(相关资料图)

背景

说起master选举，最开始想到的可能就是zookeeper，但有些场景zookeeper的使用过于繁重和复杂，又由于etcd是基于Raft的分布式K/V存储，强一致性的K/V读写是核心。所以造就了etcd可以用于master的选举的场景。

原理

etcd clientv3 concurrency对选举进行了封装

import github.com/coreos/etcd/clientv3/concurrency

上篇文章介绍了etcd使用txn实现分布式锁，而节点选举也要依靠Txn进行创建key的CAS操作zookeeper是利用创建临时有序节点的方式，etcd也同样的是在prefix path下创建相应的key,并且key的revision也是有序。同样也是通过watch机制进行通知。

应用

//封装type etcdLeader struct {sess *concurrency.Session    //会话 客户端的租约会话，会做KeepAliveelec *concurrency.Election    //选举}func Campaign(ctx context.Context, client *clientv3.Client, prefix, value string) (*etcdLeader, error) {if value == "" {value = fmt.Sprintf("%s-%d", MyHostName, timestampMs())}s, err := concurrency.NewSession(client)if err != nil {return nil, fmt.Errorf("failed to generate session: %v", err)}prefix = "/openresty/" + "-concurrency/" + strings.TrimPrefix(prefix, "/")elec := concurrency.NewElection(s, prefix)return &etcdLeader{sess: s,elec: elec,}, elec.Campaign(ctx, value) // blocked until elected}//elec.Campaign 会参加选举，直到它被选中、发生错误或上下文被取消// 调用elec.Proclaim 常用来检测当前主节点是否生效func (l *etcdLeader) Proclaim(ctx context.Context, value string) error {if l.elec == nil {return fmt.Errorf("already closed")}if value == "" {value = fmt.Sprintf("%s-%d", MyHostName, timestampMs())}return l.elec.Proclaim(ctx, value)}//调用elec.Resign 退出masterfunc (l *etcdLeader) Resign(ctx context.Context) error {if l.elec == nil {return nil}defer l.sess.Close()defer func() {l.elec = nill.sess = nil}()return l.elec.Resign(ctx)}

测试

//模拟进程处理func process(processID string) {cli, err := getEtcdCli("open")  //获得cliif err != nil {fmt.Println(err)panic(processID)}ctx := context.Background()maintainer, err := Campaign(context.Background(), cli, "process", "")  //会阻塞，直到该节点选为masterif err != nil {fmt.Println("process " + processID + " Campaign err" + err.Error())return}fmt.Println(processID + "is maintainer")ticker := time.NewTicker(time.Second)count := 0defer ticker.Stop()for {select {case <-ctx.Done():err = maintainer.Resign(context.Background())  //超时退出if err != nil {fmt.Printf("error occured when resign lfe_sync: %v", err)}returncase <-ticker.C:err = maintainer.Proclaim(ctx, "") if err != nil {fmt.Printf("error occured when proclaim lfe_sync: %v", err)maintainer.Resign(ctx)return}count++//do sthfmt.Println("processID ", processID+" ", count)if count == 5 {      //主动退出err = maintainer.Resign(context.Background())if err != nil {fmt.Printf("error occured when resign lfe_sync: %v", err)}return}}}}

总结与思考

如果从Campaign源码可以看出，比起上一篇利用txn实现分布式锁的实现，选主增加watch机制，减少轮训获取锁的过程。在本场景中选主和分布式锁的使用并没有什么差别，只是选主减少轮训，性能更好。本次代码地址：https://github.com/zhaoshoucheng/hodgepodge/blob/main/etcd/master.go