hadoop的HA机制+zookeeper

 关于hadoop的HA配置以及wordcount测试

一，简单环境配置

1，查看centos版本位数：

$>getconf LONG_BIT,

2，桌面模式和文本模式之间进行切换：

1），在终端命令行进行设置时只能暂时改变模式，

     $>init 3    表示切换到文本模式

     $>init 5    表示切换到桌面模式

2），永久改变模式需要修改配置文件，进入到etc目录下

      $>sudo nano inittab   修改该文件最后一行

     若需要文本模式则改为 id:3:initdefault

     若需要桌面模式则改为：id:5:initdefault

注意：改为桌面模式时最好加大内存，改为4G左右

3，配置静态ip,修改主机名，配置hosts文件

    1），进入桌面模式下进行配置，

2），查看配置的静态ip

     cd /etc/sysconfig/networking/devices/ifcfg-eth2

3）,修改主机名

          nano /etc/sysconfig/network  本机的主机名

4，域名映射的配置：

      nano /etc/hosts  集群中的主机域名映射表

5，检验配置是否生效

ping hostname观察网络配置是否生效

6，关闭防火墙

          1）查看防火墙状态

                   service iptables status

          2）关闭防火墙

             service iptables stop

          3）查看防火墙开机启动状态

                   chkconfig iptables --list

          4）关闭防火墙开机启动

                   chkconfig iptables off

7，配置ssh免登陆

           1）生成ssh免登陆密钥

              进入到我的home目录

                 cd ~/.ssh

ssh-keygen -t rsa （四个回车）

           2）执行完这个命令后，会生成两个文件id_rsa（私钥）、id_rsa.pub（公钥）

              将公钥拷贝到要免登陆的机器上

                 ssh-copy-id localhost

二，配置HA

1,让两个NN节点在某个时间只能有一个节点正常响应客户端请求，响应请求的必须为ACTIVE状态的那一台

2，standby状态的节点（第二个namenode）必须能够无缝的切换为ACTIVE状态，意味着两个NN必须时时刻刻保持一致

3，管理edits，写了一个分布式应用qjournal:它依赖于zookeeper实现

4，如何避免状态切换时发生brain split现象

   fencing: ssh发送kill指令

   执行自定义shell脚本

5，部署方式：

3台机器：

第一台                 第二台                              第三台

nn1                    nn2

zkfc1                  zkfc2                                   dn1                                                                   

zk1,                   zk2,                                    zk3

jn1,                   jn2,                                     jn3

RM                     RM                                    NM

7台机器：

1、        namenode                    zkfc

2,         namenode                    zkfc

3,         resourcemanager

4,         resourcemanager

5,         zookeeper  journalnode  datanode  nodemanager

6,         zookeeper  journalnode  datanode  nodemanager

7,         zookeeper  journalnode  datanode  nodemanager

hadoop2.0已经发布了稳定版本了，增加了很多特性，比如HDFS HA、YARN等。最新的hadoop-2.4.1又增加了YARN HA

前期准备就不详细说了，我下面的HA配置选择的是3个节点

1.修改Linux主机名

2.修改IP

3.修改主机名和IP的映射关系

4.关闭防火墙

5.ssh免登陆

6.安装JDK，配置环境变量等

集群规划：

         主机名              IP                       安装的软件                                          运行的进程

         s104               192.168.43.104  jdk、hadoop,zookeeper1            NN1,zkfc1,RM1,QPM,JN1 

         s106               192.168.43.106     jdk、hadoop,zookeeper2      NN2,zkfc2,RM2,QPM,JN2 

         s108               192.168.43.108     jdk、hadoop、zookeeper3        DN,NM,QPM,JN3     

说明：

1.在hadoop2.0中通常由两个NameNode组成，一个处于active状态，另一个处于standby状态。Active NameNode对外提供服务，而Standby NameNode则不对外提供服务，仅同步active namenode的状态，以便能够在它失败时快速进行切换。

         hadoop2.0官方提供了两种HDFS HA的解决方案，一种是NFS，另一种是QJM。这里我们使用简单的QJM。在该方案中，主备NameNode之间通过一组JournalNode同步元数据信息，一条数据只要成功写入多数JournalNode即认为写入成功。通常配置奇数个JournalNode

         这里还配置了一个zookeeper集群，用于ZKFC（DFSZKFailoverController）故障转移，当Active NameNode挂掉了，会自动切换Standby NameNode为standby状态

2.hadoop-2.2.0中依然存在一个问题，就是ResourceManager只有一个，存在单点故障，hadoop-2.4.1解决了这个问题，有两个ResourceManager，一个是Active，一个是Standby，状态由zookeeper进行协调

安装步骤：

1.安装配置zooekeeper集群（在s104上）

1).上传zk安装包到~/目录下

2).解压 tar -zxvf zookeeper-3.4.5.tar.gz -C app

3),进入zookeeper-3.4.5，然后进入到文件夹conf进行配置（先在一台节点上配置）

3.1添加一个zoo.cfg配置文件

                  $>zookeeper-3.4.5/conf

                  $>mv zoo_sample.cfg zoo.cfg

3.2修改配置文件（zoo.cfg）

                        dataDir=/home/hadoop/app/zookeeper-3.4.5/data

                         server.1=itcast05:2888:3888

                         server.2=itcast06:2888:3888

                         server.3=itcast07:2888:3888

3.3在（dataDir=/home/hadoop/app/zookeeper-3.4.5/data）创建一个myid文件，里面内容是server.N中的N（server.2里面内容为2）

                       echo "1" > myid

3.4将配置好的zk拷贝到其他节点

                       scp -r /home/hadoop/app/zookeeper-3.4.5/ s106:/home/hadoop/app

                       scp -r /home/hadoop/app/zookeeper-3.4.5/ s108:/home/hadoop/app

3.5注意：在其他节点上一定要修改myid的内容

                      在s106应该讲myid的内容改为2 （echo "2" > myid）

                      在s108应该讲myid的内容改为3 （echo "3" > myid）

4).启动集群(后面配置完成进行集群测试时再关闭，然后按步骤启动)

                分别启动zk,进入到zookeeper的bin目录下

       ./zkServer.sh start

           默认端口号查看：netstat -nltp | grep 2181

           查看运行状态： ./zkServer.sh status

5),设置成功后，进入命令行客户端，连接到集群访问数据，做测试

           cd zookeeper-3.4.5/bin          ./zkCli.sh

6)，zookeeper管理客户存放的数据采用的是类似于文件树的结构，每一个节点叫做node。

           注意：zookeeper集群能够开启，开启的节点数必须超过配置的一半    

2.安装配置hadoop集群（在s104上操作）

2.1解压

                            tar -zxvf hadoop-2.4.1.tar.gz -C app

2.2配置HDFS（hadoop2.0所有的配置文件都在$HADOOP_HOME/etc/hadoop目录下）

                     #将hadoop添加到环境变量中

                            vim /etc/profile

                            export JAVA_HOME=/home/hadoop/app/jdk1.7.0_65

                            export HADOOP_HOME=/home/hadoop/app/hadoop-2.4.1

                            export PATH=$PATH:$JAVA_HOME/bin:$HADOOP_HOME/bin

                     #hadoop2.0的配置文件全部在$HADOOP_HOME/etc/hadoop下

                            cd /home/hadoop/app/hadoop-2.4.1/etc/hadoop

2.2.1修改hadoop-env.sh

                                     export JAVA_HOME=/home/hadoop/app/jdk1.7.0_65

2.2.2修改core-site.xml

<configuration>

                             <!-- 指定hdfs的nameservice为ns1 -->

                                               <property>

                                                        <name>fs.defaultFS</name>

                                                        <value>hdfs://ns1/</value>

                                               </property>

                                               <!-- 指定hadoop临时目录 -->

                                               <property>

                                                        <name>hadoop.tmp.dir</name>

                                                        <value>/home/hadoop/app/hadoop-2.4.1/tmp</value>

                                               </property>

                                               <!-- 指定zookeeper地址 -->

                                               <property>

                                                        <name>ha.zookeeper.quorum</name>

                                                        <value>s104:2181,s106:2181,s108:2181</value>

                                               </property>

                                     </configuration>

2.2.3修改hdfs-site.xml

                                     <configuration>

                                     <!--指定hdfs的nameservice为ns1，需要和core-site.xml中的保持一致 -->

                                               <property>

                                                        <name>dfs.nameservices</name>

                                                        <value>ns1</value>

                                               </property>

                                     <!-- ns1下面有两个NameNode，分别是nn1，nn2 -->

                                               <property>

                                                        <name>dfs.ha.namenodes.ns1</name>

                                                        <value>nn1,nn2</value>

                                               </property>

                                               <!-- nn1的RPC通信地址 -->

                                               <property>

                                                        <name>dfs.namenode.rpc-address.ns1.nn1</name>

                                                        <value>s104:9000</value>

                                               </property>

                                               <!-- nn1的http通信地址 -->

                                               <property>

                                                        <name>dfs.namenode.http-address.ns1.nn1</name>

                                                        <value>s104:50070</value>

                                               </property>

                                              <!-- nn2的RPC通信地址 -->

                                               <property>

                                                        <name>dfs.namenode.rpc-address.ns1.nn2</name>

                                                        <value>s106:9000</value>

                                               </property>

                                               <!-- nn2的http通信地址 -->

                                               <property>

                                                        <name>dfs.namenode.http-address.ns1.nn2</name>

                                                        <value>s106:50070</value>

                                               </property>

                                               <!-- 指定NameNode的元数据在JournalNode上的存放位置 -->

                                               <property>

                                                        <name>dfs.namenode.shared.edits.dir</name>

                                                        <value>qjournal://s104:8485;s106:8485;s108:8485/ns1</value>

                                               </property>

                                               <!-- 指定JournalNode在本地磁盘存放数据的位置 -->

                                               <property>

                                                        <name>dfs.journalnode.edits.dir</name>

                                                        <value>/home/hadoop/app/hadoop-2.4.1/journaldata</value>

                                               </property>

                                               <!-- 开启NameNode失败自动切换 -->

                                               <property>

                                                        <name>dfs.ha.automatic-failover.enabled</name>

                                                        <value>true</value>

                                               </property>

                                               <!-- 配置失败自动切换实现方式 -->

                                               <property>

<name>dfs.client.failover.proxy.provider.ns1</name>

<value>org.apache.hadoop.hdfs.server.namenode.ha.ConfiguredFailoverProxyProvider</value>

</property>

<!-- 配置隔离机制方法，多个机制用换行分割，即每个机制暂用一行-->

                                               <property>

                                                        <name>dfs.ha.fencing.methods</name>

                                                        <value>

                                                                 sshfence

                                                                 shell(/bin/true)

                                                        </value>

                                               </property>

         <!-- 使用sshfence隔离机制时需要ssh免登陆 -->

                                               <property>

                                                        <name>dfs.ha.fencing.ssh.private-key-files</name>

                                                        <value>/home/hadoop/.ssh/id_rsa</value>

                                               </property>

                                               <!-- 配置sshfence隔离机制超时时间 -->

                                               <property>

                                                        <name>dfs.ha.fencing.ssh.connect-timeout</name>

                                                        <value>30000</value>

                                               </property>

                                     </configuration>

2.2.4修改mapred-site.xml

                                     <configuration>

                                               <!-- 指定mr框架为yarn方式 -->

                                               <property>

                                                        <name>mapreduce.framework.name</name>

                                                        <value>yarn</value>

                                               </property>

                                     </configuration>     

2.2.5修改yarn-site.xml

                                     <configuration>

                                                        <!-- 开启RM高可用 -->

                                                        <property>

                                                           <name>yarn.resourcemanager.ha.enabled</name>

                                                           <value>true</value>

                                                        </property>

                                                        <!-- 指定RM的cluster id -->

                                                        <property>

                                                           <name>yarn.resourcemanager.cluster-id</name>

                                                           <value>yrc</value>

                                                        </property>

                                                        <!-- 指定RM的名字 -->

                                                        <property>

                                                           <name>yarn.resourcemanager.ha.rm-ids</name>

                                                           <value>rm1,rm2</value>

                                                        </property>

                                                        <!-- 分别指定RM的地址 -->

                                                        <property>

                                                           <name>yarn.resourcemanager.hostname.rm1</name>

                                                           <value>s104</value>

                                                        </property>

                                                        <property>

                                                           <name>yarn.resourcemanager.hostname.rm2</name>

                                                           <value>s106</value>

                                                        </property>

                                                        <!-- 指定zk集群地址 -->

                                                        <property>

                                                           <name>yarn.resourcemanager.zk-address</name>

                                                           <value>s104:2181,s106:2181,s108:2181</value>

                                                        </property>

                                                        <property>

                                                           <name>yarn.nodemanager.aux-services</name>

                                                           <value>mapreduce_shuffle</value>

                                                        </property>

                                     </configuration>     

2.2.6修改slaves(slaves是指定子节点的位置，因为要在s104上启动HDFS、在s104又启动yarn，所以s104上的slaves文件既指定的是datanode的位置，又指定的是nodemanager的位置)

                                     s108

2.2.7配置免密码登陆

                   #首先要配置s104到s106,s108的免密码登陆

                   #在s104上生产一对钥匙

                      ssh-keygen -t rsa

                   #将公钥拷贝到其他节点，包括自己

                      ssh-coyp-id s104

                     ssh-coyp-id s106

                      ssh-coyp-id s108

                   #配置s106到s104,s108的免密码登陆(两个namenode之间要配置ssh免密码登陆，别忘了配置s106到s104的免登陆)

                   #在s106上生产一对钥匙

                      ssh-keygen -t rsa

                   #将公钥拷贝到其他节点

                      ssh-coyp-id s104

                      ssh-coyp-id s106

                      ssh-coyp-id s108                                    

2.4将配置好的hadoop拷贝到其他节点

                            scp -r /home/hadoop/app/hadoop-2.4.1/ hadoop@s106:/home/hadoop/app/

                            scp -r /home/hadoop/app/hadoop-2.4.1/ hadoop@s108:/home/hadoop/app/

###注意：严格按照下面的步骤

2.5启动zookeeper集群（分别在s104,s105,s106上启动zk）

                            cd /weekend/zookeeper-3.4.5/bin/

                            ./zkServer.sh start

                            #查看状态：一个leader，两个follower

                            ./zkServer.sh status        

2.6启动journalnode（分别在在s104,s105,s106上执行）

                            cd /home/hadoop/app/hadoop-2.4.1

                            sbin/hadoop-daemon.sh start journalnode

                            #运行jps命令检验，s104,s105,s106上多了JournalNode进程

2.7格式化HDFS

                            #在s104上执行命令:

                            hdfs namenode -format

                          #格式化后会在根据core-site.xml中的hadoop.tmp.dir配置生成个文件，这里我配置的是/home/hadoop/app/hadoop-2.4.1/tmp，将其拷贝到s106 的/home/hadoop/app/hadoop-2.4.1/下。

                           scp -r tmp/ s106:/home/hadoop/app/hadoop-2.4.1/

                            ##也可以这样，建议hdfs namenode –bootstrapStandby

2.8格式化ZKFC(在s104上执行即可)

                            hdfs zkfc –formatZK

2.9启动HDFS(在s104上执行)

                            sbin/start-dfs.sh     

2.10，我配置的是3个节点的HA配置，所以namenode和resourcemanager放在同一个节点上，我配置的两个resourcemanager放在s104,s106

          在s104上启动 sbin/start-yarn.sh

              到此，hadoop-2.4.1配置完毕，可以统计浏览器访问:

                       http://192.168.43.104:50070

                       NameNode ‘s104:9000‘ (active)

                       http://192.168.43.106:50070

                       NameNode ‘s106:9000‘ (standby)

         验证YARN：

                   运行一下hadoop提供的demo中的WordCount程序：

                   hadoop jar share/hadoop/mapreduce/hadoop-mapreduce-examples-2.4.1.jar wordcount /user/input /user/output

          OK，大功告成！！！     

测试集群工作状态的一些指令：

bin/hdfs dfsadmin -report        查看hdfs的各节点状态信息

bin/hdfs haadmin -getServiceState nn1           获取一个namenode节点的HA状态

sbin/hadoop-daemon.sh start namenode  单独启动一个namenode进程

./hadoop-daemon.sh start zkfc   单独启动一个zkfc进程

HA配置成功，实验测试过程如下：

HA架构：

 S104节点

S106节点

S108节点

开始active节点为s104

当s104被我们kill掉时，active节点自动变为s106,如下图

手动启动那个挂掉的NameNode

                   sbin/hadoop-daemon.sh start namenode

                   通过浏览器访问：http://192.168.43.104:50070，如下图，该节点变为standby

         datanode节点一直是s108         

验证yarn, 运行一下hadoop提供的demo中的WordCount程序：

hadoop jar share/hadoop/mapreduce/hadoop-mapreduce-examples-2.4.1.jar wordcount  /profile  /out

         OK，大功告成！！！

最后wordcount的结果如下图所示

网页中生成的结果如图所示

原文：http://www.cnblogs.com/zjf-293916/p/6823341.html