[ https://issues.apache.org/jira/browse/HDFS-1459?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Hajo Nils Krabbenhöft updated HDFS-1459: ---------------------------------------- Description: First, here's my source code accessing the HDFS: final FSDataInputStream indexFile = getFile(bucketPathStr, Integer.toString(hashTableId) + ".index"); indexFile.seek(bucketId * 4); int bucketStart = ByteSwapper.swap(indexFile.readInt()); int bucketEnd = ByteSwapper.swap(indexFile.readInt()); final FSDataInputStream dataFile = getFile(bucketPathStr, Integer.toString(hashTableId) + ".data"); dataFile.seek(bucketStart * (2 + Hasher.getConfigHashLength()) * 4); for (int hash = bucketStart; hash < bucketEnd; hash++) { int RimageIdA = ByteSwapper.swap(dataFile.readInt()); int RimageIdB = ByteSwapper.swap(dataFile.readInt()); ....... read hash of length Hasher.getConfigHashLength() and work with it .... } As you can see, i am reading the range to be read from an X.index file and then read these rows from X.data. The index file is always exactly 6.710.888 bytes in length. As for the data file, everything works fine with 50 different 1.35 GB (22 blocks) data files and it fails every time i tried with 50 different 2.42 GB (39 blocks) data files. So the cause of the bug is clearly dependent on the file size. I checked for ulimit and for the number of network connections and they are both not maxed out when the error occurs. The stack trace i get is: java.lang.NullPointerException at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.readBuffer(DFSClient.java:1703) at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1755) at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1680) at java.io.DataInputStream.readInt(DataInputStream.java:370) ... at org.apache.hadoop.mapred.MapRunner.run(MapRunner.java:50) at org.apache.hadoop.mapred.MapTask.runOldMapper(MapTask.java:358) at org.apache.hadoop.mapred.MapTask.run(MapTask.java:307) at org.apache.hadoop.mapred.Child.main(Child.java:170) which leads me to believe that DFSClient.blockSeekTo returns with a non-null chosenNode but with blockReader = null. Since the exact same jar works flawlessly with small data files and fails reliably with big data files, i'm wondering how this could possibly dependent on the file's size or block count (DFSClient.java line 1628+): s = socketFactory.createSocket(); NetUtils.connect(s, targetAddr, socketTimeout); s.setSoTimeout(socketTimeout); Block blk = targetBlock.getBlock(); blockReader = BlockReader.newBlockReader(s, src, blk.getBlockId(), blk.getGenerationStamp(), offsetIntoBlock, blk.getNumBytes() - offsetIntoBlock, buffersize, verifyChecksum, clientName); return chosenNode; was: First, here's my source code accessing the HDFS: final FSDataInputStream indexFile = getFile(bucketPathStr, Integer.toString(hashTableId) + ".index"); indexFile.seek(bucketId * 4); int bucketStart = ByteSwapper.swap(indexFile.readInt()); int bucketEnd = ByteSwapper.swap(indexFile.readInt()); final FSDataInputStream dataFile = getFile(bucketPathStr, Integer.toString(hashTableId) + ".data"); dataFile.seek(bucketStart * (2 + Hasher.getConfigHashLength()) * 4); for (int hash = bucketStart; hash < bucketEnd; hash++) { int RimageIdA = ByteSwapper.swap(dataFile.readInt()); int RimageIdB = ByteSwapper.swap(dataFile.readInt()); ....... read hash of length Hasher.getConfigHashLength() and work with it .... } As you can see, i am reading the range to be read from an X.index file and then read these rows from X.data. The index file is always exactly 6.710.888 bytes in length. As for the data file, everything works fine with 50 different 1.35 GB (22 blocks) data files and it fails every time i tried with 50 different 2.42 GB (39 blocks) data files. So the cause of the bug is clearly dependent on the file size. I checked for ulimit and for the number of network connections and they are both not maxed out when the error occurs. The stack trace i get is: java.lang.NullPointerException at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.readBuffer(DFSClient.java:1703) at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1755) at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1680) at java.io.DataInputStream.readInt(DataInputStream.java:370) ... at org.apache.hadoop.mapred.MapRunner.run(MapRunner.java:50) at org.apache.hadoop.mapred.MapTask.runOldMapper(MapTask.java:358) at org.apache.hadoop.mapred.MapTask.run(MapTask.java:307) at org.apache.hadoop.mapred.Child.main(Child.java:170) which leads me to believe that DFSClient.blockSeekTo returns with a non-null chosenNode but with blockReader = null. Since the exact same jar works flawlessly with small data files and fails reliably with big data files, i'm wondering how this could possibly dependent on the file's size or block count (DFSClient.java line 1628+): s = socketFactory.createSocket(); NetUtils.connect(s, targetAddr, socketTimeout); s.setSoTimeout(socketTimeout); Block blk = targetBlock.getBlock(); blockReader = BlockReader.newBlockReader(s, src, blk.getBlockId(), blk.getGenerationStamp(), offsetIntoBlock, blk.getNumBytes() - offsetIntoBlock, buffersize, verifyChecksum, clientName); return chosenNode; Summary: NullPointerException in DataInputStream.readInt caused by reaching xceiverCount (was: NullPointerException in DataInputStream.readInt) The NPE on client side is caused by this on datanode side: 2010-10-20 15:31:17,177 ERROR org.apache.hadoop.hdfs.server.datanode.DataNode: DatanodeRegistration(10.17.5.3:50010, storageID=DS-266784496-78.46.65.54-50010-1287004808819, infoPort=50075, ipcPort=50020):DataXceiver java.io.IOException: xceiverCount 257 exceeds the limit of concurrent xcievers 256 at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:88) at java.lang.Thread.run(Thread.java:619) > NullPointerException in DataInputStream.readInt caused by reaching > xceiverCount > ------------------------------------------------------------------------------- > > Key: HDFS-1459 > URL: https://issues.apache.org/jira/browse/HDFS-1459 > Project: Hadoop HDFS > Issue Type: Bug > Affects Versions: 0.20.1 > Environment: Debian 64 bit > Cloudera Hadoop > Reporter: Hajo Nils Krabbenhöft > > First, here's my source code accessing the HDFS: > final FSDataInputStream indexFile = getFile(bucketPathStr, > Integer.toString(hashTableId) + ".index"); > indexFile.seek(bucketId * 4); > int bucketStart = ByteSwapper.swap(indexFile.readInt()); > int bucketEnd = ByteSwapper.swap(indexFile.readInt()); > final FSDataInputStream dataFile = getFile(bucketPathStr, > Integer.toString(hashTableId) + ".data"); > dataFile.seek(bucketStart * (2 + Hasher.getConfigHashLength()) * 4); > for (int hash = bucketStart; hash < bucketEnd; hash++) { > int RimageIdA = ByteSwapper.swap(dataFile.readInt()); > int RimageIdB = ByteSwapper.swap(dataFile.readInt()); > ....... read hash of length Hasher.getConfigHashLength() and work with > it .... > } > As you can see, i am reading the range to be read from an X.index file and > then read these rows from X.data. The index file is always exactly 6.710.888 > bytes in length. > As for the data file, everything works fine with 50 different 1.35 GB (22 > blocks) data files and it fails every time i tried with 50 different 2.42 GB > (39 blocks) data files. So the cause of the bug is clearly dependent on the > file size. > I checked for ulimit and for the number of network connections and they are > both not maxed out when the error occurs. The stack trace i get is: > java.lang.NullPointerException > at > org.apache.hadoop.hdfs.DFSClient$DFSInputStream.readBuffer(DFSClient.java:1703) > at > org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1755) > at > org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1680) > at java.io.DataInputStream.readInt(DataInputStream.java:370) > ... > at org.apache.hadoop.mapred.MapRunner.run(MapRunner.java:50) > at org.apache.hadoop.mapred.MapTask.runOldMapper(MapTask.java:358) > at org.apache.hadoop.mapred.MapTask.run(MapTask.java:307) > at org.apache.hadoop.mapred.Child.main(Child.java:170) > which leads me to believe that DFSClient.blockSeekTo returns with a non-null > chosenNode but with blockReader = null. > Since the exact same jar works flawlessly with small data files and fails > reliably with big data files, i'm wondering how this could possibly dependent > on the file's size or block count (DFSClient.java line 1628+): > s = socketFactory.createSocket(); > NetUtils.connect(s, targetAddr, socketTimeout); > s.setSoTimeout(socketTimeout); > Block blk = targetBlock.getBlock(); > blockReader = BlockReader.newBlockReader(s, src, blk.getBlockId(), > blk.getGenerationStamp(), > offsetIntoBlock, blk.getNumBytes() - offsetIntoBlock, > buffersize, verifyChecksum, clientName); > return chosenNode; -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.