[MAPREDUCE-1144] JT should not hold lock while writing user history logs to DFS

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Affects Version/s: 0.20.1
• Component/s: jobtracker
• Assignee: Unassigned
• Reporter: Todd Lipcon
• Created: 23/Oct/09 14:11
• Updated: 19/Mar/15 23:14
• Resolved: 29/Jul/14 19:07

Description

I've seen behavior a few times now where the DFS is being slow for one reason or another, and the JT essentially locks up waiting on it while one thread tries for a long time to write history files out. The stack trace blocking everything is:

Thread 210 (IPC Server handler 10 on 7277):
State: WAITING
Blocked count: 171424
Waited count: 1209604
Waiting on java.util.LinkedList@407dd154
Stack:
java.lang.Object.wait(Native Method)
java.lang.Object.wait(Object.java:485)
org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.flushInternal(DFSClient.java:3122)
org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.closeInternal(DFSClient.java:3202)
org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.close(DFSClient.java:3151)
org.apache.hadoop.fs.FSDataOutputStream$PositionCache.close(FSDataOutputStream.java:67)
org.apache.hadoop.fs.FSDataOutputStream.close(FSDataOutputStream.java:106)
sun.nio.cs.StreamEncoder.implClose(StreamEncoder.java:301)
sun.nio.cs.StreamEncoder.close(StreamEncoder.java:130)
java.io.OutputStreamWriter.close(OutputStreamWriter.java:216)
java.io.BufferedWriter.close(BufferedWriter.java:248)
java.io.PrintWriter.close(PrintWriter.java:295)
org.apache.hadoop.mapred.JobHistory$JobInfo.logFinished(JobHistory.java:1349)
org.apache.hadoop.mapred.JobInProgress.jobComplete(JobInProgress.java:2167)
org.apache.hadoop.mapred.JobInProgress.completedTask(JobInProgress.java:2111)
org.apache.hadoop.mapred.JobInProgress.updateTaskStatus(JobInProgress.java:873)
org.apache.hadoop.mapred.JobTracker.updateTaskStatuses(JobTracker.java:3598)
org.apache.hadoop.mapred.JobTracker.processHeartbeat(JobTracker.java:2792)
org.apache.hadoop.mapred.JobTracker.heartbeat(JobTracker.java:2581)
sun.reflect.GeneratedMethodAccessor14.invoke(Unknown Source)

We should try not to do external IO while holding the JT lock, and instead write the data to an in-memory buffer, drop the lock, and then write.

Issue Links

• relates to: MAPREDUCE-5711 When JobTracker writing JobHistory to HDFS, it may hung for long time due to DataNode error