[HIVE-6309] Hive incorrectly removes TaskAttempt output files if MRAppMaster fails once

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 0.13.0

Environment

hadoop 2.2

• Assignee: Chun Chen
• Reporter: Chun Chen
• Created: 26/Jan/14 17:10
• Updated: 30/Jan/14 18:59
• Resolved: 30/Jan/14 18:59

Description

We recently upgrade to hadoop2.2 and sometimes find some tables lost several data files after a mid night ETL process. We find that these MapReduce jobs which generate the partial tables have something in common that the MRAppMaster of which all had failed once and the tables all left only a single data file 000000_1000.

The following log in hive.log give us some clues of what's going on with the incorrectly deleted data files.

$ grep 'hive_2014-01-24_12-33-18_507_6790415670781610350' hive.log
2014-01-24 12:52:43,140 WARN  exec.Utilities (Utilities.java:removeTempOrDuplicateFiles(1535)) - Duplicate taskid file removed: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000001_1000 with length 824627293. Existing file: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000000_1000 with length 824860643
2014-01-24 12:52:43,142 WARN  exec.Utilities (Utilities.java:removeTempOrDuplicateFiles(1535)) - Duplicate taskid file removed: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000002_1000 with length 824681826. Existing file: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000000_1000 with length 824860643
2014-01-24 12:52:43,149 WARN  exec.Utilities (Utilities.java:removeTempOrDuplicateFiles(1535)) - Duplicate taskid file removed: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000003_1000 with length 824830450. Existing file: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000000_1000 with length 824860643
2014-01-24 12:52:43,151 WARN  exec.Utilities (Utilities.java:removeTempOrDuplicateFiles(1535)) - Duplicate taskid file removed: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000004_1000 with length 824753882. Existing file: hdfs://hadoop00.lf.sankuai.com:9000/tmp/hive-scratch/hive-sankuai/hive_2014-01-24_12-33-18_507_6790415670781610350/_tmp.-ext-10000.intermediate/000000_1000 with length 824860643

We find that it's because nextAttemptNumber in hadoop2.2 is bigger than 1000 and hive doesn't correctly extract task id from filename. See the following code in org.apache.hadoop.mapreduce.v2.app.job.impl.TaskImpl.java
and ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java

// org.apache.hadoop.mapreduce.v2.app.job.impl.TaskImpl.java
    // All the new TaskAttemptIDs are generated based on MR
    // ApplicationAttemptID so that attempts from previous lives don't
    // over-step the current one. This assumes that a task won't have more
    // than 1000 attempts in its single generation, which is very reasonable.
    nextAttemptNumber = (appAttemptId - 1) * 1000;

// ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java
   /**
    * The first group will contain the task id. The second group is the optional extension. The file
    * name looks like: "0_0" or "0_0.gz". There may be a leading prefix (tmp_). Since getTaskId() can
    * return an integer only - this should match a pure integer as well. {1,3} is used to limit
    * matching for attempts #'s 0-999.
    */
   private static final Pattern FILE_NAME_TO_TASK_ID_REGEX =
       Pattern.compile("^.*?([0-9]+)(_[0-9]{1,3})?(\\..*)?$");

And with the bellow reasons, extract this value for attempt numbers >= 1000 :

>>> re.match("^.*?([0-9]+)(_[0​-9])?(\\..*)?$", 'part-r-000000_2').group(1)
'000000'
>>> re.match("^.*?([0-9]+)(_[0​-9])?(\\..*)?$", 'part-r-000000_1001').group(1)
'1001'

Issue Links

• is related to: HIVE-2309 Incorrect regular expression for extracting task id from filename