[MAPREDUCE-1171] Lots of fetch failures

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 0.21.0
• Fix Version/s: 0.21.0
• Component/s: task

Release Note

Added two expert level configuration properties.
1. "mapreduce.reduce.shuffle.notify.readerror" to know whether to send notification to JobTracker after every read error or not. If the configuration is false, read errors are treated similar to connection errors.
2. "mapreduce.reduce.shuffle.maxfetchfailures" to specify the maximum number of the fetch failures after which the failure will be notified to JobTracker.

• Assignee: Amareshwari Sriramadasu
• Reporter: Christian Kunz
• Created: 30/Oct/09 01:53
• Updated: 24/Aug/10 21:19
• Resolved: 11/Dec/09 04:12

Description

Since we upgraded to hadoop-0.20.1 from hadoop0.18.3, we see lot of more map task failures because of 'Too many fetch-failures'.

One of our jobs makes hardly any progress, because of 3000 reduces not able to get map output of 2 trailing maps (with about 80GB output each), which repeatedly are marked as failures because of reduces not being able to get their map output.
One difference to hadoop-0.18.3 seems to be that reduce tasks report a failed mapoutput fetch even after a single try when it was a read error (cr.getError().equals(CopyOutputErrorType.READ_ERROR). I do not think this is a good idea, as trailing map tasks will be attacked by all reduces simultaneously.

Here is a log output of a reduce task:

2009-10-29 21:38:36,148 WARN org.apache.hadoop.mapred.ReduceTask: attempt_200910281903_0028_r_000000_0 copy failed: attempt_200910281903_0028_m_002781_1 from some host
2009-10-29 21:38:36,148 WARN org.apache.hadoop.mapred.ReduceTask: java.net.SocketTimeoutException: Read timed out        at java.net.SocketInputStream.socketRead0(Native Method)
        at java.net.SocketInputStream.read(SocketInputStream.java:129)
        at java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
        at java.io.BufferedInputStream.read1(BufferedInputStream.java:258)
        at java.io.BufferedInputStream.read(BufferedInputStream.java:317)
        at sun.net.www.http.HttpClient.parseHTTPHeader(HttpClient.java:687)
        at sun.net.www.http.HttpClient.parseHTTP(HttpClient.java:632)
        at sun.net.www.protocol.http.HttpURLConnection.getInputStream(HttpURLConnection.java:1064)
        at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.getInputStream(ReduceTask.java:1496)
        at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.getMapOutput(ReduceTask.java:1377)
        at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.copyOutput(ReduceTask.java:1289)
        at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.run(ReduceTask.java:1220)

2009-10-29 21:38:36,149 INFO org.apache.hadoop.mapred.ReduceTask: Task attempt_200910281903_0028_r_000000_0: Failed fetch #1 from attempt_200910281903_0028_m_002781_1
2009-10-29 21:38:36,149 INFO org.apache.hadoop.mapred.ReduceTask: Failed to fetch map-output from attempt_200910281903_0028_m_002781_1 even after MAX_FETCH_RETRIES_PER_MAP retries...  or it is a read error,  reporting to the JobTracker.

Also I saw a few log messages which look suspicious as if successfully fetched map output is discarded because of the map being marked as failed (because of too many fetch failures). This would make the situation even worse.

2009-10-29 22:07:28,729 INFO org.apache.hadoop.mapred.ReduceTask: header: attempt_200910281903_0028_m_001076_0, compressed len: 21882555, decompressed len: 23967845
2009-10-29 22:07:28,729 INFO org.apache.hadoop.mapred.ReduceTask: Shuffling 23967845 bytes (21882555 raw bytes) into RAM from attempt_200910281903_0028_m_001076_0
2009-10-29 22:07:43,602 INFO org.apache.hadoop.mapred.ReduceTask: Read 23967845 bytes from map-output for attempt_200910281903_0028_m_001076_0
2009-10-29 22:07:43,602 INFO org.apache.hadoop.mapred.ReduceTask: Rec #1 from attempt_200910281903_0028_m_001076_0 -> (20, 39772) from some host
...
2009-10-29 22:10:07,220 INFO org.apache.hadoop.mapred.ReduceTask: Ignoring obsolete output of FAILED map-task: 'attempt_200910281903_0028_m_001076_0'

Issue Links

• relates to: HADOOP-3327 Shuffling fetchers waited too long between map output fetch re-tries