HDFS-HA高可用 | Yarn-HA

 HDFS-HA

HA（High Available），即高可用（7*24小时不中断服务）

单点故障即有一台机器挂了导致全部都挂了；HA就是解决单点故障，就是针对NameNode；

主Active：读写、从standby只读；所依赖的服务都必须是高可用的；

两种解决共享空间的方案：NFS、QJM（主流的）

奇数台机器；QJM跟zookeeper（数据全局一致；半数以上的机器存活就可以提供服务）高可用的方式一模一样，；

QJM也是基于Paxos算法，系统容错不能超过n-1/2， 5台容错2台；

 

 这个架构只能手动决定哪个是active哪个是standby；必须只能有一个active！！如果出现两个NameNode，即两个都是active，它可能还不报错，那么可能会导致整个集群的数据都是错的，问题很严重！两个AA的情况叫脑裂（split brain缩写sb）。

standby要想变成active，要确保active那个，需要安全可靠的zookeeper（文件系统+通知机制）第三方来联系两方  ---> 实现故障的自动转移；

 由Zkfc来联系zookeeper，并不是namenode直接联系，（zookeeper客户端）；HA是hadoop2.0才有的，namenode在1.0时就有了；没有把zkfc写进namenode是为了保持NameNode的健壮性，没有zkfc之前就已经运行的很好了（鲁棒性）；NameNode和Zkfc虽然是两个进程但它们是绑定到一起的。

两个zkfc怎么决定谁初始化就是active呢，谁快谁就是active； 是active状态它会在zookeeper中有一个临时节点，zkfc会尝试看看zookeeper中有没有这个临时节点，如果没有我就变成这个临时节点，成为active，慢的一看有了，就变成standby；

NameNode发生假死，zkfc就会把zookeeper中的临时节点删除，去通知另外一个namenode的zkfc，让它去成为active，这个namenode就会去强行杀死假死的namenode，防止脑裂！如果杀不死就自定义一个脚本强制它关机，成功之后才会变成active。

现在合并fsimage是由standby来完成的，没有secondaryNameNode；

 

在module目录下创建一个ha文件夹
mkdir ha

将/opt/module/下的 hadoop-2.7.2拷贝到/opt/module/ha目录下
cp -r hadoop-2.7.2/ /opt/module/ha/
删除data logs等文件

 配置core-site.xml

<configuration>
    <!-- 把两个NameNode）的地址组装成一个集群mycluster -->
    <property>
        <name>fs.defaultFS</name>
        <value>hdfs://mycluster</value>
    </property>

    <!-- 指定hadoop运行时产生文件的存储目录 -->
    <property>
        <name>hadoop.tmp.dir</name>
        <value>/opt/module/ha/hadoop-2.7.2/data/tmp</value>
    </property>

    <property>
        <name>ha.zookeeper.quorum</name>
        <value>hadoop101:2181,hadoop102:2181,hadoop103:2181</value>
    </property>

</configuration>

配置hdfs-site.xml

<configuration>
    <!-- 完全分布式集群名称 -->
    <property>
        <name>dfs.nameservices</name>
        <value>mycluster</value>
    </property>

    <!-- 集群中NameNode节点都有哪些 -->
    <property>
        <name>dfs.ha.namenodes.mycluster</name>
        <value>nn1,nn2</value>
    </property>

    <!-- nn1的RPC通信地址 -->
    <property>
        <name>dfs.namenode.rpc-address.mycluster.nn1</name>
        <value>hadoop101:9000</value>
    </property>

    <!-- nn2的RPC通信地址 -->
    <property>
        <name>dfs.namenode.rpc-address.mycluster.nn2</name>
        <value>hadoop102:9000</value>
    </property>

    <!-- nn1的http通信地址 -->
    <property>
        <name>dfs.namenode.http-address.mycluster.nn1</name>
        <value>hadoop101:50070</value>
    </property>

    <!-- nn2的http通信地址 -->
    <property>
        <name>dfs.namenode.http-address.mycluster.nn2</name>
        <value>hadoop102:50070</value>
    </property>

    <!-- 指定NameNode元数据在JournalNode上的存放位置 -->
    <property>
        <name>dfs.namenode.shared.edits.dir</name>
    <value>qjournal://hadoop101:8485;hadoop102:8485;hadoop103:8485/mycluster</value>
    </property>

    <!-- 配置隔离机制，即同一时刻只能有一台服务器对外响应 -->
    <property>
        <name>dfs.ha.fencing.methods</name>
        <value>sshfence</value>
    </property>

    <!-- 使用隔离机制时需要ssh无秘钥登录-->
    <property>
        <name>dfs.ha.fencing.ssh.private-key-files</name>
        <value>/home/kris/.ssh/id_rsa</value>
    </property>

    <!-- 声明journalnode服务器存储目录-->
    <property>
        <name>dfs.journalnode.edits.dir</name>
        <value>/opt/module/ha/hadoop-2.7.2/data/jn</value>
    </property>

    <!-- 关闭权限检查-->
    <property>
        <name>dfs.permissions.enable</name>
        <value>false</value>
    </property>

    <!-- 访问代理类：client，mycluster，active配置失败自动切换实现方式-->
    <property>
          <name>dfs.client.failover.proxy.provider.mycluster</name>
    <value>org.apache.hadoop.hdfs.server.namenode.ha.ConfiguredFailoverProxyProvider</value>
    </property>

    <property>
        <name>dfs.ha.automatic-failover.enabled</name>
        <value>true</value>
    </property>

</configuration>

发送到其他机器
xsync ha

启动HDFS-HA集群

1. 在各个JournalNode节点上，输入以下命令启动journalnode服务
    sbin/hadoop-daemons.sh start journalnode //加个s就可以3台一块启动；都启动之后才能格式化namenode；只能格式化一次！
2. 在[nn1]上，对其进行格式化，并启动
    bin/hdfs namenode -format 
19/02/13 02:15:00 INFO util.GSet: Computing capacity for map NameNodeRetryCache
19/02/13 02:15:00 INFO util.GSet: VM type = 64-bit
19/02/13 02:15:00 INFO util.GSet: 0.029999999329447746% max memory 889 MB = 273.1 KB
19/02/13 02:15:00 INFO util.GSet: capacity = 2^15 = 32768 entries
19/02/13 02:15:01 INFO namenode.FSImage: Allocated new BlockPoolId: BP-26035536-192.168.1.101-1549995301800
19/02/13 02:15:01 INFO common.Storage: Storage directory /opt/module/ha/hadoop-2.7.2/data/tmp/dfs/name has been successfully formatted.
19/02/13 02:15:02 INFO namenode.NNStorageRetentionManager: Going to retain 1 images with txid >= 0
19/02/13 02:15:02 INFO util.ExitUtil: Exiting with status 0
19/02/13 02:15:02 INFO namenode.NameNode: SHUTDOWN_MSG: 
/************************************************************
SHUTDOWN_MSG: Shutting down NameNode at hadoop101/192.168.1.101
************************************************************/
  启动namenode:  sbin/hadoop-daemon.sh start namenode

3.    在[nn2]上，同步nn1的元数据信息
    bin/hdfs namenode -bootstrapStandby
......
STARTUP_MSG: build = Unknown -r Unknown; compiled by 'root' on 2017-05-22T10:49Z
STARTUP_MSG: java = 1.8.0_144
************************************************************/
19/02/21 17:56:25 INFO namenode.NameNode: registered UNIX signal handlers for [TERM, HUP, INT]
19/02/21 17:56:25 INFO namenode.NameNode: createNameNode [-bootstrapStandby]
=====================================================
About to bootstrap Standby ID nn2 from:
Nameservice ID: mycluster
Other Namenode ID: nn1
Other NN's HTTP address: http://hadoop101:50070
Other NN's IPC address: hadoop101/192.168.1.101:9000
Namespace ID: 411281390
Block pool ID: BP-1462258257-192.168.1.101-1550740170734
Cluster ID: CID-d20dda0d-49d1-48f4-b9e8-2c99b72a15c2
Layout version: -63
isUpgradeFinalized: true
=====================================================
19/02/13 02:16:51 INFO common.Storage: Storage directory /opt/module/ha/hadoop-2.7.2/data/tmp/dfs/name has been successfully formatted.
19/02/13 02:16:51 INFO namenode.TransferFsImage: Opening connection to http://hadoop101:50070/imagetransfer?getimage=1&txid=0&storageInfo=-63:1640720426:0:CID-81cbaa0d-6a6f-4932-98ba-ff2a46d87514
19/02/13 02:16:51 INFO namenode.TransferFsImage: Image Transfer timeout configured to 60000 milliseconds
19/02/13 02:16:52 INFO namenode.TransferFsImage: Transfer took 0.02s at 0.00 KB/s
19/02/13 02:16:52 INFO namenode.TransferFsImage: Downloaded file fsimage.ckpt_0000000000000000000 size 351 bytes.
19/02/13 02:16:52 INFO util.ExitUtil: Exiting with status 0
19/02/13 02:16:52 INFO namenode.NameNode: SHUTDOWN_MSG: 
/************************************************************
SHUTDOWN_MSG: Shutting down NameNode at hadoop102/192.168.1.102
************************************************************/
出现这个信息是：（只需格式化一次；）
　　Re-format filesystem in Storage Directory /opt/module/ha/hadoop-2.7.2/data/tmp/dfs/name ? (Y or N) N
Format aborted in Storage Directory /opt/module/ha/hadoop-2.7.2/data/tmp/dfs/name
19/02/21 19:06:50 INFO util.ExitUtil: Exiting with status 5  ##这个是退出状态！
19/02/21 19:06:50 INFO namenode.NameNode: SHUTDOWN_MSG: 
/************************************************************
SHUTDOWN_MSG: Shutting down NameNode at hadoop102/192.168.1.102
************************************************************/

4.    启动[nn2]
    sbin/hadoop-daemon.sh start namenode
6.    在[nn1]上，启动所有datanode； hadoop-daemons.sh是3台都启动datanode
    sbin/hadoop-daemons.sh start datanode
################手动切换namenode          
7.    将[nn1]切换为Active
    bin/hdfs haadmin -transitionToActive nn1
5.    查看是否Active
    bin/hdfs haadmin -getServiceState nn1

    http://hadoop101:50070/dfshealth.html#tab-overview
    http://hadoop102:50070/dfshealth.html#tab-overview

配置HDFS-HA自动故障转移（直接启动不需先按手动的启动方式）

配置好自动故障转移后手动模式就不可用了

Automatic failover is enabled for NameNode at hadoop102/192.168.1.102:9000
Refusing to manually manage HA state, since it may cause a split-brain scenario or other incorrect state.
If you are very sure you know what you are doing, please 
specify the --forcemanual flag.

配置HDFS-HA自动故障转移
1.    具体配置 ；在上个基础上添加如下：
    （1）在hdfs-site.xml中增加
<property>
    <name>dfs.ha.automatic-failover.enabled</name>
    <value>true</value>
</property>
    （2）在core-site.xml文件中增加
<property>
    <name>ha.zookeeper.quorum</name>
    <value>hadoop101:2181,hadoop102:2181,hadoop103:2181</value>
</property>
2.    启动
    （1）关闭所有HDFS服务：
sbin/stop-dfs.sh
    （2）启动Zookeeper集群：
bin/zkServer.sh start
    （3）初始化HA在Zookeeper中状态：(只初始化一次)
bin/hdfs zkfc -formatZK

19/02/13 02:04:02 INFO zookeeper.ClientCnxn: Socket connection established to hadoop102/192.168.1.102:2181, initiating session
19/02/13 02:04:02 INFO zookeeper.ClientCnxn: Session establishment complete on server hadoop102/192.168.1.102:2181, sessionid = 0x268e2dee6e40000, negotiated timeout = 5000
19/02/13 02:04:02 INFO ha.ActiveStandbyElector: Session connected.
19/02/13 02:04:02 INFO ha.ActiveStandbyElector: Successfully created /hadoop-ha/mycluster in ZK.
19/02/13 02:04:02 INFO zookeeper.ZooKeeper: Session: 0x268e2dee6e40000 closed
19/02/13 02:04:02 INFO zookeeper.ClientCnxn: EventThread shut down

    （4）启动HDFS服务：
sbin/start-dfs.sh

hadoop101: starting namenode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-namenode-hadoop101.out
hadoop102: starting namenode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-namenode-hadoop102.out
hadoop102: starting datanode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-datanode-hadoop102.out
hadoop101: starting datanode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-datanode-hadoop101.out
hadoop103: starting datanode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-datanode-hadoop103.out
Starting journal nodes [hadoop101 hadoop102 hadoop103]
hadoop103: starting journalnode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-journalnode-hadoop103.out
hadoop102: starting journalnode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-journalnode-hadoop102.out
hadoop101: starting journalnode, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-journalnode-hadoop101.out
Starting ZK Failover Controllers on NN hosts [hadoop101 hadoop102]
hadoop101: starting zkfc, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-zkfc-hadoop101.out
hadoop102: starting zkfc, logging to /opt/module/ha/hadoop-2.7.2/logs/hadoop-kris-zkfc-hadoop102.out

3.    验证
    （1）将Active NameNode进程kill
    kill -9 namenode的进程id   ---> 另外一个namenode上位；( 杀掉之后，可单独启动: sbin/hadoop-daemon.sh start namenode，启动之后它不会变成之前的active；而是standby)
     (2) 将DFSZKFailoverController的进程kill
      kill -9 8819   --->尽管它的namenode没有挂，但另外一个namenode也会上位，它变成standby
      再启动它要先停止->sbin/stop-dfs.sh  sbin/start-dfs.sh 
    （3）将Active NameNode机器断开网络
    sudo service network stop  --->  把网络断开之后，配的隔离机制是sshfence尝试远程登录区杀敌hadoop101，直到把101网络连接上才杀死，102才成为active；

 

断网上位很容易脑裂，那边网已恢复就炸了；如果真的断网，它这个是不会自动切的是为了防脑裂。虽然可以通过配置断网可以直接上，但很危险；断网上不去反而安全。

 

[zk: localhost:2181(CONNECTED) 0] ls /
[zookeeper, hadoop-ha]
[zk: localhost:2181(CONNECTED) 1] ls /
[zookeeper, hadoop-ha]
[zk: localhost:2181(CONNECTED) 2] ls /hadoop-ha
[mycluster]
[zk: localhost:2181(CONNECTED) 3] ls /hadoop-ha/mycluster
[ActiveBreadCrumb, ActiveStandbyElectorLock] 选举的关键节点，谁占领了这个节点谁就是active； 
[zk: localhost:2181(CONNECTED) 4] get /hadoop-ha/mycluster/ActiveStandbyElectorLock

        myclusternn1 hadoop101 �F(�>
cZxid = 0x10000000f
ctime = Wed Feb 13 02:16:22 CST 2019
mZxid = 0x10000000f
mtime = Wed Feb 13 02:16:22 CST 2019
pZxid = 0x10000000f
cversion = 0
dataVersion = 0
aclVersion = 0
ephemeralOwner = 0x268e2dee6e40002  临时节点，
dataLength = 33
numChildren = 0

进程

hadoop101                hadoop102                hadoop103
 NameNode                NameNode
 JournalNode            JournalNode                JournalNode
 DataNode                DataNode                DataNode
 DFSZKFailoverController DFSZKFailoverController ZooKeeperMain（bin/zkCli.sh ,启动zookeeperd客户端）
 ResourceManager         ResourceManager
 NodeManager             NodeManager            NodeManager    

QuorumPeerMain            QuorumPeerMain            QuorumPeerMain (bin/zkServer.sh start 启动zookeeper服务器)

DFSZKFailoverController是Hadoop-2.7.0中HDFS NameNode HA实现的中心组件，它负责整体的故障转移控制等。
它是一个守护进程，通过main()方法启动，继承自ZKFailoverController。  zkfc
 使用JournalNode实现两个NameNode(Active和Standby)之间数据的共享
 

YARN-HA配置

YARN-HA工作机制

 Yarn原生根zookeeper兼容很好，配置比较简单；

 配置YARN-HA集群

  yarn-site.xml

    <property>
        <name>yarn.nodemanager.aux-services</name>
        <value>mapreduce_shuffle</value>  //混合服务还是shuffle，用shuffle确定reducer获取数据的方式
    </property>

    <!--启用resourcemanager ha-->
    <property>
        <name>yarn.resourcemanager.ha.enabled</name>
        <value>true</value>
    </property>
 
    <!--声明两台resourcemanager的地址-->
    <property>
        <name>yarn.resourcemanager.cluster-id</name>
        <value>cluster-yarn1</value>
    </property>

    <property>
        <name>yarn.resourcemanager.ha.rm-ids</name>
        <value>rm1,rm2</value>
    </property>

    <property>
        <name>yarn.resourcemanager.hostname.rm1</name>
        <value>hadoop101</value>
    </property>

    <property>
        <name>yarn.resourcemanager.hostname.rm2</name>
        <value>hadoop102</value>
    </property>
 
    <!--指定zookeeper集群的地址--> 
    <property>
        <name>yarn.resourcemanager.zk-address</name>
        <value>hadoop101:2181,hadoop102:2181,hadoop103:2181</value>
    </property>

    <!--启用自动恢复--> 
    <property>
        <name>yarn.resourcemanager.recovery.enabled</name>
        <value>true</value>
    </property>
 
    <!--指定resourcemanager的状态信息存储在zookeeper集群--> 
    <property>
        <name>yarn.resourcemanager.store.class</name>     
    <value>org.apache.hadoop.yarn.server.resourcemanager.recovery.ZKRMStateStore</value>
    </property>

yarn-site.xml 在hadoop101中配置好将其分发到其他机器：xsync etc/ 

启动zookeeper；3台其他都启动；
　　[kris@hadoop101 ~]$ /opt/module/zookeeper-3.4.10/bin/zkServer.sh start
启动hdfs 
　　sbin/start-dfs.sh

启动YARN 
（1）在hadoop101中执行：
　　sbin/start-yarn.sh
[kris@hadoop101 hadoop-2.7.2]$ sbin/start-yarn.sh 
starting yarn daemons
starting resourcemanager, logging to /opt/module/ha/hadoop-2.7.2/logs/yarn-kris-resourcemanager-hadoop101.out
hadoop101: starting nodemanager, logging to /opt/module/ha/hadoop-2.7.2/logs/yarn-kris-nodemanager-hadoop101.out
hadoop103: starting nodemanager, logging to /opt/module/ha/hadoop-2.7.2/logs/yarn-kris-nodemanager-hadoop103.out
hadoop102: starting nodemanager, logging to /opt/module/ha/hadoop-2.7.2/logs/yarn-kris-nodemanager-hadoop102.out


（2）在hadoop102中执行：
　　sbin/yarn-daemon.sh start resourcemanager //脚本只在上面一台上启动了，hadoop102上的resourcemanager要手动启；
[kris@hadoop102 hadoop-2.7.2]$ sbin/yarn-daemon.sh start resourcemanager
starting resourcemanager, logging to /opt/module/ha/hadoop-2.7.2/logs/yarn-kris-resourcemanager-hadoop102.out


（3）查看服务状态
　　bin/yarn rmadmin -getServiceState rm1  //rm1就是hadoop101;  rm2为hadoop102
[kris@hadoop102 hadoop-2.7.2]$ bin/yarn rmadmin -getServiceState rm1
active
[kris@hadoop102 hadoop-2.7.2]$ bin/yarn rmadmin -getServiceState rm2
standby

################
hdfs-ha & yarn-ha
[kris@hadoop101 hadoop-2.7.2]$ jpsall
-------hadoop101-------
14288 ResourceManager
14402 NodeManager
13396 NameNode
14791 Jps
13114 QuorumPeerMain
13946 DFSZKFailoverController
13516 DataNode
13743 JournalNode
-------hadoop102-------
9936 NameNode
10352 ResourceManager
9569 JournalNode
9698 DFSZKFailoverController
9462 DataNode
9270 QuorumPeerMain
10600 Jps
10202 NodeManager
-------hadoop103-------
9073 Jps
8697 JournalNode
8522 QuorumPeerMain
8909 NodeManager
8590 DataNode

　　http://hadoop101:8088/cluster     http://hadoop102:8088/cluster -->它会重定向到hadoop101