[ZooKeeper]连接中断,watch恢复,心跳和客户端超时

shawnmei

　　前一篇文章分析了server端主动超时session的情况，接下来看一下client和server网络暂时中断的情况。
　　1.和server主动关闭连接一样，client抛出EndOfStreamException异常，此时客户端状态还是CONNECTED
　　2.SendThread处理异常，清理连接，将当前所有请求置为失败，错误码是CONNECTIONLOSS
　　3.发送Disconnected状态通知
　　4.选下一个server重连
　　5.连上之后发送ConnectRequest，sessionid和password是当前session的数据
　　6.server端处理，分leader和follower，由于此时client端重试比较快，session还没超时，所以leader和follower端session校验成功。如果这个时候session正好超时了，则校验失败，client会抛出sessionExpired异常并退出
　　7.server端返回成功的ConnectResponse
　　8.client收到相应，发送SyncConnected状态通知给watcher
　　9.client发送SetWatches包，重建watch

//可以通过配置禁止重建watch
if (!disableAutoWatchReset) {
//当前的所有watch
List<String> dataWatches = zooKeeper.getDataWatches();
List<String> existWatches = zooKeeper.getExistWatches();
List<String> childWatches = zooKeeper.getChildWatches();
if (!dataWatches.isEmpty()
|| !existWatches.isEmpty() || !childWatches.isEmpty()) {
//发送重建请求
SetWatches sw = new SetWatches(lastZxid,
prependChroot(dataWatches),
prependChroot(existWatches),
prependChroot(childWatches));
RequestHeader h = new RequestHeader();
h.setType(ZooDefs.OpCode.setWatches);
h.setXid(-8);
Packet packet = new Packet(h, new ReplyHeader(), sw, null, null);
outgoingQueue.addFirst(packet);
}
}
　　10.server端收到setWatches请求，如果是follower，直接进入FinalRequestProcessor处理，无需proposal

case OpCode.setWatches: {
lastOp = "SETW";
SetWatches setWatches = new SetWatches();
// XXX We really should NOT need this!!!!
request.request.rewind();
ByteBufferInputStream.byteBuffer2Record(request.request, setWatches);
long relativeZxid = setWatches.getRelativeZxid();
//添加watch
zks.getZKDatabase().setWatches(relativeZxid,
setWatches.getDataWatches(),
setWatches.getExistWatches(),
setWatches.getChildWatches(), cnxn);
break;
}

//添加watch的时候判断watch是否需要触发
public void setWatches(long relativeZxid, List<String> dataWatches,
List<String> existWatches, List<String> childWatches,
Watcher watcher) {
for (String path : dataWatches) {
DataNode node = getNode(path);
WatchedEvent e = null;
if (node == null) {
e = new WatchedEvent(EventType.NodeDeleted,
KeeperState.SyncConnected, path);
} else if (node.stat.getCzxid() > relativeZxid) {
e = new WatchedEvent(EventType.NodeCreated,
KeeperState.SyncConnected, path);
} else if (node.stat.getMzxid() > relativeZxid) {
e = new WatchedEvent(EventType.NodeDataChanged,
KeeperState.SyncConnected, path);
}
if (e != null) {
watcher.process(e);
} else {
this.dataWatches.addWatch(path, watcher);
}
}
for (String path : existWatches) {
DataNode node = getNode(path);
WatchedEvent e = null;
if (node == null) {
// This is the case when the watch was registered
} else if (node.stat.getMzxid() > relativeZxid) {
e = new WatchedEvent(EventType.NodeDataChanged,
KeeperState.SyncConnected, path);
} else {
e = new WatchedEvent(EventType.NodeCreated,
KeeperState.SyncConnected, path);
}
if (e != null) {
watcher.process(e);
} else {
this.dataWatches.addWatch(path, watcher);
}
}
for (String path : childWatches) {
DataNode node = getNode(path);
WatchedEvent e = null;
if (node == null) {
e = new WatchedEvent(EventType.NodeDeleted,
KeeperState.SyncConnected, path);
} else if (node.stat.getPzxid() > relativeZxid) {
e = new WatchedEvent(EventType.NodeChildrenChanged,
KeeperState.SyncConnected, path);
}
if (e != null) {
watcher.process(e);
} else {
this.childWatches.addWatch(path, watcher);
}
}
}
　　11.如果是leader，则多了一层PrepRequestProcessor的处理，检查session是否还在
　　再来看看客户端主动超时Session和心跳的情况，SendThread主线程

public void run() {
clientCnxnSocket.introduce(this,sessionId);
clientCnxnSocket.updateNow();
clientCnxnSocket.updateLastSendAndHeard();
//selector的select超时时间，每次循环都会重新计算
int to;
long lastPingRwServer = System.currentTimeMillis();
while (state.isAlive()) {
try {
......
//session建立之后，to为读超时减去读空闲时间
if (state.isConnected()) {
......
to = readTimeout - clientCnxnSocket.getIdleRecv();
} else {
to = connectTimeout - clientCnxnSocket.getIdleRecv();
}
//如果client长时间没收到server的packet，会导致读空闲时间很长，超过读超时，直接抛出异常
if (to <= 0) {
throw new SessionTimeoutException(
"Client session timed out, have not heard from server in "
+ clientCnxnSocket.getIdleRecv() + "ms"
+ " for sessionid 0x"
+ Long.toHexString(sessionId));
}
//session建立之后，发送心跳
if (state.isConnected()) {
//如果写频繁，则写空闲时间很少，不用发送心跳
int timeToNextPing = readTimeout / 2
- clientCnxnSocket.getIdleSend();
//写少，发心跳
if (timeToNextPing <= 0) {
sendPing();
//上次发送时间
clientCnxnSocket.updateLastSend();
}
//写繁忙，不用发送心跳
else {
if (timeToNextPing < to) {
to = timeToNextPing;
}
}
}
.....
//每次doTransport都会更新now，lastHeard和lastSend则取决于是否有读写请求
clientCnxnSocket.doTransport(to, pendingQueue, outgoingQueue, ClientCnxn.this);
} catch (Throwable e) {
....
clientCnxnSocket.updateNow();
clientCnxnSocket.updateLastSendAndHeard();
}
}
}
.....
}
　　心跳包，xid为-2

        private void sendPing() {
lastPingSentNs = System.nanoTime();
RequestHeader h = new RequestHeader(-2, OpCode.ping);
queuePacket(h, null, null, null, null, null, null, null, null);
}
　　server端处理ping包，如果是follower直接进入FinalRequestProcessor处理

case OpCode.ping: {
zks.serverStats().updateLatency(request.createTime);
lastOp = "PING";
cnxn.updateStatsForResponse(request.cxid, request.zxid, lastOp,
request.createTime, System.currentTimeMillis());
//心跳包的响应xid也是-2
cnxn.sendResponse(new ReplyHeader(-2,
zks.getZKDatabase().getDataTreeLastProcessedZxid(), 0), null, "response");
return;
}
　　如果是leader，则多了一层PrepRequestProcessor的处理，检查session是否还在
　　client收到心跳包响应，啥事不做

if (replyHdr.getXid() == -2) {
// -2 is the xid for pings
if (LOG.isDebugEnabled()) {
LOG.debug("Got ping response for sessionid: 0x"
+ Long.toHexString(sessionId)
+ " after "
+ ((System.nanoTime() - lastPingSentNs) / 1000000)
+ "ms");
}
return;
}
　　以上可以看出
　　1.心跳包只有写空闲时才会发送
　　2.每次transport的时候都会更新当前时间now
　　3.lastHeard和lastSend取决于是否有读写请求
　　4.客户端session超时和连接关闭CONNECTIONLOSS处理是一样的，都会导致重试