假如现在有一个表，表里记录的是一些延迟消息，也就是在未来某时刻需要发送出去的消息，表的记录数不定，如果很多，那么需要多台机器来执行，如果很少，那么一两台就够了，这时候需要一个领导选举/任务分片的工作，下面是一个方案，使用curator来实现。
上代码：
import com.google.common.collect.Lists;
import org.apache.commons.lang3.StringUtils;
import org.apache.curator.framework.CuratorFramework;
import org.apache.curator.framework.CuratorFrameworkFactory;
import org.apache.curator.framework.recipes.cache.NodeCache;
import org.apache.curator.framework.recipes.cache.NodeCacheListener;
import org.apache.curator.framework.recipes.cache.PathChildrenCache;
import org.apache.curator.framework.recipes.cache.PathChildrenCacheEvent;
import org.apache.curator.framework.recipes.cache.PathChildrenCacheListener;
import org.apache.curator.framework.recipes.leader.LeaderLatch;
import org.apache.curator.framework.recipes.leader.LeaderLatchListener;
import org.apache.curator.retry.RetryNTimes;
import org.apache.zookeeper.CreateMode;
import java.util.List;
import java.util.UUID;
public class JobDistribute {
private static final String ZK_ADDRESS = "127.0.0.1:2181";
private static final String ZK_LEADER_PATH = "/task/exec/leader";
private static final String ZK_TASK_SPLIT_PATH = "/task/split/worker";
private static final String WORKERS_PARENT = "/task/split/";
private static List<String> allJobs = Lists.newArrayList("a","b","c","d","e","f","g","h");
public static void main(String[] args) throws Exception {
CuratorFramework client = CuratorFrameworkFactory
.newClient(ZK_ADDRESS,10000,5000,
new RetryNTimes(Integer.MAX_VALUE, 1000));
client.start();
String getJobPath = client.create().creatingParentsIfNeeded().withMode(CreateMode.EPHEMERAL_SEQUENTIAL).forPath(ZK_TASK_SPLIT_PATH);
//监听任务分配节点的值变化，根据值执行任务
final NodeCache nodeCache = new NodeCache(client, getJobPath, false);
nodeCache.getListenable().addListener(new NodeCacheListener() {
@Override
public void nodeChanged() throws Exception {
System.out.println("nodeChanged:next time do those job ==============="+nodeCache.getCurrentData());
//可以写入本地缓存，执行对应任务。。。。。
}
});
nodeCache.start();
System.out.println("getJobPath:"+getJobPath);
final String id = UUID.randomUUID().toString();
System.out.println("id:"+id);
LeaderLatch leaderLatch = new LeaderLatch(client, ZK_LEADER_PATH, id);
leaderLatch.addListener(new LeaderLatchListener() {
@Override
public void isLeader() {
System.out.println("Currently run as leader");
//成为leader之后要管理任务分片
try{
splitJobs(client);
} catch (Exception e){
e.printStackTrace();
}
try{
PathChildrenCache cache = new PathChildrenCache(client,
WORKERS_PARENT.substring(0,WORKERS_PARENT.length()-1), true);
cache.start();
cache.getListenable().addListener(new PathChildrenCacheListener(){
@Override
public void childEvent(CuratorFramework curatorFramework,
PathChildrenCacheEvent pathChildrenCacheEvent) throws Exception {
System.err.println("Event TYpe ：" + pathChildrenCacheEvent.getType());
if(pathChildrenCacheEvent.getType() == PathChildrenCacheEvent.Type.CHILD_ADDED ||
pathChildrenCacheEvent.getType() == PathChildrenCacheEvent.Type.CHILD_REMOVED){
//有新加入节点或者删除节点，重新分配任务
splitJobs(client);
}
}
});
} catch (Exception e){
e.printStackTrace();
}
}
@Override
public void notLeader() {
System.out.println("Currently run as slave");
}
});
leaderLatch.start();
Thread.sleep(Integer.MAX_VALUE);
}
//这里写了个简单粗暴的方法，其实平分任务方法很多，可以双重循环，外层循环数量大的，内层循环数量小的，进行分配
//如果希望一个任务尽量不被随意的转移，也可以有别的算法
//这里是8份任务，最多四台机器同时执行。
private static void splitJobs(CuratorFramework client) throws Exception {
List<String> workers = client.getChildren().forPath(WORKERS_PARENT.substring(0,WORKERS_PARENT.length()-1));
System.out.println("workers:"+workers);
if(workers.size()==1){
String path = WORKERS_PARENT+workers.get(0);
client.setData().forPath(path, StringUtils.join(allJobs,",").getBytes());
}else if(workers.size() ==2){
String path0 = WORKERS_PARENT+workers.get(0);
String path1 = WORKERS_PARENT+workers.get(1);
client.setData().forPath(path0, StringUtils.join(allJobs.subList(0,4),",").getBytes());
client.setData().forPath(path1, StringUtils.join(allJobs.subList(4,8),",").getBytes());
}else if(workers.size() == 3){
String path0 = WORKERS_PARENT+workers.get(0);
String path1 = WORKERS_PARENT+workers.get(1);
String path2 = WORKERS_PARENT+workers.get(2);
client.setData().forPath(path0, StringUtils.join(allJobs.subList(0,3),",").getBytes());
client.setData().forPath(path1, StringUtils.join(allJobs.subList(3,6),",").getBytes());
client.setData().forPath(path2, StringUtils.join(allJobs.subList(6,8),",").getBytes());
}else if(workers.size() == 4){
String path0 = WORKERS_PARENT+workers.get(0);
String path1 = WORKERS_PARENT+workers.get(1);
String path2 = WORKERS_PARENT+workers.get(2);
String path3 = WORKERS_PARENT+workers.get(3);
client.setData().forPath(path0, StringUtils.join(allJobs.subList(0,2),",").getBytes());
client.setData().forPath(path1, StringUtils.join(allJobs.subList(2,4),",").getBytes());
client.setData().forPath(path2, StringUtils.join(allJobs.subList(4,6),",").getBytes());
client.setData().forPath(path3, StringUtils.join(allJobs.subList(6,8),",").getBytes());
}
System.out.println("splitJobs ok");
}
}
可以使用zkCli去查看对应节点的值，可以强制leader退出，过几秒，就会选举新的leader，然后分配任务。
这只是一个简陋的雏形，想要实现高可用，还需要做很多工作。这是最近在做延迟任务时想到的。如果加上一个定时任务轮询任务表，在加上jdk delayqueue，那么，就能实现一个延迟消息系统，或者延迟任务系统。OK。
来源：oschina
链接：https://my.oschina.net/wuxiaofei/blog/3130670