Hadoop运用常见难题以及解决方法

1：Shuffle Error: Exceeded MAX_FAILED_UNIQUE_FETCHES; bailing-out
Answer：
程序里面需要打开多个文件，进行剖析，系统一般默认数量是1024，（用ulimit -a可以看到）对于正常运用是够了，但是对于程序来讲，就太少了。
修改办法：
修改2个文件。
        /etc/security/limits.conf
vi /etc/security/limits.conf
加上：
* soft nofile 102400
* hard nofile 409600
    $cd /etc/pam.d/
    $sudo vi login
        添加        session    required     /lib/security/pam_limits.so
针对第一个我纠正下答案：
这是reduce预处理阶段shuffle时获取已完成的map的输出失败次数超过上限造成的，上限默认为5。引起此的方式可能会有很多种，比如网络连接不正常，连接超时，带宽较差以及端口阻塞等。。。通常框架内网络情况较好是不会出现此错误的。
2：Too many fetch-failures
Answer:
出现这个主要是结点间的连通不够全面。
1) 检查 、/etc/hosts
   要求高手本机ip 对应 服务器名
   要求高手要包含所有的服务器ip + 服务器名
2) 检查 .ssh/authorized_keys
   要求高手包含所有服务器（包括其自身）的public key
3：处理速度特别的慢 出现map很快 但是reduce很慢 而且反复出现 reduce=0%
Answer:
结合第二点，然后
修改 conf/hadoop-env.sh 中的export HADOOP_HEAPSIZE=4000
4：能够开启datanode，但无法访问，也无法结束的错误
在重新格式化一个新的分布式文件时，需要将你NameNode上所配置的dfs.name.dir这一namenode用来存放NameNode 持久存储名字空间及事务日志的本地文件系统路径删除，同时将各DataNode上的dfs.data.dir的路径 DataNode 存放块数据的本地文件系统路径的目录也删除。如本此配置就是在NameNode上删除/home/hadoop/NameData，在DataNode上删除/home/hadoop/DataNode1和/home/hadoop/DataNode2。这是因为Hadoop在格式化一个新的分布式文件系统时，每个存储的名字空间都对应了建立时间的那个版本（可以查看/home/hadoop /NameData/current目录下的VERSION文件，上面记录了版本信息），在重新格式化新的分布式系统文件时，最好先删除NameData 目录。必须删除各DataNode的dfs.data.dir。这样才可以使namedode和datanode记录的信息版本对应。
注意：删除是个很危险的动作，不能确认的情况下不能删除！！做好删除的文件等通通备份！！
5：java.io.IOException: Could not obtain block: blk_194219614024901469_1100 file=/user/hive/warehouse/src_20090724_log/src_20090724_log
出现这种情况大多是结点断了，没有连接上。
6：java.lang.OutOfMemoryError: Java heap space
出现这种出错，明显是jvm内存不够得原因，要修改所有的datanode的jvm内存大小。
Java -Xms1024m -Xmx4096m
一般jvm的最大内存运用应该为总内存大小的一半，我们运用的8G内存，所以设置为4096m，这一值可能依旧不是最优的值。

顶，这样的贴子非常好，要置顶。附件是由Hadoop技术交流群中若冰的同学提供的相关材料：
Hadoop添加节点的方法
自己实际添加节点过程：
1. 先在slave上配置好环境，包括ssh，jdk，相关config，lib，bin等的拷贝；
2. 将新的datanode的host加到集群namenode及其他datanode中去；
3. 将新的datanode的ip加到master的conf/slaves中；
4. 重启cluster,在cluster中看到新的datanode节点；
5. 运行bin/start-balancer.sh，这个会很耗时间
备注：
1. 如果不balance，那么cluster会把新的数据都存放在新的node上，这样会降低mr的工作效率；
2. 也可调用bin/start-balancer.sh 命令运行，也可加参数 -threshold 5
   threshold 是平衡阈值，默认是10%，值越低各节点越平衡，但消耗时间也更长。
3. balancer也可以在有mr job的cluster上运行，默认dfs.balance.bandwidthPerSec很低，为1M/s。在没有mr job时，可以提高该设置加快负载均衡时间。
其他备注：
1. 必须确保slave的firewall已关闭;
2. 确保新的slave的ip已经添加到master及其他slaves的/etc/hosts中，反之也要将master及其他slave的ip添加到新的slave的/etc/hosts中
mapper及reducer个数
url地址： http://wiki.apache.org/hadoop/HowManyMapsAndReduces
HowManyMapsAndReduces
Partitioning your job into maps and reduces
Picking the appropriate size for the tasks for your job can radically change the performance of Hadoop. Increasing the number of tasks increases the framework overhead, but increases load balancing and lowers the cost of failures. At one extreme is the 1 map/1 reduce case where nothing is distributed. The other extreme is to have 1,000,000 maps/ 1,000,000 reduces where the framework runs out of resources for the overhead.
Number of Maps
The number of maps is usually driven by the number of DFS blocks in the input files. Although that causes people to adjust their DFS block size to adjust the number of maps. The right level of parallelism for maps seems to be around 10-100 maps/node, although we have taken it up to 300 or so for very cpu-light map tasks. Task setup takes awhile, so it is best if the maps take at least a minute to execute.
Actually controlling the number of maps is subtle. The mapred.map.tasks parameter is just a hint to the InputFormat for the number of maps. The default InputFormat behavior is to split the total number of bytes into the right number of fragments. However, in the default case the DFS block size of the input files is treated as an upper bound for input splits. A lower bound on the split size can be set via mapred.min.split.size. Thus, if you expect 10TB of input data and have 128MB DFS blocks, you'll end up with 82k maps, unless your mapred.map.tasks is even larger. Ultimately the [WWW] InputFormat determines the number of maps.
The number of map tasks can also be increased manually using the JobConf's conf.setNumMapTasks(int num). This can be used to increase the number of map tasks, but will not set the number below that which Hadoop determines via splitting the input data.
Number of Reduces
The right number of reduces seems to be 0.95 or 1.75 * (nodes * mapred.tasktracker.tasks.maximum). At 0.95 all of the reduces can launch immediately and start transfering map outputs as the maps finish. At 1.75 the faster nodes will finish their first round of reduces and launch a second round of reduces doing a much better job of load balancing.
Currently the number of reduces is limited to roughly 1000 by the buffer size for the output files (io.buffer.size * 2 * numReduces  * mapred.tasktracker.reduce.tasks.maximum).increasing the number of reduces increases the framework overhead, but increases load balancing and lowers the cost of failures.
  mapred.tasktracker.reduce.tasks.maximum
  2
  The maximum number of reduce tasks that will be run
  simultaneously by a task tracker.
  
单个node新加硬盘
1.修改需要新加硬盘的node的dfs.data.dir，用逗号分隔新、旧文件目录
2.重启dfs
同步hadoop 代码
hadoop-env.sh
# host:path where hadoop code should be rsync'd from.  Unset by default.
# export HADOOP_MASTER=master:/home/$USER/src/hadoop
用命令合并HDFS小文件
hadoop fs -getmerge  
重启reduce job方法
Introduced recovery of jobs when JobTracker restarts. This facility is off by default.
Introduced config parameters "mapred.jobtracker.restart.recover", "mapred.jobtracker.job.history.block.size", and "mapred.jobtracker.job.history.buffer.size".
还未验证过。
IO写操作出现
0-1246359584298, infoPort=50075, ipcPort=50020):Got exception while serving blk_-5911099437886836280_1292 to /172.16.100.165:
java.net.SocketTimeoutException: 480000 millis timeout while waiting for channel to be ready for write. ch : java.nio.channels.SocketChannel[connected local=/
172.16.100.165:50010 remote=/172.16.100.165:50930]
        at org.apache.hadoop.net.SocketIOWithTimeout.waitForIO(SocketIOWithTimeout.java:185)
        at org.apache.hadoop.net.SocketOutputStream.waitForWritable(SocketOutputStream.java:159)
        at org.apache.hadoop.net.SocketOutputStream.transferToFully(SocketOutputStream.java:198)
        at org.apache.hadoop.hdfs.server.datanode.BlockSender.sendChunks(BlockSender.java:293)
        at org.apache.hadoop.hdfs.server.datanode.BlockSender.sendBlock(BlockSender.java:387)
        at org.apache.hadoop.hdfs.server.datanode.DataXceiver.readBlock(DataXceiver.java:179)
        at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:94)
        at java.lang.Thread.run(Thread.java:619)
It seems there are many reasons that it can timeout, the example given in
HADOOP-3831 is a slow reading client.
解决办法：在hadoop-site.xml中设置dfs.datanode.socket.write.timeout=0试试；
My understanding is that this issue should be fixed in Hadoop 0.19.1 so that
we should leave the standard timeout. However until then this can help
resolve issues like the one you're seeing.
HDFS退服节点的方法
目前版本的dfsadmin的帮助信息是没写清楚的，已经file了一个bug了，正确的方法如下：
1. 将 dfs.hosts 置为当前的 slaves，文件名用完整路径，注意，列表中的节点主机名要用大名，即 uname -n 可以得到的那个。
2. 将 slaves 中要被退服的节点的全名列表放在另一个文件里，如 slaves.ex，运用 dfs.host.exclude 参数指向这个文件的完整路径
3. 运行命令 bin/hadoop dfsadmin -refreshNodes
4. web界面或 bin/hadoop dfsadmin -report 可以看到退服节点的状态是 Decomission in progress，直到需要复制的数据复制完成为止
5. 完成之后，从 slaves 里（指 dfs.hosts 指向的文件）去掉已经退服的节点
附带说一下 -refreshNodes 命令的另外三种用途：
2. 添加允许的节点到列表中（添加主机名到 dfs.hosts 里来）
3. 直接去掉节点，不做数据副本备份（在 dfs.hosts 里去掉主机名）
4. 退服的逆操作——停止 exclude 里面和 dfs.hosts 里面都有的，正在进行 decomission 的节点的退服，也就是把 Decomission in progress 的节点重新变为 Normal （在 web 界面叫 in service)
Hadoop添加节点的方法
自己实际添加节点过程：
1. 先在slave上配置好环境，包括ssh，jdk，相关config，lib，bin等的拷贝；
2. 将新的datanode的host加到集群namenode及其他datanode中去；
3. 将新的datanode的ip加到master的conf/slaves中；
4. 重启cluster,在cluster中看到新的datanode节点；
5. 运行bin/start-balancer.sh，这个会很耗时间
备注：
1. 如果不balance，那么cluster会把新的数据都存放在新的node上，这样会降低mr的工作效率；
2. 也可调用bin/start-balancer.sh 命令运行，也可加参数 -threshold 5
   threshold 是平衡阈值，默认是10%，值越低各节点越平衡，但消耗时间也更长。
3. balancer也可以在有mr job的cluster上运行，默认dfs.balance.bandwidthPerSec很低，为1M/s。在没有mr job时，可以提高该设置加快负载均衡时间。
其他备注：
1. 必须确保slave的firewall已关闭;
2. 确保新的slave的ip已经添加到master及其他slaves的/etc/hosts中，反之也要将master及其他slave的ip添加到新的slave的/etc/hosts中
hadoop 研究借鉴
1. 解决hadoop OutOfMemoryError：
   mapred.child.java.opts
   -Xmx800M -server
With the right JVM size in your hadoop-site.xml , you will have to copy this
to all mapred nodes and restart the cluster.
或者：hadoop jar jarfile [main class] -D mapred.child.java.opts=-Xmx800M
2. Hadoop java.io.IOException: Job failed! at org.apache.hadoop.mapred.JobClient.runJob(JobClient.java:1232) while indexing.
when i use nutch1.0,get this error:
Hadoop java.io.IOException: Job failed! at org.apache.hadoop.mapred.JobClient.runJob(JobClient.java:1232) while indexing.
这个也很好解决：
可以删除conf/log4j.properties，然后可以看到详细的错误报告
我这儿出现的是out of memory
解决办法是在给运行主类org.apache.nutch.crawl.Crawl加上参数：-Xms64m -Xmx512m
你的或许不是这个，但是能看到详细的错误报告就好解决了
distribute cache运用
类似一个全局变量，但是由于这个变量较大，所以不能设置在config文件中，转而运用distribute cache
具体运用方法：(详见《the definitive guide》,P240)
1. 在命令行调用时：调用-files，引入需要查询的文件(可以是local file, HDFS file(运用hdfs://xxx?)), 或者 -archives (JAR,ZIP, tar等)
% hadoop jar job.jar MaxTemperatureByStationNameUsingDistributedCacheFile \
  -files input/ncdc/metadata/stations-fixed-width.txt input/ncdc/all output
2. 程序中调用：
   public void configure(JobConf conf) {
      metadata = new NcdcStationMetadata();
      try {
        metadata.initialize(new File("stations-fixed-width.txt"));
      } catch (IOException e) {
        throw new RuntimeException(e);
      }
   }
另外一种间接的运用方法：在hadoop-0.19.0中好像没有
调用addCacheFile()或者addCacheArchive()添加文件，
运用getLocalCacheFiles() 或 getLocalCacheArchives() 获得文件
hadoop的job显示web
There are web-based interfaces to both the JobTracker (MapReduce master) and NameNode (HDFS master) which display status pages about the state of the entire system. By default, these are located at [WWW] http://job.tracker.addr:50030/ and [WWW] http://name.node.addr:50070/.
hadoop监控
OnlyXP(52388483) 131702
用nagios作告警，ganglia作监控图表即可
status of 255 error
错误类型：
java.io.IOException: Task process exit with nonzero status of 255.
        at org.apache.hadoop.mapred.TaskRunner.run(TaskRunner.java:424)
错误原因：
Set mapred.jobtracker.retirejob.interval and mapred.userlog.retain.hours to higher value. By default, their values are 24 hours. These might be the reason for failure, though I'm not sure
split size
FileInputFormat input splits: (详见 《the definitive guide》P190)
mapred.min.split.size: default=1, the smallest valide size in bytes for a file split.
mapred.max.split.size: default=Long.MAX_VALUE, the largest valid size.
dfs.block.size: default = 64M, 系统中设置为128M。
如果设置 minimum split size > block size, 会增加块的数量。(猜想从其他节点拿去数据的时候，会合并block，导致block数量增多)
如果设置maximum split size  100% when the total size of map outputs (for a
single reducer) is high "
造成原因：
在reduce的merge过程中，check progress有误差，导致status > 100%，在统计过程中就会出现以下错误：java.lang.ArrayIndexOutOfBoundsException: 3
        at org.apache.hadoop.mapred.StatusHttpServer$TaskGraphServlet.getReduceAvarageProgresses(StatusHttpServer.java:228)
        at org.apache.hadoop.mapred.StatusHttpServer$TaskGraphServlet.doGet(StatusHttpServer.java:159)
        at javax.servlet.http.HttpServlet.service(HttpServlet.java:689)
        at javax.servlet.http.HttpServlet.service(HttpServlet.java:802)
        at org.mortbay.jetty.servlet.ServletHolder.handle(ServletHolder.java:427)
        at org.mortbay.jetty.servlet.WebApplicationHandler.dispatch(WebApplicationHandler.java:475)
        at org.mortbay.jetty.servlet.ServletHandler.handle(ServletHandler.java:567)
        at org.mortbay.http.HttpContext.handle(HttpContext.java:1565)
        at org.mortbay.jetty.servlet.WebApplicationContext.handle(WebApplicationContext.java:635)
        at org.mortbay.http.HttpContext.handle(HttpContext.java:1517)
        at org.mortbay.http.HttpServer.service(HttpServer.java:954)
jira地址：https://issues.apache.org/jira/browse/HADOOP-5210
counters
3中counters：
1. built-in counters: Map input bytes, Map output records...
2. enum counters
   调用方式：
  enum Temperature {
    MISSING,
    MALFORMED
  }

reporter.incrCounter(Temperature.MISSING, 1)
   结果显示：
09/04/20 06:33:36 INFO mapred.JobClient:   Air Temperature Recor
09/04/20 06:33:36 INFO mapred.JobClient:     Malformed=3
09/04/20 06:33:36 INFO mapred.JobClient:     Missing=66136856
3. dynamic countes:
   调用方式：
   reporter.incrCounter("TemperatureQuality", parser.getQuality(),1);
   
   结果显示：
09/04/20 06:33:36 INFO mapred.JobClient:   TemperatureQuality
09/04/20 06:33:36 INFO mapred.JobClient:     2=1246032
09/04/20 06:33:36 INFO mapred.JobClient:     1=973422173
09/04/20 06:33:36 INFO mapred.JobClient:     0=1

7: Namenode in safe mode
解决方法
bin/hadoop dfsadmin -safemode leave
8:java.net.NoRouteToHostException: No route to host
j解决方法：
sudo /etc/init.d/iptables stop
9：更改namenode后，在hive中运行select 依旧指向之前的namenode地址
这是因为：When youcreate a table, hive actually stores the location of the table (e.g.
hdfs://ip:port/user/root/...) in the SDS and DBS tables in the metastore . So when I bring up a new cluster the master has a new IP, but hive's metastore is still pointing to the locations within the old
cluster. I could modify the metastore to update with the new IP everytime I bring up a cluster. But the easier and simpler solution was to just use an elastic IP for the master
所以要将metastore中的之前出现的namenode地址全部更换为现有的namenode地址

欢迎大家 把遇到的 整理下 跟贴:lol

回复 1# 摩西莫西
  非常棒阿,解决了困扰我好久的!

回复 5# stone54321277
   平常遇到的也贴出来啊，呵呵，多交流，多研究啊:lol

本帖最后由 摩西莫西 于 2009-11-25 19:37 编辑
10：
[color=]Your DataNode is started and you can create directories with bin/hadoop dfs -mkdir, but you get an error message when you try to put files into the HDFS (e.g., when you run a command like bin/hadoop dfs -put).
解决方法：
Go to the HDFS info web page (open your web browser and go to http://namenode:dfs_info_port where namenode is the hostname of your NameNode and dfs_info_port is the port you chose dfs.info.port; if followed the QuickStart on your personal computer then this URL will be http://localhost:50070). Once at that page click on the number where it tells you how many DataNodes you have to look at a list of the DataNodes in your cluster.
If it says you have used 100% of your space, then you need to free up room on local disk(s) of the DataNode(s).
If you are on Windows then this number will not be accurate (there is some kind of bug either in Cygwin's df.exe or in Windows). Just free up some more space and you should be okay. On one Windows machine we tried the disk had 1GB free but Hadoop reported that it was 100% full. Then we freed up another 1GB and then it said that the disk was 99.15% full and started writing data into the HDFS again. We encountered this bug on Windows XP SP2.
11：Your DataNodes won't start, and you see something like this in logs/*datanode*:
Incompatible namespaceIDs in /tmp/hadoop-ross/dfs/data
原因：
Your Hadoop namespaceID became corrupted. Unfortunately the easiest thing to do reformat the HDFS.
解决方法：
You need to do something like this:
bin/stop-all.sh
rm -Rf /tmp/hadoop-your-username/*
bin/hadoop namenode -format
12：
[color=]You can run Hadoop jobs written in Java (like the grep example), but your HadoopStreaming jobs (such as the Python example that fetches web page titles) won't work.
原因：
You might have given only a relative path to the mapper and reducer programs. The tutorial originally just specified relative paths, but absolute paths are required if you are running in a real cluster.
解决方法：
Use absolute paths like this from the tutorial:
bin/hadoop jar contrib/hadoop-0.15.2-streaming.jar \
  -mapper  $HOME/proj/hadoop/multifetch.py         \
  -reducer $HOME/proj/hadoop/reducer.py            \
  -input   urls/*                                  \
  -output  titles

13：
2009-01-08 10:02:40,709 ERROR metadata.Hive (Hive.java:getPartitions(499)) - javax.jdo.JDODataStoreException: Required table missing : ""PARTITIONS"" in Catalog "" Schema "". JPOX requires this table to perform its persistence operations. Either your MetaData is incorrect, or you need to enable "org.jpox.autoCreateTables"
原因：就是因为
在 hive-default.xml 里把 org.jpox.fixedDatastore 设置成 true 了

过来研究了，看到这么多高手要更加努力研究了

回复 9# ambious
这个贴意义非常大，等多了，再整理成PDF共享。