几天分析了一下三种分布式锁的实现,但是没有利用zookeeper实现一个分布式锁,因为感觉基于Zookeeper实现分布式锁还是稍微复杂的,同时也需要使用Watcher机制,所以就单独搞一篇Zookeeper实现的分布式锁。
首先,第一种实现。我们可以利用Zookeeper不能重复创建一个节点的特性来实现一个分布式锁,这看起来和redis实现分布式锁很像。但是也是有差异的,后面会详细分析。
主要流程图如下:
上面的流程很简单:
- 查看目标Node是否已经创建,已经创建,那么等待锁。
- 如果未创建,创建一个瞬时Node,表示已经占有锁。
- 如果创建失败,那么证明锁已经被其他线程占有了,那么同样等待锁。
- 当释放锁,或者当前Session超时的时候,节点被删除,唤醒之前等待锁的线程去争抢锁。
上面是一个完整的流程,简单的代码实现如下:
package com.codertom.params.engine; import com.google.common.base.Strings; import org.apache.zookeeper.*; import java.io.IOException; import java.util.concurrent.ExecutorService; import java.util.concurrent.Executors; import java.util.concurrent.locks.Lock; /** * Zookeepr实现分布式锁 */ public class LockTest { private String zkQurom = "localhost:2181"; private String lockNameSpace = "/mylock"; private String nodeString = lockNameSpace + "/test1"; private Lock mainLock; private ZooKeeper zk; public LockTest(){ try { zk = new ZooKeeper(zkQurom, 6000, new Watcher() { @Override public void process(WatchedEvent watchedEvent) { System.out.println("Receive event "+watchedEvent); if(Event.KeeperState.SyncConnected == watchedEvent.getState()) System.out.println("connection is established..."); } }); } catch (IOException e) { e.printStackTrace(); } } private void ensureRootPath() throws InterruptedException { try { if (zk.exists(lockNameSpace,true)==null){ zk.create(lockNameSpace,"".getBytes(), ZooDefs.Ids.OPEN_ACL_UNSAFE,CreateMode.PERSISTENT); } } catch (KeeperException e) { e.printStackTrace(); } } private void watchNode(String nodeString, final Thread thread) throws InterruptedException { try { zk.exists(nodeString, new Watcher() { @Override public void process(WatchedEvent watchedEvent) { System.out.println( "==" + watchedEvent.toString()); if(watchedEvent.getType() == Event.EventType.NodeDeleted){ System.out.println("Threre is a Thread released Lock=============="); thread.interrupt(); } try { zk.exists(nodeString,new Watcher() { @Override public void process(WatchedEvent watchedEvent) { System.out.println( "==" + watchedEvent.toString()); if(watchedEvent.getType() == Event.EventType.NodeDeleted){ System.out.println("Threre is a Thread released Lock=============="); thread.interrupt(); } try { zk.exists(nodeString,true); } catch (KeeperException e) { e.printStackTrace(); } catch (InterruptedException e) { e.printStackTrace(); } } }); } catch (KeeperException e) { e.printStackTrace(); } catch (InterruptedException e) { e.printStackTrace(); } } }); } catch (KeeperException e) { e.printStackTrace(); } } /** * 获取锁 * @return * @throws InterruptedException */ public boolean lock() throws InterruptedException { String path = null; ensureRootPath(); watchNode(nodeString,Thread.currentThread()); while (true) { try { path = zk.create(nodeString, "".getBytes(), ZooDefs.Ids.OPEN_ACL_UNSAFE, CreateMode.EPHEMERAL); } catch (KeeperException e) { System.out.println(Thread.currentThread().getName() + " getting Lock but can not get"); try { Thread.sleep(5000); }catch (InterruptedException ex){ System.out.println("thread is notify"); } } if (!Strings.nullToEmpty(path).trim().isEmpty()) { System.out.println(Thread.currentThread().getName() + " get Lock..."); return true; } } } /** * 释放锁 */ public void unlock(){ try { zk.delete(nodeString,-1); System.out.println("Thread.currentThread().getName() + release Lock..."); } catch (InterruptedException e) { e.printStackTrace(); } catch (KeeperException e) { e.printStackTrace(); } } public static void main(String args[]) throws InterruptedException { ExecutorService service = Executors.newFixedThreadPool(10); for (int i = 0;i<4;i++){ service.execute(()-> { LockTest test = new LockTest(); try { test.lock(); Thread.sleep(3000); } catch (InterruptedException e) { e.printStackTrace(); } test.unlock(); }); } service.shutdown(); } }
代码比较糙,但是大致的实现思路和上述一致,这里需要注意:
- 因为使用的是原生的Zookeeper API实现,Watch需要重复的设置,所以代码复杂的些。
- 唤醒直接用的Thread.interupt这样其实控制流程其实是不好的。
其实上面的实现有优点也有缺点:
优点:
实现比较简单,有通知机制,能提供较快的响应,有点类似reentrantlock的思想,对于节点删除失败的场景由Session超时保证节点能够删除掉。
缺点:
重量级,同时在大量锁的情况下会有“惊群”的问题。
“惊群”就是在一个节点删除的时候,大量对这个节点的删除动作有订阅Watcher的线程会进行回调,这对Zk集群是十分不利的。所以需要避免这种现象的发生。
解决“惊群”:
为了解决“惊群“问题,我们需要放弃订阅一个节点的策略,那么怎么做呢?
- 我们将锁抽象成目录,多个线程在此目录下创建瞬时的顺序节点,因为Zk会为我们保证节点的顺序性,所以可以利用节点的顺序进行锁的判断。
- 首先创建顺序节点,然后获取当前目录下最小的节点,判断最小节点是不是当前节点,如果是那么获取锁成功,如果不是那么获取锁失败。
- 获取锁失败的节点获取当前节点上一个顺序节点,对此节点注册监听,当节点删除的时候通知当前节点。
- 当unlock的时候删除节点之后会通知下一个节点。
上面的实现和reentrantlock的公平锁实现还是比较类似的,下面是简单的实现:
package com.codertom.params.engine; import com.google.common.base.Strings; import org.apache.zookeeper.*; import org.apache.zookeeper.data.Stat; import java.io.IOException; import java.util.Collections; import java.util.List; import java.util.concurrent.ExecutorService; import java.util.concurrent.Executors; /** * Created by zhiming on 2017-02-05. */ public class FairLockTest { private String zkQurom = "localhost:2181"; private String lockName = "/mylock"; private String lockZnode = null; private ZooKeeper zk; public FairLockTest(){ try { zk = new ZooKeeper(zkQurom, 6000, new Watcher() { @Override public void process(WatchedEvent watchedEvent) { System.out.println("Receive event "+watchedEvent); if(Event.KeeperState.SyncConnected == watchedEvent.getState()) System.out.println("connection is established..."); } }); } catch (IOException e) { e.printStackTrace(); } } private void ensureRootPath(){ try { if (zk.exists(lockName,true)==null){ zk.create(lockName,"".getBytes(), ZooDefs.Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT); } } catch (Exception e) { e.printStackTrace(); } } /** * 获取锁 * @return * @throws InterruptedException */ public void lock(){ String path = null; ensureRootPath(); try { path = zk.create(lockName+"/mylock_", "".getBytes(), ZooDefs.Ids.OPEN_ACL_UNSAFE, CreateMode.EPHEMERAL_SEQUENTIAL); lockZnode = path; List<String> minPath = zk.getChildren(lockName,false); System.out.println(minPath); Collections.sort(minPath); System.out.println(minPath.get(0)+" and path "+path); if (!Strings.nullToEmpty(path).trim().isEmpty()&&!Strings.nullToEmpty(minPath.get(0)).trim().isEmpty()&&path.equals(lockName+"/"+minPath.get(0))) { System.out.println(Thread.currentThread().getName() + " get Lock..."); return; } String watchNode = null; for (int i=minPath.size()-1;i>=0;i--){ if(minPath.get(i).compareTo(path.substring(path.lastIndexOf("/") + 1))<0){ watchNode = minPath.get(i); break; } } if (watchNode!=null){ final String watchNodeTmp = watchNode; final Thread thread = Thread.currentThread(); Stat stat = zk.exists(lockName + "/" + watchNodeTmp,new Watcher() { @Override public void process(WatchedEvent watchedEvent) { if(watchedEvent.getType() == Event.EventType.NodeDeleted){ thread.interrupt(); } try { zk.exists(lockName + "/" + watchNodeTmp,true); } catch (KeeperException e) { e.printStackTrace(); } catch (InterruptedException e) { e.printStackTrace(); } } }); if(stat != null){ System.out.println("Thread " + Thread.currentThread().getId() + " waiting for " + lockName + "/" + watchNode); } } try { Thread.sleep(1000000000); }catch (InterruptedException ex){ System.out.println(Thread.currentThread().getName() + " notify"); System.out.println(Thread.currentThread().getName() + " get Lock..."); return; } } catch (Exception e) { e.printStackTrace(); } } /** * 释放锁 */ public void unlock(){ try { System.out.println(Thread.currentThread().getName() + "release Lock..."); zk.delete(lockZnode,-1); } catch (InterruptedException e) { e.printStackTrace(); } catch (KeeperException e) { e.printStackTrace(); } } public static void main(String args[]) throws InterruptedException { ExecutorService service = Executors.newFixedThreadPool(10); for (int i = 0;i<4;i++){ service.execute(()-> { FairLockTest test = new FairLockTest(); try { test.lock(); Thread.sleep(3000); } catch (InterruptedException e) { e.printStackTrace(); } test.unlock(); }); } service.shutdown(); } }
同样上面的程序也有几点需要注意:
- Zookeeper的API没有提供直接的获取上一个节点或者最小节点的API需要我们自己实现。
- 使用了interrupt做线程的唤醒,这样不科学,因为不想将JVM的lock引进来所以没有用countdownlatch来做流程控制。
- Watch也是要重新设置的,这里使用了Watch的复用,所以代码简单些。
其实上面的实现还是很复杂的,因为你需要反复的去关注Watcher,实现一个Demo可以,做一个生产环境可用的Lock并不容易。因为你的代码bug在生产环境上会引起很严重的bug。
其实对于Zookeeper的一些常用功能是有一些成熟的包实现的,像Curator。Curator的确是足够牛逼,不仅封装了Zookeeper的常用API,也包装了很多常用Case的实现。但是它的编程风格其实还是吧比较难以接受的。
可以用Curator轻易的实现一个分布式锁:
InterProcessMutex lock = new InterProcessMutex(client, lockPath); if ( lock.acquire(maxWait, waitUnit) ) { try { // do some work inside of the critical section here } finally { lock.release(); } }
是的就这么简单,一个直接拿过来可用的轮子。
基于Zookeeper的分布式锁就说完了。基于Zookeeper实现分布式锁,其实是不常用的。虽然它实现锁十分优雅,但编程复杂,同时还要单独维护一套Zookeeper集群,频繁的Watch对Zookeeper集群的压力还是蛮大的,如果不是原有的项目以来Zookeeper,同时锁的量级比较小的话,还是不用为妙。
作者:一只小哈
链接:https://www.jianshu.com/p/5d12a01018e1
来源:简书
来源:https://www.cnblogs.com/leeego-123/p/12162220.html