标签为 Namenode 的文章

Hadoop运维笔记 之 Namenode异常停止后无法正常启动

背景:
公司在线上使用了CDH5 HA模式,有两个Namenode节点,结果其中的Standby节点因为一些关于edits文件的报错异常停止了,并且在启动的过程中一直报告找不到各种文件。

刚开始怀疑问题可能只发生在Standby本身,因此尝试了bootstrapStandby来重新初始化Standby节点,但问题依旧。
而后来因为我尝试重启ZKFC(Zookeeper Failover)服务器,导致了Active节点进行自动切换,在切换失败后打算切换回去时,也无法启动服务了,报错跟Standby节点一模一样,于是整个Hadoop集群就挂了。

问题严重,在搜遍了整个Google都找不到任何有用的信息之后,只能求助于老大。最后,老大想到一个思路,就是将fsimage(元数据)文件与edits(编辑日志)文件都反编译成文本,查看里面具体有什么内容,为什么加载edits文件时会报错。

结果,这个思路给我们带来了曙光,并最终修复了整个集群。

环境介绍:
idc2-server1: namenode, journalnode, zkfc
idc2-server2: namenode, journalnode, zkfc
idc2-server3: journalnode, resourcemanager

具体过程:
首先,是Standby Namenode上出现以下错误,然后自动异常关闭了进程:

2014-11-11 02:12:54,057 FATAL org.apache.hadoop.hdfs.server.namenode.ha.EditLogTailer: Unknown error encountered while tailing edits. Shutting down standby NN.
java.io.FileNotFoundException: File does not exist: /user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015
       at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:65)
       at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:55)
...

其中提到了"Unknown error encountered while tailing edits. Shutting down standby NN."

于是,我们尝试启动Standby Namenode服务,结果报告以下错误:

2014-11-12 04:26:28,860 INFO org.apache.hadoop.hdfs.server.namenode.EditLogInputStream: Fast-forwarding stream 'http://idc2-server10.heylinux.com:8480/getJournal?jid=idc2&segmentTxId=240823073&storageInfo=-55%3A1838233660%3A0%3ACID-d77ea84b-1b24-4bc2-ad27-7d2384d222d6' to transaction ID 240741256
2014-11-12 04:26:28,874 ERROR org.apache.hadoop.hdfs.server.namenode.FSEditLogLoader: Encountered exception on operation CloseOp [length=0, inodeId=0, path=/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015, replication=3, mtime=1415671845582, atime=1415670522749, blockSize=134217728, blocks=[], permissions=oozie:hdfs:rw-r--r--, aclEntries=null, clientName=, clientMachine=, opCode=OP_CLOSE, txid=240823292]
java.io.FileNotFoundException: File does not exist: /user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:65)
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:55)
...
2014-11-12 04:26:32,641 WARN org.apache.hadoop.hdfs.server.namenode.FSNamesystem: Encountered exception loading fsimage
java.io.FileNotFoundException: File does not exist: /user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:65)

说找不到"/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015"这个文件。
而事实上,这个文件是临时文件,不重要并且已经被删除了。
但在上面,却报告"ERROR org.apache.hadoop.hdfs.server.namenode.FSEditLogLoader: Encountered exception on operation CloseOp",可以看出是在加载edits文件,执行OP_CLOSE操作时提示找不到文件。

刚开始我们怀疑可能只是Standby上的fsimage文件或edits文件有问题,于是我们在Standby上执行了bootstrapStandby,改过程会自动从Active Namenode上获取最新的fsimage文件,并从Journalnode日志服务器上下载并执行新的edits文件。

sudo -u hdfs hadoop namenode -bootstrapStandby

但是,在初始化之后,加载edits时仍然遇到上面相同的报错。

而接下来,由于我尝试将ZKFC(Zookeeper Failover)服务重启,导致了Active Namenode自动切换到Standby,但由于Standby无法take over,所以Active Namenode切换回去的时候,也无法正常重启了,错误跟Standby启动时一样。

这样一来,整个Hadoop集群就挂了,在搜遍了整个Google都找不到任何有用的信息之后,我打电话给了老大,老大也通过上面的错误Google不到任何一条有用的信息。
于是老大尝试在edits文件中grep上面的路径,找到了一些相关的edits文件:

# cd /data1/dfs/nn/
# cp -rpa current current.backup.orig
# cd /data2/dfs/nn/
# cp -rpa current current.backup.orig
# cd /data1/dfs/nn/current
# grep attempt_1415171013961_37060_m_000015_0 *
Binary file edits_0000000000240687057-0000000000240698453 matches
Binary file edits_0000000000240823073-0000000000240838096 matches
Binary file edits_inprogress_0000000000244853266 matches

于是,我们思考能否从这些edits文件或fsimage文件中找到一些线索。
而下面的两篇文章中,提到了Hadoop自带的针对fsimage和edits文件的反编译工具:
http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-hdfs/HdfsImageViewer.html
http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-hdfs/HdfsEditsViewer.html

其中,关于edits文件的一些描述给我们带来了极大的希望:

In case there is some problem with hadoop cluster and the edits file is corrupted it is possible to save at least part of the edits file that is correct. This can be done by converting the binary edits to XML, edit it manually and then convert it back to binary.

通过以上描述,我们了解到edits文件可能会corrupted,而反编译之后手动修改,在编译回二进制格式进行替换,可以作为一种解决办法。
于是我们将上面找到的两个关联edits文件,将其复制出来并进行了反编译:

# mkdir /tmp2/
# cd /data1/dfs/nn
# cp edits_0000000000240687057-0000000000240698453 /tmp2/
# cp edits_0000000000240823073-0000000000240838096 /tmp2/
# cd /tmp2
# hdfs oev -i edits_0000000000240687057-0000000000240698453 -o edits_0000000000240687057-0000000000240698453.xml
# hdfs oev -i edits_0000000000240823073-0000000000240838096 -o edits_0000000000240823073-0000000000240838096.xml

反编译之后,生成了两个XML文件,我们在XML文件中搜索"/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015",找到了OP_CLOSE与OP_DELETE相关记录:

  <RECORD>
    <OPCODE>OP_DELETE</OPCODE>
    <DATA>
      <TXID>240818498</TXID>
      <LENGTH>0</LENGTH>
      <PATH>/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015</PATH>
      <TIMESTAMP>1415671972595</TIMESTAMP>
      <RPC_CLIENTID>4a38861d-3bee-40e6-abb6-d2b58f313781</RPC_CLIENTID>
      <RPC_CALLID>676</RPC_CALLID>
    </DATA>
  </RECORD>
  <RECORD>
    <OPCODE>OP_CLOSE</OPCODE>
    <DATA>
      <TXID>240823292</TXID>
      <LENGTH>0</LENGTH>
      <INODEID>0</INODEID>
      <PATH>/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015</PATH>
      <REPLICATION>3</REPLICATION>
      <MTIME>1415671845582</MTIME>
      <ATIME>1415670522749</ATIME>
      <BLOCKSIZE>134217728</BLOCKSIZE>
      <CLIENT_NAME></CLIENT_NAME>
      <CLIENT_MACHINE></CLIENT_MACHINE>
      <PERMISSION_STATUS>
        <USERNAME>oozie</USERNAME>
        <GROUPNAME>hdfs</GROUPNAME>
        <MODE>420</MODE>
      </PERMISSION_STATUS>
    </DATA>
  </RECORD>

可以看到,对于"/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000015_0/part-00015",OP_DELETE发生在了OP_CLOSE之前,因此执行OP_CLOSE时会提示"File does not exist"。

于是,我们尝试将OP_CLOSE这部分代码,替换成其它的内容,比如无关痛痒的修改一个现有文件的权限,并保留TXID 240823292以确保edits文件的完整性。

  <RECORD>
    <OPCODE>OP_SET_PERMISSIONS</OPCODE>
    <DATA>
      <TXID>240823292</TXID>
      <SRC>/user/oozie-heylinux/.staging/job_1415171013961_37194</SRC>
      <MODE>504</MODE>
    </DATA>
  </RECORD>

修改完成之后,再将XML文件反编译回binary格式。

# cd /tmp2/
# cp edits_0000000000240823073-0000000000240838096.xml edits_0000000000240823073-0000000000240838096.xml.orig
# vim edits_0000000000240823073-0000000000240838096.xml
# hdfs oev -i edits_0000000000240823073-0000000000240838096.xml -o edits_0000000000240823073-0000000000240838096 -p binary 

然后将binary文件同步到journalnode日志服务器中:

# cd /var/hadoop/data/dfs/jn/idc2prod/
# cp -rpa current current.backup.orig
# cd /tmp2/
# cp edits_0000000000240823073-0000000000240838096 /data1/dfs/nn/current/
# cp edits_0000000000240823073-0000000000240838096 /data2/dfs/nn/current/
# cp edits_0000000000240823073-0000000000240838096 /var/hadoop/data/dfs/jn/idc2prod/current/
# scp edits_0000000000240823073-0000000000240838096 root@idc2-server2:/var/hadoop/data/dfs/jn/idc2prod/current/
# scp edits_0000000000240823073-0000000000240838096 root@idc2-server3:/var/hadoop/data/dfs/jn/idc2prod/current/

然后启动namenode服务,可以发现,之间的错误已经不存在了,取而代之的已经变成了其它文件。

2014-11-12 08:57:13,053 INFO org.apache.hadoop.hdfs.server.namenode.EditLogInputStream: Fast-forwarding stream 'http://idc2-server1.heylinux.com:8480/getJournal?jid=idc2prod&segmentTxId=240823073&storageInfo=-55%3A1838233660%3A0%3ACID-d77ea84b-1b24-4bc2-ad27-7d2384d222d6' to transaction ID 240299210
2014-11-12 08:57:13,063 ERROR org.apache.hadoop.hdfs.server.namenode.FSEditLogLoader: Encountered exception on operation CloseOp [length=0, inodeId=0, path=/user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000018_0/part-00018, replication=3, mtime=1415671845675, atime=1415670519537, blockSize=134217728, blocks=[], permissions=oozie:hdfs:rw-r--r--, aclEntries=null, clientName=, clientMachine=, opCode=OP_CLOSE, txid=240823337]
java.io.FileNotFoundException: File does not exist: /user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000018_0/part-00018
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:65)
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:55)
...

2014-11-12 08:57:16,847 WARN org.apache.hadoop.hdfs.server.namenode.FSNamesystem: Encountered exception loading fsimage
java.io.FileNotFoundException: File does not exist: /user/dong/data/dpp/classification/gender/vw-output-train/2014-10-30-research-with-confict-fix-bug-rerun/_temporary/1/_temporary/attempt_1415171013961_37060_m_000018_0/part-00018
        at org.apache.hadoop.hdfs.server.namenode.INodeFile.valueOf(INodeFile.java:65)
...

那么,接下来,就是重复以上动作,其中有时候能找到一部分规律,可以批量将同一个目录下反复报错的文件的OP_CLOSE都替换掉。但更多的时候则是随机的文件,需要一次次修改XML文件,然后编译成binary,再启动namenode,进行针对性的修改,一直反复的进行下去,直到Namenode能够成功启动为止。

我们在具体的操作过程中,还遇到过关于OP_ADD_BLOCK的错误,造成问题的原因是由于最后一个edits文件在反编译回binary文件时出现一些关于OP_UPDATE_BLOCK的错误。
我将报错的部分通过以上方式进行了替换,才成功的将edits文件反编译回binary文件。
具体的解决办法,就是根据"Encountered exception on operation AddBlockOp"定位到OP_ADD_BLOCK相关配置并替换即可。

2014-11-12 18:07:39,070 ERROR org.apache.hadoop.hdfs.server.namenode.FSEditLogLoader: Encountered exception on operation AddBlockOp [path=/user/dong/data/dpp/classification/gender/vw-input/2014-10-30-research-with-no-confict-fix-bug-rerun/all_labelled/_temporary/1/_temporary/attempt_1415171013961_42350_m_001474_0/part-m-01474, penultimateBlock=NULL, lastBlock=blk_1109647729_35920089, RpcClientId=, RpcCallId=-2]
java.lang.IllegalStateException

最后,在Namenode启动成功后,会报告很多Block丢失,解决办法是通过fsck删除这些错误的Block。

# hadoop fsck / -files -blocks -locations | tee -a fsck.out

然后在fsck.out中获取所有Block的信息,执行"hadoop fsck / -move"加Block路径进行删除。

最后,退出safemode,生活再次变得美好起来。

# hadoop dfsadmin -safemode leave

PS: 在后来,上面的问题又再次出现了,通过分析我们发现是由于Hadoop的一个Bug导致的,从2.10-beta就开始存在了:
https://issues.apache.org/jira/browse/HDFS-6527
https://issues.apache.org/jira/browse/HDFS-6647
想要彻底避免这类问题的再次发生,必须将Hadoop升级到2.5以上版本,而对应的CDH版本则是5.2.0。

,

1 Comment