Zookeeper（六）应用实例

6.1 JAVA API

客户端要连接 Zookeeper服务器可以通过创建 org.apache.zookeeper.ZooKeeper 的一个实例对象，然后调用这个类提供的接口来和服务器交互。

ZooKeeper 主要是用来维护和监控一个目录节点树中存储的数据的状态，所有我们能够操作 ZooKeeper 和操作目录节点树大体一样，如创建一个目录节点，给某个目录节点设置数据，获取某个目录节点的所有子目录节点，给某个目录节点设置权限和监控这个目录节点的状态变化。

下面通过代码实例，来熟悉一下JavaAPI的常用方法。

package com.yc.Hadoop42_006_zookeeper;import java.util.List;import org.apache.zookeeper.CreateMode;import org.apache.zookeeper.WatchedEvent;import org.apache.zookeeper.Watcher;import org.apache.zookeeper.ZooDefs.Ids;import org.apache.zookeeper.ZooKeeper;import org.apache.zookeeper.data.Stat;public class ZookeeperDemo1 {public static void main(String[] args) throws Exception {// 定义一个监控所有节点变化的watcherWatcher watcher = new Watcher() {@Overridepublic void process(WatchedEvent event) {System.out.println("watcher receive WatchEvent, changed path: " + event.getPath() + "; changed type: "+ event.getType().name());}};/** * 初始化一个与Zookeeper的连接，三个参数：  * 1. 要连接的服务器地址，"IP:port"的格式  * 2. 会话超时时间  * 3.节点变化监视器 */ZooKeeper zk = new ZooKeeper("slave01:2181,slave02:2181,slave03:2181", 5000, watcher);// 创建节点 四个参数：1.节点路径（绝对路径） 2.节点数据 3.节点权限 4.节点类型zk.create("/newNode", "Hello World".getBytes(), Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT);System.out.println("create new node : '/newNode' ");// 判断某路径是否存在 两个参数：1.节点路径 2.是否进行监视（watcher即初始化Zookeeper时传入的watcher）Stat beforeStat = zk.exists("/newNode", true);System.out.println("stat of '/newNode' before change : " + beforeStat.toString());// 修改节点数据。三个参数：1.节点路径 2.新数据 3.版本，如果为-1，则匹配任何版本Stat afterStat = zk.setData("/newNode", "World Hello".getBytes(), -1);System.out.println("stat of '/newNode' after change : " + afterStat.toString());// 获取所有子节点数据 两个参数：1.节点路径 2.是否监视List<String> children = zk.getChildren("/", true);System.out.println("children of path '/' : " + children.toString());// 获取节点数据。三个参数：1.节点路径；2.书否监控该节点；3.版本等信息可以通过一个Stat对象来指定byte[] nameByte = zk.getData("/newNode", true, null);String name = new String(nameByte, "utf-8");System.out.println("get Data from '/newNode':" + name);//删除节点 两个参数：1.节点路径 2.版本-1可以匹配任何版本，会删除所有数据zk.delete("/newNode", -1);System.out.println("delete '/newNode'");zk.close();}}

测试结果：

watcher receive WatchEvent, changed path: null; changed type: None
create new node : '/newNode'
stat of '/newNode' before change : 219043332105,219043332105,1502095223367,1502095223367,0,0,0,0,11,0,219043332105

watcher receive WatchEvent, changed path: /newNode; changed type: NodeDataChanged
stat of '/newNode' after change : 219043332105,219043332106,1502095223367,1502095223389,1,0,0,0,11,0,219043332105

children of path '/' : [newNode, zookeeper, hbase]
get Data from '/newNode':World Hello
delete '/newNode'
watcher receive WatchEvent, changed path: /newNode; changed type: NodeDeleted
watcher receive WatchEvent, changed path: /; changed type: NodeChildrenChanged

更详细的API请参考官方网站。

 

Zookeeper 从设计模式角度来看，是一个基于观察者模式设计的分布式服务管理框架，它负责存储和管理大家都关心的数据，然后接受观察者的注册，一旦这些数据的状态发生变化，Zookeeper 将负责通知已经在 Zookeeper 上注册的那些观察者做出相应的反应。

下面通过两个ZooKeeper的典型用用场景来体会下ZooKeeper的特性与使用方法。

6.2 分布式锁

先来回顾一下多线程中的锁控制

package com.yc.Hadoop42_006_zookeeper;public class MultiThreadTest {// 以一个静态变量来模拟公共资源private static int counter = 0;// 多线程环境下，会出现并发问题public static void plus() {// 计数器加一counter++;// 线程随机休眠数毫秒，模拟现实中的耗时操作int sleepMillis = (int) (Math.random() * 100);try {Thread.sleep(sleepMillis);} catch (InterruptedException e) {e.printStackTrace();}}// 线程实现类static class CountPlus extends Thread {@Overridepublic void run() {for (int i = 0; i < 20; i++) {plus();}System.out.println(Thread.currentThread().getName() + "执行完毕：" + counter);}public CountPlus(String threadName) {super(threadName);}}public static void main(String[] args) throws Exception {// 开启五个线程CountPlus threadA = new CountPlus("threadA");threadA.start();CountPlus threadB = new CountPlus("threadB");threadB.start();CountPlus threadC = new CountPlus("threadC");threadC.start();CountPlus threadD = new CountPlus("threadD");threadD.start();CountPlus threadE = new CountPlus("threadE");threadE.start();}}

上例中，开启了五个线程，每个线程通过plus()方法对静态变量counter分别进行20次累加，预期counter最后会变成100。运行程序：

threadC执行完毕：81
threadA执行完毕：93
threadE执行完毕：97
threadD执行完毕：99
threadB执行完毕：99

可以发现，五个线程执行完毕之后，counter并没有变成100。plus()方法涉及到对公共资源的改动，但是并没有对它进行同步控制，可能会造成多个线程同时对公共资源发起改动，进而出现并发问题。问题的根源在于，上例中没有保证同一时刻只能有一个线程可以改动公共资源。

给plus()方法加上synchronized关键字，重新运行程序：

threadE执行完毕：51
threadC执行完毕：66
threadD执行完毕：86
threadA执行完毕：95
threadB执行完毕：100

可见，最终达到了预期结果。

synchronized关键字的作用是对plus()方法加入锁控制，一个线程想要执行该方法，首先需要获得锁（锁是唯一的），执行完毕后，再释放锁。如果得不到锁，该线程会进入等待池中等待，直到抢到锁才能继续执行。这样就保证了同一时刻只能有一个线程可以改动公共资源，避免了并发问题。

共享锁在同一个进程中很容易实现，可以靠Java本身提供的同步机制解决，但是在跨进程或者在不同 Server 之间就不好实现了，这时候就需要一个中间人来协调多个Server之间的各种问题，比如如何获得锁/释放锁、谁先获得锁、谁后获得锁等。

有了zookeeper的一致性文件系统，锁的问题变得容易。锁服务可以分为两类，一个是保持独占，另一个是控制时序。

对于第一类，我们将zookeeper上的一个znode看作是一把锁，通过createznode的方式来实现。所有客户端都去创建 /distribute_lock 节点，最终成功创建的那个客户端也即拥有了这把锁。厕所有言：来也冲冲，去也冲冲，用完删除掉自己创建的distribute_lock 节点就释放出锁。

对于第二类， /distribute_lock 已经预先存在，所有客户端在它下面创建临时顺序编号目录节点，和选master一样，编号最小的获得锁，用完删除，依次方便。

借助 Zookeeper 可以实现这种分布式锁：需要获得锁的 Server 创建一个 EPHEMERAL_SEQUENTIAL 目录节点，然后调用 getChildren( ) 方法获取列表中最小的目录节点，如果最小节点就是自己创建的目录节点，那么它就获得了这个锁，如果不是那么它就调用 exists( ) 方法并监控前一节点的变化，一直到自己创建的节点成为列表中最小编号的目录节点，从而获得锁。释放锁很简单，只要删除它自己所创建的目录节点就行了。

流程图：

下面我们对刚才的代码进行改造，不用synchronize关键字而是使用ZooKeeper达到锁控制的目的，模拟分布式锁的实现。

package com.yc.Hadoop42_006_zookeeper;import java.util.Collections;import java.util.List;import org.apache.zookeeper.CreateMode;import org.apache.zookeeper.KeeperException;import org.apache.zookeeper.WatchedEvent;import org.apache.zookeeper.Watcher;import org.apache.zookeeper.ZooDefs.Ids;import org.apache.zookeeper.ZooKeeper;import org.apache.zookeeper.data.Stat;public class ZkDistributedLock {// 以一个静态变量来模拟公共资源private static int counter = 0;public static void plus() {// 计数器加一counter++;// 线程随机休眠数毫秒，模拟现实中的费时操作int sleepMillis = (int) (Math.random() * 100);try {Thread.sleep(sleepMillis);} catch (InterruptedException e) {e.printStackTrace();}}// 线程实现类static class CountPlus extends Thread {private static final String LOCK_ROOT_PATH = "/Locks";private static final String LOCK_NODE_NAME = "Lock_";// 每个线程持有一个zk客户端，负责获取锁与释放锁ZooKeeper zkClient;@Overridepublic void run() {for (int i = 0; i < 20; i++) {// 访问计数器之前需要先获取锁String path = getLock();// 执行任务plus();// 执行完任务后释放锁releaseLock(path);}closeZkClient();System.out.println(Thread.currentThread().getName() + "执行完毕：" + counter);}/** * 获取锁，即创建子节点，当该节点成为序号最小的节点时则获取锁 */private String getLock() {try {// 创建EPHEMERAL_SEQUENTIAL类型节点String lockPath = zkClient.create(LOCK_ROOT_PATH + "/" + LOCK_NODE_NAME,Thread.currentThread().getName().getBytes(), Ids.OPEN_ACL_UNSAFE,CreateMode.EPHEMERAL_SEQUENTIAL);System.out.println(Thread.currentThread().getName() + " create path : " + lockPath);// 尝试获取锁tryLock(lockPath);return lockPath;} catch (Exception e) {e.printStackTrace();}return null;}/** * 该函数是一个递归函数 如果获得锁，直接返回；否则，阻塞线程，等待上一个节点释放锁的消息，然后重新tryLock */private boolean tryLock(String lockPath) throws KeeperException, InterruptedException {// 获取LOCK_ROOT_PATH下所有的子节点，并按照节点序号排序List<String> lockPaths = zkClient.getChildren(LOCK_ROOT_PATH, false);Collections.sort(lockPaths);int index = lockPaths.indexOf(lockPath.substring(LOCK_ROOT_PATH.length() + 1));if (index == 0) { // lockPath是序号最小的节点，则获取锁System.out.println(Thread.currentThread().getName() + " get lock, lockPath: " + lockPath);return true;} else { // lockPath不是序号最小的节点// 创建Watcher，监控lockPath的前一个节点Watcher watcher = new Watcher() {@Overridepublic void process(WatchedEvent event) {System.out.println(event.getPath() + " has been deleted");synchronized (this) {notifyAll();}}};String preLockPath = lockPaths.get(index - 1);Stat stat = zkClient.exists(LOCK_ROOT_PATH + "/" + preLockPath, watcher);if (stat == null) { // 由于某种原因，前一个节点不存在了（比如连接断开），重新tryLockreturn tryLock(lockPath);} else { // 阻塞当前进程，直到preLockPath释放锁，重新tryLockSystem.out.println(Thread.currentThread().getName() + " wait for " + preLockPath);synchronized (watcher) {watcher.wait();}return tryLock(lockPath);}}}/** * 释放锁，即删除lockPath节点 */private void releaseLock(String lockPath) {try {zkClient.delete(lockPath, -1);} catch (InterruptedException | KeeperException e) {e.printStackTrace();}}public void setZkClient(ZooKeeper zkClient) {this.zkClient = zkClient;}public void closeZkClient() {try {zkClient.close();} catch (InterruptedException e) {e.printStackTrace();}}public CountPlus(String threadName) {super(threadName);}}public static void setZkClient(CountPlus thread) throws Exception {ZooKeeper zkClient = new ZooKeeper("slave01:2181,slave02:2181,slave03:2181", 5000, null);thread.setZkClient(zkClient);}public static void main(String[] args) throws Exception {// 开启五个线程CountPlus threadA = new CountPlus("threadA");setZkClient(threadA);threadA.start();CountPlus threadB = new CountPlus("threadB");setZkClient(threadB);threadB.start();CountPlus threadC = new CountPlus("threadC");setZkClient(threadC);threadC.start();CountPlus threadD = new CountPlus("threadD");setZkClient(threadD);threadD.start();CountPlus threadE = new CountPlus("threadE");setZkClient(threadE);threadE.start();}}

注意：运行程序之前需要创建“/Locks”作为存放锁信息的根节点。

一旦某个Server想要获得锁，就会在/Locks”下创建一个EPHEMERAL_SEQUENTIAL类型的名为“Lock_”子节点，ZooKeeper会自动为每个子节点附加一个递增的编号，该编号为int类型，长度为10，左端以0补全。“/Locks”下会维持着这样一系列的节点：

Lock_0000000001,Lock_0000000002, Lock_0000000003, Lock_0000000004…

一旦这些创建这些节点的Server断开连接，该节点就会被清除（当然也可以主动清除）。

由于节点的编号是递增的，创建越晚排名越靠后。遵循先到先得的原则，Server创建完节点之后会检查自己的节点是不是最小的，如果是，那就获得锁，如果不是，排队等待。执行完任务之后，Server清除自己创建的节点，这样后面的节点会依次获得锁。

程序的运行结果如下：

/Locks/Lock_0000000353 has been deleted
threadB create path : /Locks/Lock_0000000358
threadD get lock, lockPath: /Locks/Lock_0000000354
threadB wait for Lock_0000000357
/Locks/Lock_0000000354 has been deleted
threadA get lock, lockPath: /Locks/Lock_0000000355
threadD create path : /Locks/Lock_0000000359
threadD wait for Lock_0000000358
/Locks/Lock_0000000355 has been deleted
threadC get lock, lockPath: /Locks/Lock_0000000356
threadA执行完毕：97
/Locks/Lock_0000000356 has been deleted
threadE get lock, lockPath: /Locks/Lock_0000000357
threadC执行完毕：98
/Locks/Lock_0000000357 has been deleted
threadB get lock, lockPath: /Locks/Lock_0000000358
threadE执行完毕：99
/Locks/Lock_0000000358 has been deleted
threadD get lock, lockPath: /Locks/Lock_0000000359
threadB执行完毕：100
threadD执行完毕：100

6.3 分布式队列

很多单机上很平常的事情，放在集群环境中都会发生质的变化。

以一个常见的生产者-消费者模型举例：有一个容量有限的邮筒，寄信者（即生产者）不断地将信件塞入邮筒，邮递员（即消费者）不断地从邮筒取出信件发往目的地。运行期间需要保证：

（1）邮筒已达上限时，寄信者停止活动，等带邮筒恢复到非满状态

（2）邮筒已空时，邮递员停止活动，等带邮筒恢复到非空状态

该邮筒用有序队列实现，保证FIFO（先进先出）特性。

在一台机器上，可以用有序队列来实现邮筒，保证FIFO（先进先出）特性，开启两个线程，一个充当寄信者，一个充当邮递员，通过wait()/notify()很容易实现上述功能。

但是，如果在跨进程或者分布式环境下呢？比如，一台机器运行生产者程序，另一台机器运行消费者程序，代表邮筒的有序队列无法跨机器共享，但是两者需要随时了解邮筒的状态（是否已满、是否已空）以及保证信件的有序（先到达的先发送）。

这种情况下，可以借助ZooKeeper实现一个分布式队列。新建一个“/mailBox”节点代表邮筒。一旦有信件到达，就在该节点下创建PERSISTENT_SEQUENTIAL类型的子节点，当子节点总数达到上限时，阻塞生产者，然后使用getChildren(String path, Watcher watcher)方法监控子节点的变化，子节点总数减少后再回复生产；而消费者每次选取序号最小的子节点进行处理，然后删除该节点，当子节点总数为0时，阻塞消费者，同样设置监控，子节点总数增加后再回复消费。

package com.yc.Hadoop42_006_zookeeper;import java.io.IOException;import java.util.Collections;import java.util.List;import org.apache.zookeeper.CreateMode;import org.apache.zookeeper.KeeperException;import org.apache.zookeeper.WatchedEvent;import org.apache.zookeeper.Watcher;import org.apache.zookeeper.ZooDefs.Ids;import org.apache.zookeeper.ZooKeeper;import org.apache.zookeeper.data.Stat;public class ZkDistributedQueue {// 邮箱上限为10封信private static final int MAILBOX_MAX_SIZE = 10;// 邮箱路径private static final String MAILBOX_ROOT_PATH = "/mailBox";// 信件节点private static final String LETTER_NODE_NAME = "letter_";// 生产者线程，负责接受信件static class Producer extends Thread {ZooKeeper zkClient;@Overridepublic void run() {while (true) {try {if (getLetterNum() == MAILBOX_MAX_SIZE) { // 信箱已满System.out.println("mailBox has been full");// 创建Watcher，监控子节点的变化Watcher watcher = new Watcher() {@Overridepublic void process(WatchedEvent event) {// 生产者已停止，只有消费者在活动，所以只可能出现发送信件的动作System.out.println("mailBox has been not full");synchronized (this) {notify(); // 唤醒生产者}}};zkClient.getChildren(MAILBOX_ROOT_PATH, watcher);synchronized (watcher) {watcher.wait(); // 阻塞生产者}} else {// 线程随机休眠数毫秒，模拟现实中的费时操作int sleepMillis = (int) (Math.random() * 1000);Thread.sleep(sleepMillis);// 接收信件，创建新的子节点String newLetterPath = zkClient.create(MAILBOX_ROOT_PATH + "/" + LETTER_NODE_NAME,"letter".getBytes(), Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT_SEQUENTIAL);System.out.println("a new letter has been received: "+ newLetterPath.substring(MAILBOX_ROOT_PATH.length() + 1) + ", letter num: "+ getLetterNum());}} catch (Exception e) {System.out.println("producer equit task becouse of exception !");e.printStackTrace();break;}}}private int getLetterNum() throws KeeperException, InterruptedException {Stat stat = zkClient.exists(MAILBOX_ROOT_PATH, null);int letterNum = stat.getNumChildren();return letterNum;}public void setZkClient(ZooKeeper zkClient) {this.zkClient = zkClient;}}// 消费者线程，负责发送信件static class Consumer extends Thread {ZooKeeper zkClient;@Overridepublic void run() {while (true) {try {if (getLetterNum() == 0) { // 信箱已空System.out.println("mailBox has been empty");// 创建Watcher，监控子节点的变化Watcher watcher = new Watcher() {@Overridepublic void process(WatchedEvent event) {// 消费者已停止，只有生产者在活动，所以只可能出现收取信件的动作System.out.println("mailBox has been not empty");synchronized (this) {notify(); // 唤醒消费者}}};zkClient.getChildren(MAILBOX_ROOT_PATH, watcher);synchronized (watcher) {watcher.wait(); // 阻塞消费者}} else {// 线程随机休眠数毫秒，模拟现实中的费时操作int sleepMillis = (int) (Math.random() * 1000);Thread.sleep(sleepMillis);// 发送信件，删除序号最小的子节点String firstLetter = getFirstLetter();zkClient.delete(MAILBOX_ROOT_PATH + "/" + firstLetter, -1);System.out.println("a letter has been delivered: " + firstLetter + ", letter num: " + getLetterNum());}} catch (Exception e) {System.out.println("consumer equit task becouse of exception !");e.printStackTrace();break;}}}private int getLetterNum() throws KeeperException, InterruptedException {Stat stat = zkClient.exists(MAILBOX_ROOT_PATH, false);int letterNum = stat.getNumChildren();return letterNum;}private String getFirstLetter() throws KeeperException, InterruptedException {List<String> letterPaths = zkClient.getChildren(MAILBOX_ROOT_PATH, false);Collections.sort(letterPaths);return letterPaths.get(0);}public void setZkClient(ZooKeeper zkClient) {this.zkClient = zkClient;}}public static void main(String[] args) throws IOException {// 开启生产者线程Producer producer = new Producer();ZooKeeper zkClientA = new ZooKeeper("slave01:2181,slave02:2181,slave03:2181", 5000, null);producer.setZkClient(zkClientA);producer.start();// 开启消费者线程Consumer consumer = new Consumer();ZooKeeper zkClientB = new ZooKeeper("slave01:2181,slave02:2181,slave03:2181", 5000, null);consumer.setZkClient(zkClientB);consumer.start();}}

测试结果：

mailBox has been not empty
a new letter has been received: letter_0000000008, letter num: 1
a letter has been delivered: letter_0000000008, letter num: 0
mailBox has been empty
mailBox has been not empty
a new letter has been received: letter_0000000009, letter num: 1
a letter has been delivered: letter_0000000009, letter num: 0
mailBox has been empty
mailBox has been not empty
a new letter has been received: letter_0000000010, letter num: 1
a letter has been delivered: letter_0000000010, letter num: 0
mailBox has been empty
mailBox has been not empty
a new letter has been received: letter_0000000011, letter num: 1
a new letter has been received: letter_0000000012, letter num: 2
a new letter has been received: letter_0000000013, letter num: 3
a new letter has been received: letter_0000000014, letter num: 4
a new letter has been received: letter_0000000015, letter num: 5
a letter has been delivered: letter_0000000011, letter num: 4
a new letter has been received: letter_0000000016, letter num: 5
a letter has been delivered: letter_0000000012, letter num: 4
a letter has been delivered: letter_0000000013, letter num: 3

上例中还有一个可以改进的地方，在分布式环境下，像MAILBOX_MAX_SIZE这类常量是被多台机器共用的，而且运行期间可能发生改变，比如邮筒上限需要从10改为20，只能停掉机器，然后改动每台机器上的参数，再重新部署。可是，如果该服务不允许停机，而且部署在数十台机器上，让参数在运行时生效且保持一致，怎么办？

这就涉及到了ZooKeeper另一个典型的应用场景——配置中心。被多台机器共享的参数可以托管在ZNode上，对该参数关心的机器在Znode上注册Watcher，一旦该参数发生变化，注册者会收到消息，然后做出相应的调整。

 

ZooKeeper的作用当然不止于此，更多的应用场景就需要使用者在实际项目中发掘跟探索了，毕竟，纸上得来终觉浅，实践出真知。