) {
Subscription sub = connection.getSubscription();
if (sub != null) {
sub.subscribe(channels);
}
}
}
}
}
//JedisSubscription.java
protected void doSubscribe(byte[]... channels) {
jedisPubSub.subscribe(channels);
}
但是启动之前 这个listening=false。故该代码不生效。再看lazyListen方法:
//RedisMessageListenerContainer.java
private void lazyListen() {
boolean debug = logger.isDebugEnabled();
boolean started = false;
if (isRunning()) {
if (!listening) {
synchronized (monitor) {
if (!listening) {
if (channelMapping.size() > 0 || patternMapping.size() > 0) {
subscriptionExecutor.execute(subscriptionTask);
listening = true;
started = true;
}
}
}
if (debug) {
if (started) {
logger.debug("Started listening for Redis messages");
} else {
logger.debug("Postpone listening for Redis messages until actual listeners are added");
}
}
}
}
}
调用addMessageListener的时候,isRunning()=false 也不生效。
最后:当@Bean构造完成的时候 ,LifeCycle进入start的时候,该Container启动。
//RedisMessageListenerContainer.java
public void start() {
if (!running) {
running = true;
// wait for the subscription to start before returning
// technically speaking we can only be notified right before the subscription starts
synchronized (monitor) {
lazyListen();
if (listening) {
try {
// wait up to 5 seconds for Subscription thread
monitor.wait(initWait);
} catch (InterruptedException e) {
// stop waiting
}
}
}
if (logger.isDebugEnabled()) {
logger.debug("Started RedisMessageListenerContainer");
}
}
}
这个时候,running=true了。
然后调用 lazyListen(),确实比较Lazy。
这个时候,启动子线程来执行订阅和监听。
subscriptionExecutor.execute(subscriptionTask);
这个subscriptionTask的构造如下:
//RedisMessageListenerContainer.java
public void run() {
synchronized (localMonitor) {
subscriptionTaskRunning = true;
}
try {
connection = connectionFactory.getConnection();
if (connection.isSubscribed()) {
throw new IllegalStateException("Retrieved connection is already subscribed; aborting listening");
}
boolean asyncConnection = ConnectionUtils.isAsync(connectionFactory);
// NB: async drivers' Xsubscribe calls block, so we notify the RDMLC before performing the actual subscription.
if (!asyncConnection) {
synchronized (monitor) {
monitor.notify();
}
}
SubscriptionPresentCondition subscriptionPresent = eventuallyPerformSubscription();
if (asyncConnection) {
SpinBarrier.waitFor(subscriptionPresent, getMaxSubscriptionRegistrationWaitingTime());
synchronized (monitor) {
monitor.notify();
}
}
} catch (Throwable t) {
handleSubscriptionException(t);
} finally {
// this block is executed once the subscription thread has ended, this may or may not mean
// the connection has been unsubscribed, depending on driver
synchronized (localMonitor) {
subscriptionTaskRunning = false;
localMonitor.notify();
}
}
}
这里connection 肯定不是subscribed。
然后他根据Redis的客户端类型来判断是否是阻塞的
如果是阻塞的类型,则唤醒一下被阻塞的Container线程。(???)
然后,最关键的是:eventuallyPerformSubscription(),最终发起订阅的,并轮询订阅的是方法。
//RDMLC
private SubscriptionPresentCondition eventuallyPerformSubscription() {
SubscriptionPresentCondition condition = null;
if (channelMapping.isEmpty()) {
condition = new PatternSubscriptionPresentCondition();
connection.pSubscribe(new DispatchMessageListener(), unwrap(patternMapping.keySet()));
} else {
if (patternMapping.isEmpty()) {
condition = new SubscriptionPresentCondition();
} else {
// schedule the rest of the subscription
subscriptionExecutor.execute(new PatternSubscriptionTask());
condition = new PatternSubscriptionPresentCondition();
}
connection.subscribe(new DispatchMessageListener(), unwrap(channelMapping.keySet()));
}
return condition;
}
以connection.subscribe()为例:即将发起订阅,注意这里是利用DispatchMessageListener做的事件分发监听器。
//JedisConnection.java
public void subscribe(MessageListener listener, byte[]... channels) {
if (isSubscribed()) {
throw new RedisSubscribedConnectionException(
"Connection already subscribed; use the connection Subscription to cancel or add new channels");
}
if (isQueueing()) {
throw new UnsupportedOperationException();
}
if (isPipelined()) {
throw new UnsupportedOperationException();
}
try {
BinaryJedisPubSub jedisPubSub = new JedisMessageListener(listener);
subscription = new JedisSubscription(listener, jedisPubSub, channels, null);
jedis.subscribe(jedisPubSub, channels);
} catch (Exception ex) {
throw convertJedisAccessException(ex);
}
}
//BinaryJedis.java
public void subscribe(BinaryJedisPubSub jedisPubSub, byte[]... channels) {
client.setTimeoutInfinite();
try {
jedisPubSub.proceed(client, channels);
} finally {
client.rollbackTimeout();
}
}
这里调用了BinaryJedisPubSub的proceed()。
这里先提出两个问题?
要订阅是不是要发起subscribe命令给Redis?发起 subscribe channel命令,然后Listener怎么办?
这里调用是jedis.subscribe(jedisPubSub, channels);而一开始 subscibeChannels的实现却不太一样?
下面看jedisPubSub:
public void proceed(Client client, byte[]... channels) {
this.client = client;
client.subscribe(channels);
client.flush();
process(client);
}
这里subscribe是再次发起订阅请求,然后process轮询检查消息。
异常处理
再看看JedisConnection类subscribe方法的异常的处理:
protected DataAccessException convertJedisAccessException(Exception ex) {
if (ex instanceof NullPointerException) {
// An NPE before flush will leave data in the OutputStream of a pooled connection
broken = true;
}
DataAccessException exception = EXCEPTION_TRANSLATION.translate(ex);
if (exception instanceof RedisConnectionFailureException) {
broken = true;
}
return exception;
}
EXCEPTION_TRANSLATION.translate(ex); 会调用:PassThroughExceptionTranslationStrategy的Convert。
public class JedisExceptionConverter implements Converter<Exception, DataAccessException> {
public DataAccessException convert(Exception ex) {
if (ex instanceof DataAccessException) {
return (DataAccessException) ex;
}
if (ex instanceof JedisDataException) {
return new InvalidDataAccessApiUsageException(ex.getMessage(), ex);
}
if (ex instanceof JedisConnectionException) {
return new RedisConnectionFailureException(ex.getMessage(), ex);
}
if (ex instanceof JedisException) {
return new InvalidDataAccessApiUsageException(ex.getMessage(), ex);
}
if (ex instanceof UnknownHostException) {
return new RedisConnectionFailureException("Unknown host " + ex.getMessage(), ex);
}
if (ex instanceof IOException) {
return new RedisConnectionFailureException("Could not connect to Redis server", ex);
}
return null;
}
}
那么,当Jedis抛错:JedisConnectionException 服务器似乎断开了连接
这个时候,subscribe 从而抛出RedisConnectionFailureException。
最后,再看RedisMessageListenerContainerd的run方法内的异常处理:
这个时候,
protected void handleSubscriptionException(Throwable ex) {
listening = false;
subscriptionTask.closeConnection();
if (ex instanceof RedisConnectionFailureException) {
if (isRunning()) {
logger.error("Connection failure occurred. Restarting subscription task after " + recoveryInterval + " ms");
sleepBeforeRecoveryAttempt();
lazyListen();
}
} else {
logger.error("SubscriptionTask aborted with exception:", ex);
}
}
到这个时候,isRunning还是true的(当且仅当LifeCycle进入close的时候,才会变成false),结果就会在 recoveryInterval ms之后,重启调用lazyListen(),再次启动订阅和监听。
实际效果
实际上,我在服务器上的错误日志中,我确实看到了
Connection failure occurred. Restarting subscription task after 5000 ms
总结
SpringData-Redis,能够解决手动处理Redis pub/sub的订阅被意外断开,导致监听失败的问题。
他能确保,服务持续监听,出现异常时,能够重新订阅并监听给定的频道。
所以,还是用框架吧,比自己手写的发布订阅更可靠。