Zookeeper源码分析:Watcher机制

1. 设置Watcher
使用Watcher需要先实现Watcher接口,并将实现类对象传递到指定方法中,如getChildren, exist等。Zookeeper允许在构造Zookeeper对象时候指定一个默认Watcher对象.getChildren和exit方法可以使用这个默认的Watcher对象,也可以指定一个新Watcher对象。

Code 1: Watcher接口

public interface Watcher {

    /**
    * Event的状态
    */
    public interface Event {
        /**
        * 在事件发生时,ZooKeeper的状态
        */
        public enum KeeperState {

            @Deprecated
            Unknown (-1),

            Disconnected (0),

            @Deprecated
            NoSyncConnected (1),

            SyncConnected (3),

            AuthFailed (4),

            ConnectedReadOnly (5),

            SaslAuthenticated(6),

            Expired (-112);

            private final int intValue; 

            KeeperState( int intValue) {
                this.intValue = intValue;
            } 

            ......
        }

        /**
        * ZooKeeper中的事件
        */
        public enum EventType {
            None (-1),
            NodeCreated (1),
            NodeDeleted (2),
            NodeDataChanged (3),
            NodeChildrenChanged (4);

            private final int intValue;    // Integer representation of value
                                            // for sending over wire
            EventType( int intValue) {
                this.intValue = intValue;
            }
            ...... 
        }
    }

    //Watcher的回调方法
    abstract public void process(WatchedEvent event);
}


 

Code 2: Zookeeper.getChildren(final String, Watcher)方法

public List<String> getChildren(final String path, Watcher watcher)
    throws KeeperException, InterruptedException
{
    final String clientPath = path;
    PathUtils. validatePath(clientPath);

    WatchRegistration wcb = null;
    //如果watcher不等于null, 构建WatchRegistration对象,
    //该对象描述了watcher和path之间的关系
    if (watcher != null) {
        wcb = new ChildWatchRegistration(watcher, clientPath);
    }
   
    //在传入的path加上root path前缀,构成服务器端的绝对路径
    final String serverPath = prependChroot(clientPath);
   
    //构建RequestHeader对象
    RequestHeader h = new RequestHeader();
    //设置操作类型为OpCode. getChildren
    h.setType(ZooDefs.OpCode. getChildren);
    //构建GetChildrenRequest对象
    GetChildrenRequest request = new GetChildrenRequest();
    //设置path
    request.setPath(serverPath);
    //设置是否使用watcher
    request.setWatch(watcher != null);
    //构建GetChildrenResponse对象
    GetChildrenResponse response = new GetChildrenResponse();
    //提交请求,并阻塞等待结果
    ReplyHeader r = cnxn.submitRequest(h, request, response, wcb);
    if (r.getErr() != 0) {
        throw KeeperException.create(KeeperException.Code. get(r.getErr()),
                clientPath);
    }
    return response.getChildren();
}


Follower的NIOServerCnxn类接到了Client的请求,会调用ZookeeperServer.processPacket()方法。该方法会构建一个Request对象,并调用第一个处理器FollowerRequestProcessor。

由于我们的请求只是一个读操作,而不是一个Quorum请求或者sync请求,所以FollowerRequestProcessor不需要调用Follower.request()方法将请求转给Leader,只需要将请求传递到下一个处理器CommitProcessor。

处理器CommitProcessor线程发现请求是读请求后,直接将Requet对象加入到toProcess队列中,在接下的循环中会调用FinalRequestProcessor.processRequest方法进行处理。

FinalRequestProcessor.processRequest方法最终会调用ZKDatabase中的读操作方法(如statNode和getData方法), 而ZKDatabase的这些方法会最终调用DataTree类的方法来获取指定path的znode信息并返回给Client端,同时也会设置Watcher。

Code 3: FinalRequestProcessor对OpCode.getData请求的处理

case OpCode. getData: {
              lastOp = "GETD";
              GetDataRequest getDataRequest = new GetDataRequest();
              ByteBufferInputStream. byteBuffer2Record(request.request,
                      getDataRequest);
              //获得znode对象
              DataNode n = zks.getZKDatabase().getNode(getDataRequest.getPath());
              //n为null, 抛出NoNodeException异常
              if (n == null) {
                  throw new KeeperException.NoNodeException();
              }
              Long aclL;
              synchronized(n) {
                  aclL = n. acl;
              }
              //检查是否有读权限
              PrepRequestProcessor. checkACL(zks, zks.getZKDatabase().convertLong(aclL),
                      ZooDefs.Perms. READ,
                      request. authInfo);
              //构建状态对象stat
              Stat stat = new Stat();
              //获得指定path的znode数据,
              //如果GetDataRequest.getWatcher()返回true, 将ServerCnxn类型对象cnxn传递进去。
              //ServerCnxn是实现了Watcher接口
              byte b[] = zks.getZKDatabase().getData(getDataRequest.getPath(), stat,
                      getDataRequest. getWatch() ? cnxn : null);
              //构建GetDataResponse对象
              rsp = new GetDataResponse(b, stat);
              break;
          }


Code 4: DataTree.getData()方法

public byte[] getData(String path, Stat stat, Watcher watcher)
        throws KeeperException.NoNodeException {
    //从nodes map中获取指定path的DataNode对象
    DataNode n = nodes.get(path);
    //如果n为null, 则抛出NoNodeException异常
    if (n == null) {
        throw new KeeperException.NoNodeException();
    }
    synchronized (n) {
        //将n的状态copy到stat中
        n.copyStat(stat);
        //如果watcher不会null, 则将(path, watcher)键值对放入dataWatchers Map里
        if (watcher != null) {
            dataWatches.addWatch(path, watcher);
        }
        //返回节点数据
        return n.data ;
    }
}


2. 修改znode数据触发Watcher
在Zookeeper二阶段提交的COMMIT阶段。当Follower从Leader那接收到一个写请求的Leader.COMMIT数据包,会调用FinalRequestProcessor.processRequest()方法。Leader本身在发送完Leader.COMMIT数据包,也会调用FinalRequestProcessor.processRequest()方法。

如果是setData修改数据请求,那么FinalRequestProcessor.processRequest()方法最终会调用到DataTree.setData方法将txn应用到指定znode上,同时触发Watcher,并发送notification给Client端。

其关SetData请求的时序图如下:

Zookeeper源码分析:Watcher机制

triggerWatcher

Code 5: DataTree.setData()方法

public Stat setData(String path, byte data[], int version, long zxid,
        long time) throws KeeperException.NoNodeException {
    Stat s = new Stat();
    //根据path, 获得DataNode对象n
    DataNode n = nodes.get(path);
    //如果n为null, 则抛出NoNodeException异常
    if (n == null) {
        throw new KeeperException.NoNodeException();
    }
    byte lastdata[] = null;
    synchronized (n) {
        lastdata = n. data;
        n. data = data;
        n. stat.setMtime(time);
        n. stat.setMzxid(zxid);
        n. stat.setVersion(version);
        n.copyStat(s);
    }
    // now update if the path is in a quota subtree.
    String lastPrefix = getMaxPrefixWithQuota(path);
    if(lastPrefix != null) {
      this.updateBytes(lastPrefix, (data == null ? 0 : data.length)
          - (lastdata == null ? 0 : lastdata.length ));
    }
    //触发Watcher
    dataWatches.triggerWatch(path, EventType.NodeDataChanged);
    return s;
}

 


Code 6: WatchManage.triggerWatcher()方法,触发Watcher。

Set<Watcher> triggerWatch(String path, EventType type, Set<Watcher> supress) {
    WatchedEvent e = new WatchedEvent(type,
            KeeperState. SyncConnected, path);
    HashSet<Watcher> watchers;
    synchronized (this ) {
        //从watchTable删除掉path对于的watcher
        watchers = watchTable.remove(path);
        if (watchers == null || watchers.isEmpty()) {
            if (LOG .isTraceEnabled()) {
                ZooTrace. logTraceMessage(LOG,
                        ZooTrace. EVENT_DELIVERY_TRACE_MASK,
                        "No watchers for " + path);
            }
            return null;
        }
        for (Watcher w : watchers) {
            HashSet<String> paths = watch2Paths.get(w);
            if (paths != null) {
                paths.remove(path);
            }
        }
    }
    //循环处理所有关于path的Watcher, 这里Watcher对象实际上就是ServerCnxn类型对象
    for (Watcher w : watchers) {
        if (supress != null && supress.contains(w)) {
            continue;
        }
        w.process(e);
    }
    return watchers;
}

 

Code 7: NIOServerCnxn.process方法,发送notification给Client端

synchronized public void process (WatchedEvent event) {
    ReplyHeader h = new ReplyHeader(-1, -1L, 0);
    if (LOG .isTraceEnabled()) {
        ZooTrace. logTraceMessage(LOG, ZooTrace.EVENT_DELIVERY_TRACE_MASK ,
                                "Deliver event " + event + " to 0x"
                                + Long. toHexString(this. sessionId)
                                + " through " + this );
    }

    // Convert WatchedEvent to a type that can be sent over the wire
    WatcherEvent e = event.getWrapper();
   
    //发送notification给Client端
    sendResponse(h, e, "notification");
}


3. 总结
Watcher具有one-time trigger的特性,在代码中我们也可以看到一个watcher被处理后会立即从watchTable中删掉。

--------------------------------------分割线 --------------------------------------

相关推荐