Curator高级API - Master选举
在分布式系统中，经常会碰到这样的场景：对于一个复杂的任务，仅需要从集群中选举出一台进行处理即可。称为’Master选举’问题。借助Zookeeper，可以比较简单的实现Master选举的功能，大体思路如下：
选择一个根节点，例如/master_select,多台机器同事向该节点创建一个子节点/master_select/lock,利用Zookeeper的特性，最终只有一台机器能够创建成功，成功的哪台机器就作为Master。
Curator也是基于这个思路，但是它将节点创建，事件监听和自动选举过程进行了封装，开发人员只需要条用简单的API就可以实现Master选举
Curator实现了两套选举API，分别是LeaderSelector和LeaderLatch。
LeaderSelector
LeaderSelector支持任务执行完成之后自动重新竞选。针对单个任务在集群中竞争执行，只允许一台服务器执行场景。
示例代码
package com.freud.zk.curator;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import org.apache.curator.RetryPolicy;
import org.apache.curator.framework.CuratorFramework;
import org.apache.curator.framework.CuratorFrameworkFactory;
import org.apache.curator.framework.recipes.leader.LeaderSelector;
import org.apache.curator.framework.recipes.leader.LeaderSelectorListener;
import org.apache.curator.framework.state.ConnectionState;
import org.apache.curator.retry.ExponentialBackoffRetry;
/**
*
* Zookeeper - Curator - Leader Election
*
* @author Freud
*
*/
public class CuratorLeaderElectionZookeeper {
private static final int SECOND = 1000;
private static int count = 1;
public static void main(String[] args) throws Exception {
ExecutorService service = Executors.newFixedThreadPool(3);
for (int i = 0; i < 3; i++) {
final int index = i;
service.submit(new Runnable() {
public void run() {
try {
new CuratorLeaderElectionZookeeper().schedule(index);
} catch (Exception e) {
e.printStackTrace();
}
}
});
}
Thread.sleep(10 * SECOND);
service.shutdownNow();
}
private void schedule(final int thread) throws Exception {
CuratorFramework client = this.getStartedClient(thread);
String path = "/leader_selector";
if (client.checkExists().forPath(path) == null) {
client.create().creatingParentsIfNeeded().forPath(path);
}
LeaderSelector selector = new LeaderSelector(client, path, new LeaderSelectorListener() {
public void stateChanged(CuratorFramework cf, ConnectionState state) {
System.out.println("Thread [" + thread + "][Callback] State changed to :" + state.name());
}
public void takeLeadership(CuratorFramework cf) throws Exception {
Thread.sleep(1 * SECOND);
System.out.println("Thread [" + thread + "]Do some business work...timestamp ["
+ System.currentTimeMillis() + "] times [" + count++ + "]");
}
});
// 自动重新部署竞选
selector.autoRequeue();
selector.start();
}
private CuratorFramework getStartedClient(final int thread) {
RetryPolicy rp = new ExponentialBackoffRetry(1 * SECOND, 3);
// Fluent风格创建
CuratorFramework cfFluent = CuratorFrameworkFactory.builder().connectString("localhost:2181")
.sessionTimeoutMs(5 * SECOND).connectionTimeoutMs(3 * SECOND).retryPolicy(rp).build();
cfFluent.start();
System.out.println("Thread [" + thread + "]Server connected...");
return cfFluent;
}
}打印结果
Thread [1]Server connected...
Thread [2]Server connected...
Thread [0]Server connected...
Thread [0]Do some business work...timestamp [1484621709106] times [1]
Thread [1]Do some business work...timestamp [1484621710198] times [2]
Thread [2]Do some business work...timestamp [1484621711329] times [3]
Thread [0]Do some business work...timestamp [1484621712396] times [4]
Thread [1]Do some business work...timestamp [1484621713438] times [5]
Thread [2]Do some business work...timestamp [1484621714586] times [6]
Thread [0]Do some business work...timestamp [1484621715688] times [7]
Thread [1]Do some business work...timestamp [1484621716725] times [8]观察结果之后发现在相同的秒数内只有一个Server执行了打印输出。
LeaderLatch
LeaderLatch支持针对竞选失败情况下的操作。并且选出的Master将为长期保持Master状态，在Master节点挂掉之后，会马上在集群中选出新的Master节点。
示例代码
package com.freud.zk.curator;
import java.text.MessageFormat;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import org.apache.curator.RetryPolicy;
import org.apache.curator.framework.CuratorFramework;
import org.apache.curator.framework.CuratorFrameworkFactory;
import org.apache.curator.framework.recipes.leader.LeaderLatch;
import org.apache.curator.framework.recipes.leader.LeaderLatchListener;
import org.apache.curator.retry.ExponentialBackoffRetry;
/**
*
* Zookeeper - Curator - Leader Latch
*
* @author Freud
*
*/
public class CuratorLeaderLatchZookeeper {
private static final int SECOND = 1000;
public static void main(String[] args) throws Exception {
ExecutorService service = Executors.newFixedThreadPool(3);
for (int i = 0; i < 3; i++) {
final int index = i;
service.submit(new Runnable() {
public void run() {
try {
new CuratorLeaderLatchZookeeper().schedule(index);
} catch (Exception e) {
e.printStackTrace();
}
}
});
}
Thread.sleep(10 * SECOND);
service.shutdownNow();
}
private void schedule(final int thread) throws Exception {
CuratorFramework client = this.getStartedClient(thread);
String path = "/leader_latch";
if (client.checkExists().forPath(path) == null) {
client.create().creatingParentsIfNeeded().forPath(path);
}
LeaderLatch latch = new LeaderLatch(client, path);
latch.addListener(new LeaderLatchListener() {
public void notLeader() {
System.out.println(MessageFormat.format("Thread [" + thread
+ "] I am not the leader... timestamp [{0}]", System.currentTimeMillis()));
}
public void isLeader() {
System.out.println(MessageFormat.format("Thread [" + thread + "] I am the leader... timestamp [{0}]",
System.currentTimeMillis()));
}
});
latch.start();
Thread.sleep(2 * (thread + 1) * SECOND);
if (latch != null) {
latch.close();
}
if (client != null) {
client.close();
}
System.out.println("Thread [" + thread + "] Server closed...");
}
private CuratorFramework getStartedClient(final int thread) {
RetryPolicy rp = new ExponentialBackoffRetry(1 * SECOND, 3);
// Fluent风格创建
CuratorFramework cfFluent = CuratorFrameworkFactory.builder().connectString("localhost:2181")
.sessionTimeoutMs(5 * SECOND).connectionTimeoutMs(3 * SECOND).retryPolicy(rp).build();
cfFluent.start();
System.out.println("Thread [" + thread + "] Server connected...");
return cfFluent;
}
}连续启动三次之后输出结果如下
打印结果
Thread [1] Server connected...
Thread [0] Server connected...
Thread [2] Server connected...
Thread [0] I am the leader... timestamp [1,484,622,048,539]
Thread [0] Server closed...
Thread [1] I am the leader... timestamp [1,484,622,050,431]
Thread [2] I am the leader... timestamp [1,484,622,052,427]
Thread [1] Server closed...
Thread [2] Server closed...观察结果之后发现在Master节点挂掉之后，会马上重新选举一个新的Master出来。
参考资料
《从PAXOS到ZOOKEEPER分布式一致性原理与实践》 - 倪超
Curator官网 : http://curator.apache.org/