[HADOOP-14219] RumenToSLS: parsing problem with crashed attempts

• Type: Bug
• Status: Patch Available
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 2.6.0
• Component/s: tools
• Assignee: Unassigned
• Reporter: Julien Vaudour
• Created: 23/Mar/17 12:56
• Updated: 24/Mar/17 18:34

Description

In case of crashed task attempts, we may have in rumen logs task attempts with null hostName and finishTime defined to -1

for example

       {
          "resourceUsageMetrics": {
            "heapUsage": 0,
            "physicalMemoryUsage": 0,
            "virtualMemoryUsage": 0,
            "cumulativeCpuUsage": 0
          },
          "vmemKbytes": [],
          "physMemKbytes": [],
          "cpuUsages": [],
          "clockSplits": [],
          "location": null,
          "sortFinished": -1,
          "shuffleFinished": -1,
          "spilledRecords": -1,
          "reduceOutputRecords": -1,
          "reduceShuffleBytes": -1,
          "fileBytesRead": -1,
          "hdfsBytesWritten": -1,
          "hdfsBytesRead": -1,
          "hostName": null,
          "finishTime": -1,
          "startTime": 1489619193378,
          "result": null,
          "attemptID": "attempt_1488896259152_410442_r_000015_1",
          "fileBytesWritten": -1,
          "mapInputRecords": -1,
          "mapInputBytes": -1,
          "mapOutputBytes": -1,
          "mapOutputRecords": -1,
          "combineInputRecords": -1,
          "reduceInputGroups": -1,
          "reduceInputRecords": -1
        }

Jackson parser will automatically consider -1 as a java.lang.Integer. However RumenToSLSConverter make the assumption than jackson has deserialize all timstamp as instance of java.lang.Long, resulting in a ClassCastException.

RumenToSLSConverter also make the assumption that hostName is not null, so we can also have a NullPointerException.