hadoop 常見錯誤及解決方案

1：Shuffle Error: Exceeded MAX_FAILED_UNIQUE_FETCHES; bailing-out

Answer：
程序裏面須要打開多個文件，進行分析，系統通常默認數量是1024，（用ulimit -a能夠看到）對於正常使用是夠了，可是對於程序來說，就太少了。
修改辦法：
修改2個文件。
          /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

針對第一個問題我糾正下答案：
這是reduce預處理階段shuffle時獲取已完成的map的輸出失敗次數超過上限形成的，上限默認爲5。引發此問題的方式可能會有不少種，好比網絡鏈接不正常，鏈接超時，帶寬較差以及端口阻塞等。。。一般框架內網絡狀況較好是不會出現此錯誤的。

2：Too many fetch-failures
Answer:
出現這個問題主要是結點間的連通不夠全面。
1) 檢查 、/etc/hosts
要求本機ip 對應 服務器名
  要求要包含全部的服務器ip + 服務器名
2) 檢查 .ssh/authorized_keys
要求包含全部服務器（包括其自身）的public key

3：處理速度特別的慢 出現map很快 可是reduce很慢 並且反覆出現 reduce=0%
Answer:
結合第二點，而後
修改 conf/hadoop-env.sh 中的export HADOOP_HEAPSIZE=4000

4：可以啓動datanode，但沒法訪問，也沒法結束的錯誤
在 從新格式化一個新的分佈式文件時，須要將你NameNode上所配置的dfs.name.dir這一namenode用來存放NameNode  持久存儲名字空間及事務日誌的本地文件系統路徑刪除，同時將各DataNode上的dfs.data.dir的路徑 DataNode  存放塊數據的本地文件系統路徑的目錄也刪除。如本此配置就是在NameNode上刪除/home/hadoop/NameData，在DataNode上 刪除/home/hadoop/DataNode1和/home/hadoop/DataNode2。這是由於Hadoop在格式化一個新的分佈式文件系 統時，每一個存儲的名字空間都對應了創建時間的那個版本（能夠查看/home/hadoop  /NameData/current目錄下的VERSION文件，上面記錄了版本信息），在從新格式化新的分佈式系統文件時，最好先刪除NameData  目錄。必須刪除各DataNode的dfs.data.dir。這樣纔可使namedode和datanode記錄的信息版本對應。
注意：刪除是個很危險的動做，不能確認的狀況下不能刪除！！作好刪除的文件等統統備份！！

5：java.io.IOException: Could not obtain block: blk_194219614024901469_1100  file=/user/hive/warehouse/src_20090724_log/src_20090724_log
出現這種狀況大可能是結點斷了，沒有鏈接上。

6：java.lang.OutOfMemoryError: Java heap space
出現這種異常，明顯是jvm內存不夠得緣由，要修改全部的datanode的jvm內存大小。
Java -Xms1024m -Xmx4096m
通常jvm的最大內存使用應該爲總內存大小的一半，咱們使用的8G內存，因此設置爲4096m，這一值可能依舊不是最優的值。

本主題由 admin 於 2009-11-20 10:50 置頂

頂，這樣的貼子很是好，要置頂。附件是由Hadoop技術交流羣中若冰的同窗提供的相關資料：  
(12.58 KB)  
 Hadoop添加節點的方法  
本身實際添加節點過程：  
1. 先在slave上配置好環境，包括ssh，jdk，相關config，lib，bin等的拷貝；  
2. 將新的datanode的host加到集羣namenode及其餘datanode中去；  
3. 將新的datanode的ip加到master的conf/slaves中；  
4. 重啓cluster,在cluster中看到新的datanode節點；  
5. 運行bin/start-balancer.sh，這個會很耗時間  
備註：  
1. 若是不balance，那麼cluster會把新的數據都存放在新的node上，這樣會下降mr的工做效率；  
2. 也可調用bin/start-balancer.sh 命令執行，也可加參數 -threshold 5  
  threshold 是平衡閾值，默認是10%，值越低各節點越平衡，但消耗時間也更長。  
3. balancer也能夠在有mr job的cluster上運行，默認dfs.balance.bandwidthPerSec很低，爲1M/s。在沒有mr job時，能夠提升該設置加快負載均衡時間。  
 
其餘備註：  
1. 必須確保slave的firewall已關閉;  
2. 確保新的slave的ip已經添加到master及其餘slaves的/etc/hosts中，反之也要將master及其餘slave的ip添加到新的slave的/etc/hosts中  
 mapper及reducer個數  
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 << heapSize). This will be fixed at some point, but  until it is it provides a pretty firm upper bound.  
The number of  reduces also controls the number of output files in the output  directory, but usually that is not important because the next map/reduce step will split them into even smaller splits for the maps.  
The  number of reduce tasks can also be increased in the same way as the map  tasks, via JobConf's conf.setNumReduceTasks(int num).  
本身的理解：  
mapper個數的設置：跟input file 有關係，也跟filesplits有關係，filesplits的上線爲dfs.block.size，下線能夠經過mapred.min.split.size設置，最後仍是由InputFormat決定。  
 
較好的建議：  
The right number of reduces seems to be 0.95 or 1.75 multiplied by (<no. of nodes> * mapred.tasktracker.reduce.tasks.maximum).increasing the  number of reduces increases the framework overhead, but increases load  balancing and lowers the cost of failures.  
< property>  
    <name>mapred.tasktracker.reduce.tasks.maximum</name>  
    <value>2</value>  
    <description>The maximum number of reduce tasks that will be run  
    simultaneously by a task tracker.  
    </description>  
< /property>  
 
 單個node新加硬盤  
1.修改須要新加硬盤的node的dfs.data.dir，用逗號分隔新、舊文件目錄  
2.重啓dfs  
 
 同步hadoop 代碼  
hadoop-env.sh  
# host:path where hadoop code should be rsync'd from.    Unset by default.  
# export HADOOP_MASTER=master:/home/$USER/src/hadoop  
 
 用命令合併HDFS小文件  
hadoop fs -getmerge <src> <dest>  
 
 重啓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".  
還未驗證過。  
 
 IO寫操做出現問題  
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.  
 
解決辦法：在hadoop-site.xml中設置dfs.datanode.socket.write.timeout=0試試；  
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退服節點的方法  
目前版本的dfsadmin的幫助信息是沒寫清楚的，已經file了一個bug了，正確的方法以下：  
1. 將 dfs.hosts 置爲當前的 slaves，文件名用完整路徑，注意，列表中的節點主機名要用大名，即 uname -n 能夠獲得的那個。  
2. 將 slaves 中要被退服的節點的全名列表放在另外一個文件裏，如 slaves.ex，使用 dfs.host.exclude 參數指向這個文件的完整路徑  
3. 運行命令 bin/hadoop dfsadmin -refreshNodes  
4. web界面或 bin/hadoop dfsadmin -report 能夠看到退服節點的狀態是 Decomission in progress，直到須要複製的數據複製完成爲止  
5. 完成以後，從 slaves 裏（指 dfs.hosts 指向的文件）去掉已經退服的節點  
 
附帶說一下 -refreshNodes 命令的另外三種用途：  
2. 添加容許的節點到列表中（添加主機名到 dfs.hosts 裏來）  
3. 直接去掉節點，不作數據副本備份（在 dfs.hosts 裏去掉主機名）  
4.  退服的逆操做——中止 exclude 裏面和 dfs.hosts 裏面都有的，正在進行 decomission 的節點的退服，也就是把  Decomission in progress 的節點從新變爲 Normal （在 web 界面叫 in service)  
 
 hadoop 學習借鑑  
1. 解決hadoop OutOfMemoryError問題：  
<property>  
  <name>mapred.child.java.opts</name>  
  <value>-Xmx800M -server</value>  
< /property>  
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.  
這個也很好解決：  
能夠刪除conf/log4j.properties，而後能夠看到詳細的錯誤報告  
我這兒出現的是out of memory  
解決辦法是在給運行主類org.apache.nutch.crawl.Crawl加上參數：-Xms64m -Xmx512m  
你的或許不是這個問題，可是能看到詳細的錯誤報告問題就好解決了  
 
 distribute cache使用  
相似一個全局變量，可是因爲這個變量較大，因此不能設置在config文件中，轉而使用distribute cache  
具體使用方法：(詳見《the definitive guide》,P240)  
1. 在命令行調用時：調用-files，引入須要查詢的文件(能夠是local file, HDFS file(使用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. 程序中調用：  
public void configure(JobConf conf) {  
       metadata = new NcdcStationMetadata();  
       try {  
           metadata.initialize(new File("stations-fixed-width.txt"));  
       } catch (IOException e) {  
           throw new RuntimeException(e);  
       }  
  }  
另一種間接的使用方法：在hadoop-0.19.0中好像沒有  
調用addCacheFile()或者addCacheArchive()添加文件，  
使用getLocalCacheFiles() 或 getLocalCacheArchives() 得到文件  
 
 hadoop的job顯示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監控  
OnlyXP(52388483) 131702  
用nagios做告警，ganglia做監控圖表便可  
 
 status of 255 error  
錯誤類型：  
java.io.IOException: Task process exit with nonzero status of 255.  
           at org.apache.hadoop.mapred.TaskRunner.run(TaskRunner.java:424)  
 
錯誤緣由：  
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: (詳見 《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, 系統中設置爲128M。  
若是設置 minimum split size > block size, 會增長塊的數量。(猜測從其餘節點拿去數據的時候，會合並block，致使block數量增多)  
若是設置maximum split size < block size, 會進一步拆分block。  
 
split size = max(minimumSize, min(maximumSize, blockSize));  
其中 minimumSize < blockSize < maximumSize.  
 
 sort by value  
hadoop 不提供直接的sort by value方法，由於這樣會下降mapreduce性能。  
但能夠用組合的辦法來實現，具體實現方法見《the definitive guide》, P250  
基本思想：  
1. 組合key/value做爲新的key；  
2. 重載partitioner，根據old key來分割；  
conf.setPartitionerClass(FirstPartitioner.class);  
3. 自定義keyComparator：先根據old key排序，再根據old value排序；  
conf.setOutputKeyComparatorClass(KeyComparator.class);  
4. 重載GroupComparator, 也根據old key 來組合；    conf.setOutputValueGroupingComparator(GroupComparator.class);  
 
 small input files的處理  
對於一系列的small files做爲input file，會下降hadoop效率。  
有3種方法能夠將small file合併處理：  
1. 將一系列的small files合併成一個sequneceFile，加快mapreduce速度。  
詳見WholeFileInputFormat及SmallFilesToSequenceFileConverter,《the definitive guide》, P194  
2. 使用CombineFileInputFormat集成FileinputFormat，可是未實現過；  
3. 使用hadoop archives(相似打包)，減小小文件在namenode中的metadata內存消耗。(這個方法不必定可行，因此不建議使用)  
方法：  
  將/my/files目錄及其子目錄歸檔成files.har，而後放在/my目錄下  
bin/hadoop archive -archiveName files.har /my/files /my  
 
查看files in the archive:  
  bin/hadoop fs -lsr har://my/files.har  
 
 skip bad records  
JobConf conf = new JobConf(ProductMR.class);  
conf.setJobName("ProductMR");  
conf.setOutputKeyClass(Text.class);  
conf.setOutputValueClass(Product.class);  
conf.setMapperClass(Map.class);  
conf.setReducerClass(Reduce.class);  
conf.setMapOutputCompressorClass(DefaultCodec.class);  
conf.setInputFormat(SequenceFileInputFormat.class);  
conf.setOutputFormat(SequenceFileOutputFormat.class);  
String objpath = "abc1";  
SequenceFileInputFormat.addInputPath(conf, new Path(objpath));  
SkipBadRecords.setMapperMaxSkipRecords(conf, Long.MAX_VALUE);  
SkipBadRecords.setAttemptsToStartSkipping(conf, 0);  
SkipBadRecords.setSkipOutputPath(conf, new Path("data/product/skip/"));  
String output = "abc";  
SequenceFileOutputFormat.setOutputPath(conf, new Path(output));  
JobClient.runJob(conf);  
 
For skipping failed tasks try : mapred.max.map.failures.percent  
 
 restart 單個datanode  
若是一個datanode 出現問題，解決以後須要從新加入cluster而不重啓cluster，方法以下：  
bin/hadoop-daemon.sh start datanode  
bin/hadoop-daemon.sh start jobtracker  
 
 reduce exceed 100%  
"Reduce Task Progress shows > 100% when the total size of map outputs (for a  
single reducer) is high "  
形成緣由：  
在reduce的merge過程當中，check progress有偏差，致使status > 100%，在統計過程當中就會出現如下錯誤：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：  
 
 counters  
3中counters：  
1. built-in counters: Map input bytes, Map output records...  
2. enum counters  
調用方式：  
    enum Temperature {  
   MISSING,  
   MALFORMED  
    }  
 
reporter.incrCounter(Temperature.MISSING, 1)  
結果顯示：  
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:  
調用方式：  
reporter.incrCounter("TemperatureQuality", parser.getQuality(),1);  
 
結果顯示：  
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  
解決方法  
bin/hadoop dfsadmin -safemode leave  
 
 8:java.net.NoRouteToHostException: No route to host  
j解決方法：  
sudo /etc/init.d/iptables stop  
 
 9：更改namenode後，在hive中運行select 依舊指向以前的namenode地址  
這是由於：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  
因此要將metastore中的以前出現的namenode地址所有更換爲現有的namenode地址

 
 10：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).  
解決方法：  
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  
緣由：  
Your Hadoop namespaceID became corrupted. Unfortunately the easiest thing to do reformat the HDFS.  
解決方法：  
You need to do something like this:  
bin/stop-all.sh  
rm -Rf /tmp/hadoop-your-username/*  
bin/hadoop namenode -format  
 12：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.  
緣由：  
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.  
解決方法：  
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"  
緣由：就是由於在 hive-default.xml 裏把 org.jpox.fixedDatastore 設置成 true 了

starting namenode, logging to /home/hadoop/HadoopInstall/hadoop/bin/../logs/hadoop-hadoop-namenode-hadoop.out  
localhost: starting datanode, logging to /home/hadoop/HadoopInstall/hadoop/bin/../logs/hadoop-hadoop-datanode-hadoop.out  
localhost: starting secondarynamenode, logging to  /home/hadoop/HadoopInstall/hadoop/bin/../logs/hadoop-hadoop-secondarynamenode-hadoop.out  
localhost: Exception in thread "main" java.lang.NullPointerException  
localhost:        at org.apache.hadoop.net.NetUtils.createSocketAddr(NetUtils.java:130)  
localhost:        at org.apache.hadoop.dfs.NameNode.getAddress(NameNode.java:116)  
localhost:        at org.apache.hadoop.dfs.NameNode.getAddress(NameNode.java:120)  
localhost:        at org.apache.hadoop.dfs.SecondaryNameNode.initialize(SecondaryNameNode.java:124)  
localhost:        at org.apache.hadoop.dfs.SecondaryNameNode.<init>(SecondaryNameNode.java:108)  
localhost:        at org.apache.hadoop.dfs.SecondaryNameNode.main(SecondaryNameNode.java:460)

14：09/08/31 18:25:45 INFO hdfs.DFSClient: Exception in  createBlockOutputStream java.io.IOException:Bad connect ack with  firstBadLink 192.168.1.11:50010  
> 09/08/31 18:25:45 INFO hdfs.DFSClient: Abandoning block blk_-8575812198227241296_1001  
> 09/08/31 18:25:51 INFO hdfs.DFSClient: Exception in createBlockOutputStream java.io.IOException:  
Bad connect ack with firstBadLink 192.168.1.16:50010  
> 09/08/31 18:25:51 INFO hdfs.DFSClient: Abandoning block blk_-2932256218448902464_1001  
> 09/08/31 18:25:57 INFO hdfs.DFSClient: Exception in createBlockOutputStream java.io.IOException:  
Bad connect ack with firstBadLink 192.168.1.11:50010  
> 09/08/31 18:25:57 INFO hdfs.DFSClient: Abandoning block blk_-1014449966480421244_1001  
> 09/08/31 18:26:03 INFO hdfs.DFSClient: Exception in createBlockOutputStream java.io.IOException:  
Bad connect ack with firstBadLink 192.168.1.16:50010  
> 09/08/31 18:26:03 INFO hdfs.DFSClient: Abandoning block blk_7193173823538206978_1001  
> 09/08/31 18:26:09 WARN hdfs.DFSClient: DataStreamer Exception: java.io.IOException: Unable  
to create new block.  
>             at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.nextBlockOutputStream(DFSClient.java:2731)  
>             at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.access$2000(DFSClient.java:1996)  
>             at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream$DataStreamer.run(DFSClient.java:2182)  
>    
> 09/08/31 18:26:09 WARN hdfs.DFSClient: Error Recovery for block blk_7193173823538206978_1001  
bad datanode[2] nodes == null  
> 09/08/31 18:26:09 WARN hdfs.DFSClient: Could not get block locations. Source file "/user/umer/8GB_input"  
- Aborting...  
> put: Bad connect ack with firstBadLink 192.168.1.16:50010  
 
 
解決方法：  
I have resolved the issue:  
What i did:  
 
1) '/etc/init.d/iptables stop' -->stopped firewall  
2) SELINUX=disabled in '/etc/selinux/config' file.-->disabled selinux  
I worked for me after these two changes

解決jline.ConsoleReader.readLine在Windows上不生效問題方法  
在 CliDriver.java的main()函數中，有一條語句reader.readLine，用來讀取標準輸入，但在Windows平臺上該語句老是 返回null，這個reader是一個實例jline.ConsoleReader實例，給Windows Eclipse調試帶來不便。  
咱們能夠經過使用java.util.Scanner.Scanner來替代它，將原來的  
while ((line=reader.readLine(curPrompt+"> ")) != null)  
複製代碼  
替換爲：  
Scanner sc = new Scanner(System.in);  
while ((line=sc.nextLine()) != null)  
複製代碼  
從新編譯發佈，便可正常從標準輸入讀取輸入的SQL語句了。  
 
 Windows eclispe調試hive報does not have a scheme錯誤可能緣由  
一、Hive配置文件中的「hive.metastore.local」配置項值爲false，須要將它修改成true，由於是單機版  
二、沒有設置HIVE_HOME環境變量，或設置錯誤  
三、 「does not have a  scheme」極可能是由於找不到「hive-default.xml」。使用Eclipse調試Hive時，遇到找不到hive- default.xml的解決方法：http://bbs.hadoopor.com/thread-292-1-1.html

一、中文問題  
   從url中解析出中文,但hadoop中打印出來還是亂碼?咱們曾經覺得hadoop是不支持中文的，後來通過查看源代碼，發現hadoop僅僅是不支持以gbk格式輸出中文而己。  
    這是TextOutputFormat.class中的代碼，hadoop默認的輸出都是繼承自FileOutputFormat來 的，FileOutputFormat的兩個子類一個是基於二進制流的輸出，一個就是基於文本的輸出TextOutputFormat。  
public class TextOutputFormat<K, V> extends FileOutputFormat<K, V> {  
    protected static class LineRecordWriter<K, V>  
   implements RecordWriter<K, V> {  
   private static final String utf8 = 「UTF-8″;//這裏被寫死成了utf-8  
   private static final byte[] newline;  
   static {  
       try {  
           newline = 「\n」.getBytes(utf8);  
       } catch (UnsupportedEncodingException uee) {  
           throw new IllegalArgumentException(」can’t find 」 + utf8 + 」 encoding」);  
       }  
   }  
…  
public LineRecordWriter(DataOutputStream out, String keyValueSeparator) {  
       this.out = out;  
       try {  
           this.keyValueSeparator = keyValueSeparator.getBytes(utf8);  
       } catch (UnsupportedEncodingException uee) {  
           throw new IllegalArgumentException(」can’t find 」 + utf8 + 」 encoding」);  
       }  
   }  
…  
private void writeObject(Object o) throws IOException {  
       if (o instanceof Text) {  
           Text to = (Text) o;  
           out.write(to.getBytes(), 0, to.getLength());//這裏也須要修改  
      } else {  
           out.write(o.toString().getBytes(utf8));  
       }  
   }  
…  
}  
能夠看出hadoop默認的輸出寫死爲utf-8，所以若是decode中文正確，那麼將Linux客戶端的character設爲utf-8是能夠看到中文的。由於hadoop用utf-8的格式輸出了中文。  
   由於大多數數據庫是用gbk來定義字段的，若是想讓hadoop用gbk格式輸出中文以兼容數據庫怎麼辦？  
   咱們能夠定義一個新的類：  
public class GbkOutputFormat<K, V> extends FileOutputFormat<K, V> {  
    protected static class LineRecordWriter<K, V>  
   implements RecordWriter<K, V> {  
//寫成gbk便可  
private static final String gbk = 「gbk」;  
private static final byte[] newline;  
   static {  
       try {  
           newline = 「\n」.getBytes(gbk);  
       } catch (UnsupportedEncodingException uee) {  
           throw new IllegalArgumentException(」can’t find 」 + gbk + 」 encoding」);  
       }  
   }  
…  
public LineRecordWriter(DataOutputStream out, String keyValueSeparator) {  
       this.out = out;  
       try {  
           this.keyValueSeparator = keyValueSeparator.getBytes(gbk);  
       } catch (UnsupportedEncodingException uee) {  
           throw new IllegalArgumentException(」can’t find 」 + gbk + 」 encoding」);  
       }  
   }  
…  
private void writeObject(Object o) throws IOException {  
       if (o instanceof Text) {  
//            Text to = (Text) o;  
//            out.write(to.getBytes(), 0, to.getLength());  
//        } else {  
           out.write(o.toString().getBytes(gbk));  
       }  
   }  
…  
}  
而後在mapreduce代碼中加入conf1.setOutputFormat(GbkOutputFormat.class)  
便可以gbk格式輸出中文。  
 
二、某次正常運行mapreduce實例時,拋出錯誤  
 
java.io.IOException: All datanodes xxx.xxx.xxx.xxx:xxx are bad. Aborting…  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream.processDatanodeError(DFSClient.java:2158)  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream.access$1400(DFSClient.java:1735)  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream$DataStreamer.run(DFSClient.java:1889)  
java.io.IOException: Could not get block locations. Aborting…  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream.processDatanodeError(DFSClient.java:2143)  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream.access$1400(DFSClient.java:1735)  
at org.apache.hadoop.dfs.DFSClient$DFSOutputStream$DataStreamer.run(DFSClient.java:1889)  
經查明，問題緣由是linux機器打開了過多的文件致使。用命令ulimit -n能夠發現linux默認的文件打開數目爲1024，修改/ect/security/limit.conf，增長hadoop soft 65535  
 
再從新運行程序（最好全部的datanode都修改），問題解決  
 
三、運行一段時間後hadoop不能stop-all.sh的問題，顯示報錯  
no tasktracker to stop ，no datanode to stop  
問 題的緣由是hadoop在stop的時候依據的是datanode上的mapred和dfs進程號。而默認的進程號保存在/tmp下，linux默認會每 隔一段時間（通常是一個月或者7天左右）去刪除這個目錄下的文件。所以刪掉hadoop-hadoop-jobtracker.pid和hadoop- hadoop-namenode.pid兩個文件後，namenode天然就找不到datanode上的這兩個進程了。  
在配置文件中的export HADOOP_PID_DIR能夠解決這個問題  

 
問題：  
Incompatible namespaceIDs in /usr/local/hadoop/dfs/data: namenode namespaceID = 405233244966; datanode namespaceID = 33333244  
緣由：  
在 每次執行hadoop namenode  -format時，都會爲NameNode生成namespaceID,，可是在hadoop.tmp.dir目錄下的DataNode仍是保留上次的 namespaceID，由於namespaceID的不一致，而致使DataNode沒法啓動，因此只要在每次執行hadoop namenode  -format以前，先刪除hadoop.tmp.dir目錄就能夠啓動成功。請注意是刪除hadoop.tmp.dir對應的本地目錄，而不是HDFS 目錄。

Problem: Storage directory not exist  
2010-02-09 21:37:53,203 INFO org.apache.hadoop.hdfs.server.common.Storage: Storage directory D:\hadoop\run\dfs_name_dir does not exist.  
2010-02-09 21:37:53,203 ERROR org.apache.hadoop.hdfs.server.namenode.FSNamesystem: FSNamesystem initialization failed.  
org.apache.hadoop.hdfs.server.common.InconsistentFSStateException: Directory D:\hadoop\run\dfs_name_dir is in an inconsistent state:  storage directory does not exist or is not accessible.  
solution: 是由於存儲目錄D:\hadoop\run\dfs_name_dir不存在，因此只須要手動建立好這個目錄便可。  
 Problem: NameNode is not formatted  
solution: 是由於HDFS尚未格式化，只須要運行hadoop namenode -format一下，而後再啓動便可

bin/hadoop jps後報以下異常： Exception in thread "main" java.lang.NullPointerException            at sun.jvmstat.perfdata.monitor.protocol.local.LocalVmManager.activeVms(LocalVmManager.java:127)            at sun.jvmstat.perfdata.monitor.protocol.local.MonitoredHostProvider.activeVms(MonitoredHostProvider.java:133)            at sun.tools.jps.Jps.main(Jps.java:45)緣由爲： 系統根目錄/tmp文件夾被刪除了。從新創建/tmp文件夾便可。bin/hive中出現 unable to    create log directory /tmp/...也多是這個緣由