前言：Hadoop NameNode节点维护了整个HDFS集群所有的元数据信息，一旦NameNode发生脑裂，或者服务不可用，整个HDFS集群都将处于不可用状态。对于线上生产环境，将会造成不可估量的损失。因此，对于NameNode节点的状态监控不可或缺，一旦发生脑裂或者服务异常停止，都能够及时通知运维介入。当然，NameNode本身有Fencing机制，Hadoop默认Fencing机制不做任何动作，这就需要开发人员自主配置Fencing机制需要执行的操作。
本文要旨：① 如何监控NameNode的状态切换； ② 自主配置Fencing机制；
集群环境：Ambari 2.2.1， HDP 2.3.4，Hadoop 2.7.1
【NameNode状态监控】
① 访问url获取集群Namenode状态：
http://${hostname}:50070/jmx?qry=Hadoop:service=NameNode,name=NameNodeStatus
返回如下结果：
{
"beans" : [ {
"name" : "Hadoop:service=NameNode,name=NameNodeStatus",
"modelerType" : "org.apache.hadoop.hdfs.server.namenode.NameNode",
"State" : "active",
"NNRole" : "NameNode",
"HostAndPort" : "hostname1:8020",
"SecurityEnabled" : false,
"LastHATransitionTime" : 1519300091101,
"BytesWithFutureGenerationStamps" : 0
} ]
}
② 依次访问两个namenode节点，返回信息通过python脚本解析，获取节点状态:
# -*- coding: utf-8 -*-
import json
import urllib
import sys
def get_namenode_status(hostname):
jmxport = 50070
url = "http://{0}:{1}/jmx?qry=Hadoop:service=NameNode,name=NameNodeStatus".format(hostname, jmxport)
try:
response = urllib.urlopen(url)
jsonString = json.load(response)
nnstatus = jsonString.get('beans', [{}])[0].get('State', '')
if 'active' == nnstatus:
return 'active'
elif 'standby' == nnstatus:
return 'standby'
else:
return
except:
return
if __name__ == "__main__":
hostname = (sys.argv[1] if len(sys.argv) > 1 else None)
if hostname:
print get_namenode_status(hostname)
else:
print None
脚本返回三种状态：active(主)，standby(备)，None(服务异常，无法获取状态)
③ 通过shell脚本根据返回状态判断是否报警：
#!/bin/bash
bin=`dirname "$0"`
code_path=`cd "$bin";pwd`
file_name="last_ha_status"
export monitor_info=""
nn1=`python $code_path/namenodestatus.py hostname1`
nn2=`python $code_path/namenodestatus.py hostname2`
last_status_hdfs=`cat /tmp/$file_name`
current_status_hdfs=$nn1,$nn2
if [[ "standby" == $nn1 ]] && [[ "standby" == $nn2 ]]; then
monitor_info="hostname1:"$nn1" hostname2:"$nn2" 上次HA状态信息:"$last_status_hdfs" 当前HA状态信息:"$current_status_hdfs
echo $nn1,$nn2 > /tmp/$file_name
#todo 调用报警接口，打印monitor_info内容
exit 1
elif [[ "active" == $nn1 ]] && [[ "active" == $nn2 ]]; then
monitor_info="hostname1:"$nn1" hostname2:"$nn2" 上次HA状态信息:"$last_status_hdfs" 当前HA状态信息:"$current_status_hdfs
echo $nn1,$nn2 > /tmp/$file_name
#todo 调用报警接口，打印monitor_info内容
exit 1
elif [[ "None" == $nn1 ]] || [[ "None" == $nn2 ]]; then
monitor_info="hostname1:"$nn1" hostname2:"$nn2" 上次HA状态信息:"$last_status_hdfs" 当前HA状态信息:"$current_status_hdfs
echo $nn1,$nn2 > /tmp/$file_name
#todo 调用报警接口，打印monitor_info内容
exit 1
else
#状态正常，将正常状态写入本地文件
echo $nn1,$nn2 > /tmp/$file_name
exit 0
fi
报警阈值：
① 两个NameNode都处于standby状态；
② 两个NameNode都处于active状态；
③ 任何一个NameNode服务停止；
NameNode的主备状态获取途径除了通过访问50070获取JMX监控信息外，还可以通过hdfs haadmin命令获取：
nn1=`ssh hdfs@hostname1 hdfs haadmin -Dipc.client.connect.max.retries.on.timeouts=3 -Dipc.client.connect.timeout=3000 -GetServiceState nn1`
nn2=`ssh hdfs@hostname2 hdfs haadmin -Dipc.client.connect.max.retries.on.timeouts=3 -Dipc.client.connect.timeout=3000 -GetServiceState nn2`
备注①：执行监控脚本的用户A需要设置访问hdfs用户的免密码登录，当然也可以开通用户A访问hdfs的权限；
备注②：hdfs core-site.xml默认配置定义了socket连接最大重试次数为45，以及每次等待连接时间为20s。这种情况下 如果NameNode节点异常，需要等待45 * 20s才能返回相应结果。对于监控脚本，这么长的等待时间显然不合理，因此需要我们自定义最长允许的等待时间。
<property>
<name>ipc.client.connect.timeout</name>
<value>20000</value>
<description>Indicates the number of milliseconds a client will wait for the socket to establish a server connection.
</description>
</property>
<property>
<name>ipc.client.connect.max.retries.on.timeouts</name>
<value>45</value>
<description>Indicates the number of retries a client will make on socket timeout to establish a server connection.
</description>
</property>
【自定义Fencing机制】
参考官方文档：https://hadoop.apache.org/docs/r2.7.1/hadoop-project-dist/hadoop-hdfs/HDFSHighAvailabilityWithQJM.html
在分布式系统中脑裂又称为双主现象，由于Zookeeper的“假死”，长时间的垃圾回收或其它原因都可能导致双Active NameNode现象，此时两个NameNode都可以对外提供服务，无法保证数据一致性。对于生产环境，这种情况的出现是毁灭性的，必须通过自带的隔离（Fencing）机制预防这种现象的出现。
Hadoop目前主要提供两种隔离措施：
sshfence：SSH to the Active NameNode and kill the process；
shellfence：run an arbitrary shell command to fence the Active NameNode；
只有在成功地执行完成fencing之后，选主成功的ActiveStandbyElector才会回调ZKFC的becomeActive方法将对应的NameNode切换为Active，开始对外提供服务。
2018-02-22 18:57:11,865 INFO ha.NodeFencer (NodeFencer.java:fence(91)) - ====== Beginning Service Fencing Process... ======
2018-02-22 18:57:11,865 INFO ha.NodeFencer (NodeFencer.java:fence(94)) - Trying method 1/1: org.apache.hadoop.ha.ShellCommandFencer(/usr/custom/hdfs/restart_namenode.sh)
2018-02-22 18:57:11,871 INFO ha.ShellCommandFencer (ShellCommandFencer.java:tryFence(99)) - Launched fencing command '/usr/custom/hdfs/restart_namenode.sh' with pid 21772
2018-02-22 18:57:11,894 WARN ha.ShellCommandFencer (StreamPumper.java:pump(88)) - [PID 21772] /usr/cus...menode.sh: % Total % Received % Xferd Average Speed Time Time Time Current
2018-02-22 18:57:11,894 WARN ha.ShellCommandFencer (StreamPumper.java:pump(88)) - [PID 21772] /usr/cus...menode.sh: Dload Upload Total Spent Left Speed
2018-02-22 18:57:11,894 WARN ha.ShellCommandFencer (StreamPumper.java:pump(88)) - [PID 21772] /usr/cus...menode.sh:
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: {
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: "href" : "http://hostname1:8080/api/v1/clusters/cbas_cluster/requests/5363",
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: "Requests" : {
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: "id" : 5363,
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: "status" : "Accepted"
2018-02-22 18:57:11,965 WARN ha.ShellCommandFencer (StreamPumper.java:pump(88)) - [PID 21772] /usr/cus...menode.sh: 0 0 0 0 0 346 0 343k --:--:-- --:--:-- --:--:-- 343k
2018-02-22 18:57:11,965 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: }
2018-02-22 18:57:11,966 INFO ha.ShellCommandFencer (StreamPumper.java:pump(86)) - [PID 21772] /usr/cus...menode.sh: }
2018-02-22 18:57:11,965 WARN ha.ShellCommandFencer (StreamPumper.java:pump(88)) - [PID 21772] /usr/cus...menode.sh: 495 149 149 149 0 346 2073 4815 --:--:-- --:--:-- --:--:-- 0
2018-02-22 18:57:11,967 INFO ha.NodeFencer (NodeFencer.java:fence(98)) - ====== Fencing successful by method org.apache.hadoop.ha.ShellCommandFencer(/usr/custom/hdfs/restart_namenode.sh) ======
2018-02-22 18:57:11,967 INFO ha.ActiveStandbyElector (ActiveStandbyElector.java:writeBreadCrumbNode(878)) - Writing znode /hadoop-ha/cbas/ActiveBreadCrumb to indicate that the local node is the most recent active...
2018-02-22 18:57:11,969 INFO ha.ZKFailoverController (ZKFailoverController.java:becomeActive(380)) - Trying to make NameNode at hostname1/xx.x.xx.xxx:8021 active...
2018-02-22 18:57:13,710 INFO ha.ZKFailoverController (ZKFailoverController.java:becomeActive(387)) - Successfully transitioned NameNode at hostname1/xx.x.xx.xxx:8021 to active state
本章只介绍shellfence机制，因为作者所使用集群通过Ambari管理，因此可以调用ambari自带的Rest API达到重启NameNode节点的效果。
Ambari Rest API官网文档：https://cwiki.apache.org/confluence/display/AMBARI/Restart+Host+Components
#!/bin/bash
curl -u admin:${admin_password} -H 'X-Requested-By: ambari' -X POST -d '
{
"RequestInfo":{
"command":"RESTART",
"context":"Restart NameNode",
"operation_level":{
"level":"HOST",
"cluster_name":"${cluster_name}"
}
},
"Requests/resource_filters":[
{
"service_name":"HDFS",
"component_name":"NAMENODE",
"hosts":"${hostname1}"
}
]
}' http://${ambari_hostname}:8080/api/v1/clusters/${cluster_name}/requests
① 将脚本存放在两台NameNode所在服务器上：/usr/custom/hdfs/restart_namenode.sh，注意脚本的可执行权限，并修改相应变量以达到远程重启另外一个NameNode服务的目的；
② 修改hdfs配置文件：
Amabri入口：Ambari UI => HDFS => Configs => Advanced => Custom hdfs-site
<property>
<name>dfs.ha.fencing.methods</name>
<value>shell(/usr/custom/hdfs/restart_namenode.sh)</value>
</property>
结束语：对于生产环境的大数据平台建设，NameNode HA模式的状态切换都应积极跟进，查找状态发生切换的根本原因，避免或者是减少切换的频率。状态监控和自定义Fencing机制都只是NameNode异常发生后的处理措施。即使发生状态切换，也应该实现自动化，避免人工干预，用户无感知。