hadoop中的一次集群任务执行超时问题查找过程

问题背景

本次进行一个项目的重构，在某些活动数据量比较大的情况下，会偶尔出现1200s超时的情况，如下：

AttemptID:attempt_1410771599055_11709_m_000033_0 Timed out after 1200 secs

而hadoop会不断启动备份任务进行重试，重试也许成功，但失败的概率还是比较大：

经过分析，hadoop的任务都有个超时时间，使用下面的参数设置，表示1200s后如果没有进展，就会任务该任务超时，将其状态设置为FAILED。

-Dmapreduce.task.timeout=1200000

到底因为什么原因导致超时？为了继续分析这个问题，我们将这个参数设置得非常之大。

调整超时参数

将超时时间设置为24小时之后，发现任务不会FAILED，但是其执行了大概40多个小时，仍然还没有执行完成。

还好我们在任务执行过程中打了不少的log，以帮助分析问题。经过日志的分析，我们发现有下面的现象：

2014-09-22 00:17:29,005 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 77477) is collected!
2014-09-22 00:17:29,005 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 77477) is collected!
2014-09-22 00:17:29,005 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 77477) is collected!
2014-09-22 01:17:29,054 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 120096) is collected!
2014-09-22 01:17:29,064 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 120096) is collected!
2014-09-22 01:17:29,064 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 120096) is collected!
2014-09-22 01:17:29,064 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 120096) is collected!
...
2014-09-22 01:17:36,590 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 164747) is collected!
2014-09-22 01:17:36,590 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 164747) is collected!
2014-09-22 01:17:36,590 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 164747) is collected!
2014-09-22 01:17:36,590 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 164747) is collected!
2014-09-22 02:17:36,674 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 158198) is collected!
2014-09-22 02:17:36,683 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 158198) is collected!
2014-09-22 02:17:36,683 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 158198) is collected!
2014-09-22 02:17:36,683 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 158198) is collected!
....
2014-09-22 02:17:40,888 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 203233) is collected!
2014-09-22 02:17:40,888 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 203233) is collected!
2014-09-22 02:17:40,888 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 203233) is collected!
2014-09-22 03:17:40,925 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 79188) is collected!
2014-09-22 03:17:40,934 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 79188) is collected!
2014-09-22 03:17:40,934 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 79188) is collected!
2014-09-22 03:17:40,934 INFO [main] com.xxx.yo.phase1.Phase1Mapper: history(caid: 2000037, superid: 79188) is collected!

日志分析得出的结论便是，程序总会在某个时间点休息3600秒（大概1个小时），然后再执行一会儿，便又休息3600秒。

hadoop configuration中得出初步结论

我们在这1个小时对该java进程进行监控，发现该进程在此期间(jstack命令查看其日志)，一直在一个点等待：

"main" prio=10 tid=0x000000000293f000 nid=0x1e06 runnable [0x0000000041b20000]
   java.lang.Thread.State: RUNNABLE
at sun.nio.ch.EPollArrayWrapper.epollWait(Native Method)
at sun.nio.ch.EPollArrayWrapper.poll(EPollArrayWrapper.java:228)
at sun.nio.ch.EPollSelectorImpl.doSelect(EPollSelectorImpl.java:81)
at sun.nio.ch.SelectorImpl.lockAndDoSelect(SelectorImpl.java:87)
- locked <0x00000006e243c3f0> (a sun.nio.ch.Util$2)
- locked <0x00000006e243c3e0> (a java.util.Collections$UnmodifiableSet)
- locked <0x00000006e243c1a0> (a sun.nio.ch.EPollSelectorImpl)
at sun.nio.ch.SelectorImpl.select(SelectorImpl.java:98)
at org.apache.hadoop.net.SocketIOWithTimeout$SelectorPool.select(SocketIOWithTimeout.java:335)
at org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:157)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:161)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.readChannelFully(PacketReceiver.java:258)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doReadFully(PacketReceiver.java:209)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doRead(PacketReceiver.java:171)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.receiveNextPacket(PacketReceiver.java:102)
at org.apache.hadoop.hdfs.RemoteBlockReader2.readNextPacket(RemoteBlockReader2.java:170)
at org.apache.hadoop.hdfs.RemoteBlockReader2.read(RemoteBlockReader2.java:135)
- locked <0x00000006e12dcc78> (a org.apache.hadoop.hdfs.RemoteBlockReader2)
at org.apache.hadoop.hdfs.DFSInputStream$ByteArrayStrategy.doRead(DFSInputStream.java:642)
at org.apache.hadoop.hdfs.DFSInputStream.readBuffer(DFSInputStream.java:698)
- eliminated <0x00000006e12dcc18> (a org.apache.hadoop.hdfs.DFSInputStream)
at org.apache.hadoop.hdfs.DFSInputStream.readWithStrategy(DFSInputStream.java:752)
at org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:793)
- locked <0x00000006e12dcc18> (a org.apache.hadoop.hdfs.DFSInputStream)
at java.io.DataInputStream.read(DataInputStream.java:149)
at com.xxx.app.MzSequenceFile$PartInputStream.read(MzSequenceFile.java:451)
at org.apache.hadoop.io.compress.DecompressorStream.getCompressedData(DecompressorStream.java:159)
at org.apache.hadoop.io.compress.DecompressorStream.decompress(DecompressorStream.java:143)
at org.apache.hadoop.io.compress.DecompressorStream.read(DecompressorStream.java:85)
at java.io.DataInputStream.readFully(DataInputStream.java:195)
at org.apache.hadoop.io.Text.readFields(Text.java:292)
at org.apache.hadoop.io.serializer.WritableSerialization$WritableDeserializer.deserialize(WritableSerialization.java:71)
at org.apache.hadoop.io.serializer.WritableSerialization$WritableDeserializer.deserialize(WritableSerialization.java:42)
at com.xxx.app.MzSequenceFile$Reader.deserializeValue(MzSequenceFile.java:672)
at com.xxx.app.MzSequenceFile$Reader.next(MzSequenceFile.java:684)
at com.xxx.app.MzSequenceFile$Reader.next(MzSequenceFile.java:692)
at com.xxx.yo.io.CombineFileRawLogReader.streamNext(CombineFileRawLogReader.java:284)
at com.xxx.yo.io.CombineFileRawLogReader.next(CombineFileRawLogReader.java:342)
at com.xxx.yo.io.CampaignRawLogReader.next(CampaignRawLogReader.java:73)
at com.xxx.yo.io.CampaignRawLogReader.next(CampaignRawLogReader.java:23)
at org.apache.hadoop.mapred.MapTask$TrackedRecordReader.moveToNext(MapTask.java:197)
- locked <0x00000006e01dd3e0> (a org.apache.hadoop.mapred.MapTask$TrackedRecordReader)
at org.apache.hadoop.mapred.MapTask$TrackedRecordReader.next(MapTask.java:183)
- locked <0x00000006e01dd3e0> (a org.apache.hadoop.mapred.MapTask$TrackedRecordReader)
at org.apache.hadoop.mapred.MapRunner.run(MapRunner.java:52)
at org.apache.hadoop.mapred.MapTask.runOldMapper(MapTask.java:429)
at org.apache.hadoop.mapred.MapTask.run(MapTask.java:341)
at org.apache.hadoop.mapred.YarnChild$2.run(YarnChild.java:162)
at java.security.AccessController.doPrivileged(Native Method)
at javax.security.auth.Subject.doAs(Subject.java:415)
at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1491)

当时初步怀疑是JDK版本问题，Java NIO也确实存在着臭名昭著的epoll空轮询导致CPU100%问题，但这个bug在JDK6中的高版本就已经解决了，更何况我们使用的是1.7。而且我们通过top -p 进程id的方式查看其CPU占用率为0，排除了该bug。

经过日志的初步分析，发现3600s这个线索，从job的configuration中，初步查找出参数dfs.client.socket-timeout，单位毫秒。

-Ddfs.client.socket-timeout=3600000

试验性地将这个参数修改为60ms，可以看出出现超时的概率非常大，但会不断重试以继续:

2014-09-26 12:53:03,184 WARN [main] org.apache.hadoop.hdfs.DFSClient: Failed to connect to /192.168.7.22:50010 for block, add to deadNodes and continue.
 java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch :
 java.nio.channels.SocketChannel[connected local=/192.168.7.17:22051 remote=/192.168.7.22:50010]
 java.net.SocketTimeoutException: 60 millis timeout while waiting for channel to be ready for read. ch :
 java.nio.channels.SocketChannel[connected local=/192.168.7.17:22051 remote=/192.168.7.22:50010]
at org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:164)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:161)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:131)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:118)
at java.io.FilterInputStream.read(FilterInputStream.java:83)
at org.apache.hadoop.hdfs.protocolPB.PBHelper.vintPrefixed(PBHelper.java:1490)
at org.apache.hadoop.hdfs.RemoteBlockReader2.newBlockReader(RemoteBlockReader2.java:392)
at org.apache.hadoop.hdfs.BlockReaderFactory.newBlockReader(BlockReaderFactory.java:131)
at org.apache.hadoop.hdfs.DFSInputStream.getBlockReader(DFSInputStream.java:1108)
at org.apache.hadoop.hdfs.DFSInputStream.blockSeekTo(DFSInputStream.java:533)
at org.apache.hadoop.hdfs.DFSInputStream.readWithStrategy(DFSInputStream.java:749)
at org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:793)
at org.apache.hadoop.hdfs.DFSInputStream.read(DFSInputStream.java:601)
at java.io.DataInputStream.readInt(DataInputStream.java:387)
at com.xxx.app.MzSequenceFile$Reader.init(MzSequenceFile.java:521)
at com.xxx.app.MzSequenceFile$Reader.<init>(MzSequenceFile.java:515)
at com.xxx.app.MzSequenceFile$Reader.<init>(MzSequenceFile.java:505)
at com.xxx.yo.io.CombineFileRawLogReader.<init>(CombineFileRawLogReader.java:146)
at com.xxx.yo.io.CampaignRawLogReader.next(CampaignRawLogReader.java:64)
at com.xxx.yo.io.CampaignRawLogReader.next(CampaignRawLogReader.java:22)
at org.apache.hadoop.mapred.MapTask$TrackedRecordReader.moveToNext(MapTask.java:197)
at org.apache.hadoop.mapred.MapTask$TrackedRecordReader.next(MapTask.java:183)
at org.apache.hadoop.mapred.MapRunner.run(MapRunner.java:52)
at org.apache.hadoop.mapred.MapTask.runOldMapper(MapTask.java:429)
at org.apache.hadoop.mapred.MapTask.run(MapTask.java:341)
at org.apache.hadoop.mapred.YarnChild$2.run(YarnChild.java:162)
at java.security.AccessController.doPrivileged(Native Method)
at javax.security.auth.Subject.doAs(Subject.java:415)
at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1491)
at org.apache.hadoop.mapred.YarnChild.main(YarnChild.java:157)

于是，最终将这个参数设置为60s，这其实也是集群最终默认的超时时间，由于之前一次不明就里的优化，导致了后续这些问题的发生，因此在调整参数时，一定要注意了解清楚该参数造成的影响。

简要分析的结论

在Mapper端读取HDFS上的文件时，可能由于网络原因（由于我们的Split切分地比较大，因此不可能做到完全数据本地化）导致读取数据超时，原来居然设置成1个小时，而任务的超时时间仅设置为20分钟，因此只要发生读取数据超时，就必然会引起任务超时。

通过这次分析过程，学到了很多查找问题的方式，包括通过现象分析规律，得到线索，最终查找问题的原因。快速测试，不能忽略哪怕一个小的Exception，不行就是分析hadoop的源码，掌握其运行时行为。

对于悬挂的task，如果task tracker在一段时间（默认是10min，可以通过mapred.task.timeout属性值来设置，单位是毫秒）内一直没有收到它的进度报告，则把它标记为失效，这时候就会出现前面所说的超时错误。

TaskTracker通过心跳包告知JobTracker某个task attempt失败了，则JobTracker把该task尽量分配给另外一个task tracker来执行。如果同一个task连续4次（该值可以通过mapred.map.max.attempts和mapred.reduce.max.attempts属性值来设置）都执行失败，那JobTracker就不会再做更多的尝试了，本次job也就宣告失败了。