hdfs的HA以及Yarn的HA高可用

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tHA(High Available), 高可用，是保证业务连续性的有效解决方案，一般有两个或两个以上的节点，分为活动节点（Active）及备用节点（Standby）。通常把正在执行业务的称为活动节点，而作为活动节点的一个备份的则称为备用节点。当活动节点出现问题，导致正在运行的业务（任务）不能正常运行时，备用节点此时就会侦测到，并立即接续活动节点来执行业务。从而实现业务的不中断或短暂中断。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tHadoop1.X版本，NN是HDFS集群的单点故障点，每一个集群只有一个NN,如果这个机器或进程不可用，整个集群就无法使用。为了解决这个问题，出现了一堆针对HDFS HA的解决方案（如：Linux HA, VMware FT, shared NAS+NFS, BookKeeper, ","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"QJM/Quorum Journal Manager","attrs":{}},{"type":"text","text":", BackupNode等）。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t在HA具体实现方法不同情况下，HA框架的流程是一致的, 不一致的就是如何存储、管理、同步edits编辑日志文件。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t在Active NN和Standby NN之间要有个共享的存储日志的地方，Active NN把edit Log写到这个共享的存储日志的地方，Standby NN去读取日志然后执行，这样Active和Standby NN内存中的HDFS元数据保持着同步。一旦发生主从切换Standby NN可以尽快接管Active NN的工作。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"9.1 Namenode HA介绍","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\thadoop2.x之后，Clouera提出了QJM/Qurom Journal Manager，这是一个基于Paxos算法（分布式一致性算法）实现的HDFS HA方案，它给出了一种较好的解决思路和方案,QJM主要优势如下：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t不需要配置额外的高共享存储，降低了复杂度和维护成本。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t消除spof(单点故障)。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t系统鲁棒性(Robust)的程度可配置、可扩展。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t基本原理就是用2N+1台 JournalNode 存储EditLog，每次写数据操作有>=N+1返回成功时即认为该次写成功，数据不会丢失了。当然这个算法所能容忍的是最多有N台机器挂掉，如果多于N台挂掉，这个算法就失效了。这个原理是基于Paxos算法。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t在HA架构里面SecondaryNameNode已经不存在了，为了保持standby NN时时的与Active NN的元数据保持一致，他们之间交互通过JournalNode进行操作同步。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t任何修改操作在 Active NN上执行时，JournalNode进程同时也会记录修改log到至少半数以上的JN中，这时 Standby NN 监测到JN 里面的同步log发生变化了会读取 JN 里面的修改log，然后同步到自己的目录镜像树里面，如下图：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t当发生故障时，Active的 NN 挂掉后，Standby NN 会在它成为Active NN 前，读取所有的JN里面的修改日志，这样就能高可靠的保证与挂掉的NN的目录镜像树一致，然后无缝的接替它的职责，维护来自客户端请求，从而达到一个高可用的目的。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t在HA模式下，datanode需要确保同一时间有且只有一个NN能命令DN。为此：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t每个NN改变状态的时候，向DN发送自己的状态和一个序列号。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tDN在运行过程中维护此序列号，当failover时，新的NN在返回DN心跳时会返回自己的active状态和一个更大的序列号。DN接收到这个返回则认为该NN为新的active。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t如果这时原来的active NN恢复，返回给DN的心跳信息包含active状态和原来的序列号，这时DN就会拒绝这个NN的命令。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Failover Controller :","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tHA模式下，会将FailoverController部署在每个NameNode的节点上，作为一个单独的进程用来监视NN的健康状态。","attrs":{}},{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"FailoverController主要包括三个组件:","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tHealthMonitor: 监控NameNode是否处于unavailable或unhealthy状态。当前通过RPC调用NN相应的方法完成。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tActiveStandbyElector: 监控NN在ZK中的状态。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\tZKFailoverController: 订阅HealthMonitor 和ActiveStandbyElector 的事件，并管理NN的状态,另外zkfc还负责解决fencing（也就是脑裂问题）。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"​\t上述三个组件都在跑在一个JVM中，这个JVM与NN的JVM在同一个机器上。但是两个独立的进程。一个典型的HA集群，有两个NN组成，每个NN都有自己的ZKFC进程。","attrs":{}}]}]}