[YARN-2024] IOException in AppLogAggregatorImpl does not give stacktrace and leaves aggregated TFile in a bad state.

• Type: Sub-task
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 0.23.10, 2.4.0
• Component/s: log-aggregation
• Target Version/s: 3.5.0
• Assignee: Xuan Gong
• Reporter: Eric Payne
• Created: 06/May/14 16:32
• Updated: 23/Mar/24 09:23

Description

Multiple issues were encountered when AppLogAggregatorImpl encountered an IOException in AppLogAggregatorImpl#uploadLogsForContainer while aggregating yarn-logs for an application that had very large (>150G each) error logs.

• An IOException was encountered during the LogWriter#append call, and a message was printed, but no stacktrace was provided. Message: "ERROR: Couldn't upload logs for container_nnnnnnnnnnnnn_nnnnnnn_nn_nnnnnn. Skipping this container."
• After the IOExceptin, the TFile is in a bad state, so subsequent calls to LogWriter#append fail with the following stacktrace:
  2014-04-16 13:29:09,772 LogAggregationService #17907 ERROR org.apache.hadoop.yarn.YarnUncaughtExceptionHandler: Thread ThreadLogAggregationService #17907,5,main threw an Exception.
  java.lang.IllegalStateException: Incorrect state to start a new key: IN_VALUE
  at org.apache.hadoop.io.file.tfile.TFile$Writer.prepareAppendKey(TFile.java:528)
  at org.apache.hadoop.yarn.logaggregation.AggregatedLogFormat$LogWriter.append(AggregatedLogFormat.java:262)
  at org.apache.hadoop.yarn.server.nodemanager.containermanager.logaggregation.AppLogAggregatorImpl.uploadLogsForContainer(AppLogAggregatorImpl.java:128)
  at org.apache.hadoop.yarn.server.nodemanager.containermanager.logaggregation.AppLogAggregatorImpl.doAppLogAggregation(AppLogAggregatorImpl.java:164)
  ...
• At this point, the yarn-logs cleaner still thinks the thread is aggregating, so the huge yarn-logs never get cleaned up for that application.