dubbo使用的zk客户端

在使用dubbo的过程中，当注册中心的数据修改后，新的配置是怎样刷到consumer和provider的？本文以consumer为例，进行分析。

dubbo使用的是zkclient的jar，而zkclient依赖zookeeper的jar。

示例代码：

public class MyI0zk {
    public static void main(String[] args) {
        ZkClient zkclient = new ZkClient("127.0.0.1:2181");
        int countChildren = zkclient.countChildren("/");
        System.out.println(countChildren);
    }
}

调用栈如下图：

在connect方法中创建了ZooKeeper对象：

// void org.I0Itec.zkclient.ZkConnection.connect(Watcher watcher)
public void connect(Watcher watcher) {
    _zookeeperLock.lock();
    try {
        if (_zk != null) {
            throw new IllegalStateException("zk client has already been started");
        }
        try {
            LOG.debug("Creating new ZookKeeper instance to connect to " + _servers + ".");
            _zk = new ZooKeeper(_servers, _sessionTimeOut, watcher);
        } catch (IOException e) {
            throw new ZkException("Unable to connect to " + _servers, e);
        }
    } finally {
        _zookeeperLock.unlock();
    }
}

ZkClient自身就是一个Watcher：

//省略其他代码
public class ZkClient implements Watcher {
    public void process(WatchedEvent event) {
        LOG.debug("Received event: " + event);
        _zookeeperEventThread = Thread.currentThread();
 
        boolean stateChanged = event.getPath() == null;
        boolean znodeChanged = event.getPath() != null;
        boolean dataChanged = event.getType() == EventType.NodeDataChanged || event.getType() == EventType.NodeDeleted || event.getType() == EventType.NodeCreated
                || event.getType() == EventType.NodeChildrenChanged;
 
        getEventLock().lock();
        try {
 
            // We might have to install child change event listener if a new node was created
            if (getShutdownTrigger()) {
                LOG.debug("ignoring event '{" + event.getType() + " | " + event.getPath() + "}' since shutdown triggered");
                return;
            }
            if (stateChanged) {
                processStateChanged(event);
            }
            if (dataChanged) {
                processDataOrChildChange(event);
            }
        } finally {
            if (stateChanged) {
                getEventLock().getStateChangedCondition().signalAll();
 
                // If the session expired we have to signal all conditions, because watches might have been removed and
                // there is no guarantee that those
                // conditions will be signaled at all after an Expired event
                // TODO PVo write a test for this
                if (event.getState() == KeeperState.Expired) {
                    getEventLock().getZNodeEventCondition().signalAll();
                    getEventLock().getDataChangedCondition().signalAll();
                    // We also have to notify all listeners that something might have changed
                    fireAllEvents();
                }
            }
            if (znodeChanged) {
                getEventLock().getZNodeEventCondition().signalAll();
            }
            if (dataChanged) {
                getEventLock().getDataChangedCondition().signalAll();
            }
            getEventLock().unlock();
            LOG.debug("Leaving process event");
        }
    }
 
}

1. 再看看dubbo的consumer创建ZkClient的调用栈：

2. consumer的ZkClient注册childListener，等待节点内容改变的通知，对[/dubbo/com.zhang.HelloService/providers, /dubbo/com.zhang.HelloService/configurators, /dubbo/com.zhang.HelloService/routers]分别进行注册：

实际调用exists和getChildren方法，对应zk的stat path watch和ls path watch命令，即当/dubbo/com.zhang.HelloService/providers的值或者子节点改变时，客户端会收到通知：

// ZkClient
public List<String> watchForChilds(final String path) {
    if (_zookeeperEventThread != null && Thread.currentThread() == _zookeeperEventThread) {
        throw new IllegalArgumentException("Must not be done in the zookeeper event thread.");
    }
    return retryUntilConnected(new Callable<List<String>>() {
        @Override
        public List<String> call() throws Exception {
            exists(path, true);
            try {
                return getChildren(path, true);
            } catch (ZkNoNodeException e) {
                // ignore, the "exists" watch will listen for the parent node to appear
            }
            return null;
        }
    });
}
 
public <T> T retryUntilConnected(Callable<T> callable) throws ZkInterruptedException, IllegalArgumentException, ZkException, RuntimeException {
    if (_zookeeperEventThread != null && Thread.currentThread() == _zookeeperEventThread) {
        throw new IllegalArgumentException("Must not be done in the zookeeper event thread.");
    }
    while (true) {
        try {
            return callable.call();
        } catch (ConnectionLossException e) {
            // we give the event thread some time to update the status to 'Disconnected'
            Thread.yield();
            waitUntilConnected();
        } catch (SessionExpiredException e) {
            // we give the event thread some time to update the status to 'Expired'
            Thread.yield();
            waitUntilConnected();
        } catch (KeeperException e) {
            throw ZkException.create(e);
        } catch (InterruptedException e) {
            throw new ZkInterruptedException(e);
        } catch (Exception e) {
            throw ExceptionUtil.convertToRuntimeException(e);
        }
    }
}

3. consumer收到通知，重新加watch

consumer的ZkClient收到通知后，会将通知包装成ZkEvent事件，发送给ZkEventThread，ZkEvent的run方法里重新地加上watch：

private void fireChildChangedEvents(final String path, Set<IZkChildListener> childListeners) {
    try {
        // reinstall the watch
        for (final IZkChildListener listener : childListeners) {
            _eventThread.send(new ZkEvent("Children of " + path + " changed sent to " + listener) {
 
                @Override
                public void run() throws Exception {
                    try {
                        // if the node doesn't exist we should listen for the root node to reappear
                        // 在函数内部，根据是否有listeners，判断加watch
                        exists(path);
                        List<String> children = getChildren(path);
                        listener.handleChildChange(path, children);
                    } catch (ZkNoNodeException e) {
                        listener.handleChildChange(path, null);
                    }
                }
            });
        }
    } catch (Exception e) {
        LOG.error("Failed to fire child changed event. Unable to getChildren.  ", e);
    }
}

4. 从ZkEvent的run方法进入，刷新配置：

ZkEventThread线程：

class ZkEventThread extends Thread {
    private static final Logger LOG = Logger.getLogger(ZkEventThread.class);
    private BlockingQueue<ZkEvent> _events = new LinkedBlockingQueue<ZkEvent>();
    private static AtomicInteger _eventId = new AtomicInteger(0);
 
    static abstract class ZkEvent {
        private String _description;
        public ZkEvent(String description) {
            _description = description;
        }
        public abstract void run() throws Exception;
        @Override
        public String toString() {
            return "ZkEvent[" + _description + "]";
        }
    }
 
    ZkEventThread(String name) {
        setDaemon(true);
        setName("ZkClient-EventThread-" + getId() + "-" + name);
    }
 
    @Override
    public void run() {
        LOG.info("Starting ZkClient event thread.");
        try {
            while (!isInterrupted()) {
                ZkEvent zkEvent = _events.take();
                int eventId = _eventId.incrementAndGet();
                LOG.debug("Delivering event #" + eventId + " " + zkEvent);
                try {
                    zkEvent.run();
                } catch (InterruptedException e) {
                    interrupt();
                } catch (ZkInterruptedException e) {
                    interrupt();
                } catch (Throwable e) {
                    LOG.error("Error handling event " + zkEvent, e);
                }
                LOG.debug("Delivering event #" + eventId + " done");
            }
        } catch (InterruptedException e) {
            LOG.info("Terminate ZkClient event thread.");
        }
    }
 
    public void send(ZkEvent event) {
        if (!isInterrupted()) {
            LOG.debug("New event: " + event);
            _events.add(event);
        }
    }
}