分布式选主 -- 利用Mysql ACID和Lease协议实现选主和高可用
在实际生产开发中,遇到一些多节点共存,需要选主,并且要实现HA自动容错的场景,思考了写方法拿出来和大家分享一下。
- Lease协议,Mysql ACID
- 高可用选主方案设计
- 适用场景
- Java语言实现描述
- 进一步优化
系统中有很多应用场景要类似主从架构,主服务器(Master)对外提供服务,从服务器(Salve)热备份,不提供服务但随时活着,如果Master出现宕机或者网络问题,Slave即可接替Master对外服务,并由Slave提升为Master(新主)。典型的多节点共存, 但只能同时存在一个主,并且所有节点的状态能统一维护。
大家一定首先想到了著名的Paxos算法( http://baike.baidu.com/view/8438269.htm)。简单的说,Paxos通过每个节点的投票算法,来决议一个事情,当多余1/2个节点都投票通过时,Paxos产生一个唯一结果的决议,并通知各个节点维护这个信息。例如Paxos的选主,首先产生一个关于某个节点希望当Master的投票,然后各个节点给出反馈,最终Paxos集群维护唯一的Master的结论。Zookeeper就是Paxos的一种实现。这种场景最适合用zookeeper来选主, 但zookeeper有个明显的缺点,当存活的节点小于zookeeper集群的1/2时,就不能工作了。比如zk有10各节点,那么必须满足可用的节点大于5才可。
在实际环境中,如果对Master要求不是那么严格的话,可以通过某些改进和取舍来达到目的。比如可能在秒级别允许Master暂时不能访问、选主时间内可能存在一定的冲突但通过再次选主即可。本人设计了一个简易的利用Mysql一致性和简易版Lease来workaround。
Mysql ACID保证了一条数据记录的一致性、完整性,不会出现多进程读写的一致性问题和唯一正确性。Lease协议(协议细节可以Google之)通过向Master发送一个lease(租期)包,Master在这个lease期之内充当主角色,如果lease期到了则再次去申请lease,如果lease期到了,但是网络除了问题,这时Master可以i主动下线,让其他节点去竞选Master。举个例子,三个节点A、B、C经过第一轮选主之后,A成为Master,它获得了10秒的lease,当前时间假设是00:00:00,那么它Master地位可以用到00:00:10,当时间到达00:00:10时,A、B、C会重新进行Master选举,每个节点都有可能成为Master(从工程的角度触发,A继续为Master的概率更大),如果这时候A的网络断了,不能联通B、C的集群了, 那么A会自动下线,不会去竞争,这样就不会出现“脑裂”的现象。
---------------------------------------------- 华丽的分割线 ----------------------------------------------
设计方案如下:(server代表集群中的一台机器,也可看作一个进程,server之间是平等的)
- 各个server之间用ntpserver时间同步(保证服务器之间秒级同步即可)
- 各个server持有一个唯一ID号(ip+进程号),通过此id唯一标识一个server实例
- 各个server定义一个lease租期,单位为秒
- Mysql唯一表唯一一条记录维护全局Master的信息,ACID保证一致性
- Master Server每半个lease期向Mysql更新如上的唯一一条记录,并更新心跳,维护Master状态
- Slaver Server每半个lease周期从mysql获取Master Server信息,如果数据库中Master的Lease超过了当前时间(heartbeat_time+ lease > current_time),则申请当Master。
这其中比较棘手的问题是:
1、由于数据库访问和休眠的时间(lease的一半),有时延的存在,要处理Mysql异常、网络异常。
2、可能存在同时抢占Master的server,这个时候就需要一个验证机制保证为抢到Master的server自动退位为Slaver
下面给出图实例 :(10.0.0.1为Master)
10.0.0.1 crash了。mysql中维护的10.0.0.1的主信息已过期,其他节点去抢占
各个节点再次读取数据库,查看是否是自己抢占成功了:
之后,10.0.0.3作为Master对外服务。此时如果10.0.0.1重启,可作为Slaver。如果10.0.0.1因为网络分化或者网络异常而不能维护心跳,则在超过自身lease时自动停止服务,不会出现“双Master”的现象。
每个Server遵循如下流程:
数据库设计:
某一时刻,数据库中Master的信息:
当前时间: 45分15秒
当前Master Lease :6秒
当前Master Lease可用到: 45分21秒
---------------------------------------------- 华丽的分割线 ----------------------------------------------
3、适用的场景
一、生命周期内可使用Mysql、并且各个server之间时间同步。
二、需要集群中选出唯一主对外提供服务,其他节点作为slaver做standby,主lease过期时竞争为Master
三、对比zookeeper,可满足如果集群挂掉一半节点,也可正常工作的情况,比如只有一主一备。
四、允许选主操作在秒级容错的系统,选主的时候可能有lease/2秒的时间窗口,此时服务可能不可用。
五、允许lease/2秒内出现极限双Master情况,但是概率很小。
---------------------------------------------- 华丽的分割线 ----------------------------------------------
4、Java语言实现描述
一些配置信息和时间相关、休眠周期相关的时间变量
final long interval = lease / intervalDivisor; long waitForLeaseChallenging = 0L; lease = lease / 1000L; long challengeFailTimes = 0L; long takeRest = 0L; long dbExceptionTimes = 0L; long offlineTime = 0L; Random rand = new Random(); Status stateMechine = Status.START; long activeNodeLease = 0L; long activeNodeTimeStamp = 0L;
数据库异常的处理:
KeepAlive keepaliveNode = null; try { /* first of all get it from mysql */ keepaliveNode = dbService.accquireAliveNode(); if (stateMechine != Status.START && keepaliveNode==null) throw new Exception(); // recount , avoid network shake dbExceptionTimes = 0L; } catch (Exception e) { log.fatal("[Scanner] Database Exception with times : " + dbExceptionTimes++); if (stateMechine == Status.OFFLINE) { log.warn("[Scanner] Database Exception , OFFLINE "); } else if (dbExceptionTimes >= 3) { log.fatal("[Scanner] Database Exception , Node Offline Mode Active , uniqueid : " + uniqueID); stateMechine = Status.OFFLINE; dbExceptionTimes = 0L; offlineTime = System.currentTimeMillis(); online = false; } else continue; }
总的循环和状态机的变迁:
while (true) { SqlSession session = dbConnecction.openSession(); ActionScanMapper dbService = session.getMapper(ActionScanMapper.class); KeepAlive keepaliveNode = null; try { /* first of all get it from mysql */ keepaliveNode = dbService.accquireAliveNode(); if (stateMechine != Status.START && keepaliveNode==null) throw new Exception(); // recount , avoid network shake dbExceptionTimes = 0L; } catch (Exception e) { log.fatal("[Scanner] Database Exception with times : " + dbExceptionTimes++); if (stateMechine == Status.OFFLINE) { log.warn("[Scanner] Database Exception , OFFLINE "); } else if (dbExceptionTimes >= 3) { log.fatal("[Scanner] Database Exception , Node Offline Mode Active , uniqueid : " + uniqueID); stateMechine = Status.OFFLINE; dbExceptionTimes = 0L; offlineTime = System.currentTimeMillis(); online = false; } else continue; } try { activeNodeLease = keepaliveNode!=null ? keepaliveNode.getLease() : activeNodeLease; activeNodeTimeStamp = keepaliveNode!=null ? keepaliveNode.getTimestamp() : activeNodeTimeStamp; takeRest = interval; switch (stateMechine) { case START: if (keepaliveNode == null) { log.fatal("[START] Accquire node is null , ignore "); // if no node register here , we challenge it stateMechine = Status.CHALLENGE_REGISTER; takeRest = 0; } else { // check the lease , wether myself or others if (activeNodeLease < timestampGap(activeNodeTimeStamp)) { log.warn("[START] Lease Timeout scanner for uniqueid : " + uniqueID + ", timeout : " + timestampGap(activeNodeTimeStamp)); if (keepaliveNode.getStatus().equals(STAT_CHALLENGE)) stateMechine = Status.HEARTBEAT; else { stateMechine = Status.CHALLENGE_MASTER; takeRest = 0; } } else if (keepaliveNode.getUniqueID().equals(uniqueID)) { // I'am restart log.info("[START] Restart Scanner for uniqueid : " + uniqueID + ", timeout : " + timestampGap(activeNodeTimeStamp)); stateMechine = Status.HEARTBEAT; } else { log.info("[START] Already Exist Keepalive Node with uniqueid : " + uniqueID); stateMechine = Status.HEARTBEAT; } } break; case HEARTBEAT: /* uniqueID == keepaliveNode.uniqueID */ if (keepaliveNode.getUniqueID().equals(uniqueID)) { if (activeNodeLease < timestampGap(activeNodeTimeStamp)) { // we should challenge now , without nessesary to checkout Status[CHALLENGE] log.warn("[HEARTBEAT] HEART BEAT Lease is timeout for uniqueid : " + uniqueID + ", time : " + timestampGap(activeNodeTimeStamp)); stateMechine = Status.CHALLENGE_MASTER; takeRest = 0; break; } else { // lease ok , just update mysql keepalive status dbService.updateAliveNode(keepaliveNode.setLease(lease)); online = true; log.info("[HEARTBEAT] update equaled keepalive node , uniqueid : " + uniqueID + ", lease : " + lease + "s, remain_usable : " + ((activeNodeTimeStamp * 1000L + lease * 1000L) - System.currentTimeMillis()) + " ms"); } } else { /* It's others , let's check lease */ if (activeNodeLease < timestampGap(activeNodeTimeStamp)) { if (keepaliveNode.getStatus().equals(STAT_CHALLENGE)) { waitForLeaseChallenging = (long) (activeNodeLease * awaitFactor); if ((waitForLeaseChallenging) < timestampGap(activeNodeTimeStamp)) { log.info("[HEARTBEAT] Lease Expired , Diff[" + timestampGap(activeNodeTimeStamp) + "] , Lease[" + activeNodeLease + "]"); stateMechine = Status.CHALLENGE_MASTER; takeRest = 0; } else { log.info("[HEARTBEAT] Other Node Challenging , We wait for a moment ..."); } } else { log.info("[HEARTBEAT] Lease Expired , Diff[" + timestampGap(activeNodeTimeStamp) + "] , lease[" + activeNodeLease + "]"); stateMechine = Status.CHALLENGE_MASTER; takeRest = 0; } } else { online = false; log.info("[HEARTBEAT] Exist Active Node On The Way with uniqueid : " + keepaliveNode.getUniqueID() + ", lease : " + keepaliveNode.getLease()); } } break; case CHALLENGE_MASTER: dbService.challengeAliveNode(new KeepAlive().setUniqueID(uniqueID).setLease(lease)); online = false; // wait for the expired node offline automatic // and others also have changce to challenge takeRest = activeNodeLease; stateMechine = Status.CHALLENGE_COMPLETE; log.info("[CHALLENGE_MASTER] Other Node is timeout[" + timestampGap(activeNodeTimeStamp) + "s] , I challenge with uniqueid : " + uniqueID + ", lease : " + lease + ", wait : " + lease); break; case CHALLENGE_REGISTER: dbService.registerNewNode(new KeepAlive().setUniqueID(uniqueID).setLease(lease)); online = false; // wait for the expired node offline automatic // and others also have changce to challenge takeRest = activeNodeLease; stateMechine = Status.CHALLENGE_COMPLETE; log.info("[CHALLENGE_REGISTER] Regiter Keepalive uniqueid : " + uniqueID + ", lease : " + lease); break; case CHALLENGE_COMPLETE : if (keepaliveNode.getUniqueID().equals(uniqueID)) { dbService.updateAliveNode(keepaliveNode.setLease(lease)); online = true; log.info("[CHALLENGE_COMPLETE] I Will be the Master uniqueid : " + uniqueID); // make the uptime correct stateMechine = Status.HEARTBEAT; } else { online = false; log.warn("[CHALLENGE_COMPLETE] So unlucky , Challenge Failed By Other Node with uniqueid : " + keepaliveNode.getUniqueID()); if (challengeFailTimes++ >= (rand.nextLong() % maxChallenge) + minChallenge) { // need't challenge anymore in a long time takeRest=maxChallengeAwaitInterval; stateMechine = Status.HEARTBEAT; challengeFailTimes = 0L; log.info("[CHALLENGE_COMPLETE] Challenge Try Times Used Up , let's take a long rest !"); } else { stateMechine = Status.HEARTBEAT; log.info("[CHALLENGE_COMPLETE] Challenge Times : " + challengeFailTimes + ", Never Give Up , to[" + stateMechine + "]"); } } break; case OFFLINE : log.fatal("[Scanner] Offline Mode Node with uniqueid : " + uniqueID); if (System.currentTimeMillis() - offlineTime >= maxOfflineFrozen) { // I am relive forcely log.info("[Scanner] I am relive to activie node , uniqueid : " + uniqueID); stateMechine = Status.HEARTBEAT; offlineTime = 0L; } else if (keepaliveNode != null) { // db is reconnected stateMechine = Status.HEARTBEAT; offlineTime = 0L; log.info("[Scanner] I am relive to activie node , uniqueid : " + uniqueID); } break; default : System.exit(0); } session.commit(); session.close(); if (takeRest != 0) Thread.sleep(takeRest); log.info("[Scanner] State Stage [" + stateMechine + "]"); } catch (InterruptedException e) { log.fatal("[System] Thread InterruptedException : " + e.getMessage()); } finally { log.info("[Scanner] UniqueID : " + uniqueID + ", Mode : " + (online?"online":"offline")); } } } enum Status { START, HEARTBEAT, CHALLENGE_MASTER, CHALLENGE_REGISTER, CHALLENGE_COMPLETE, OFFLINE }