[HIVE-4885] Alternative object serialization for execution plan in hive testing

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.10.0, 0.11.0
• Fix Version/s: 0.12.0
• Component/s: CLI
• Assignee: Xuefu Zhang
• Reporter: Xuefu Zhang
• Created: 18/Jul/13 21:04
• Updated: 15/Oct/13 23:30
• Resolved: 13/Aug/13 12:57

Description

Currently there are a lot of test cases involving in comparing execution plan, such as those in TestParse suite. XmlEncoder is used to serialize the generated plan by hive, and store it in the file for file diff comparison. However, XmlEncoder is tied with Java compiler, whose implementation may change from version to version. Thus, upgrade the compiler can generate a lot of fake test failures. The following is an example of diff generated when running hive with JDK7:

Begin query: case_sensitivity.q
diff -a /data/4/hive-local/a2307.halxg.cloudera.com-hiveptest-2/cdh-source/build/ql/test/logs/positive/case_sensitivity.q.out /data/4/hive-local/a2307.halxg.cloudera.com-hiveptest-2/cdh-source/ql/src/test/results/compiler/parse/case_sensitivity.q.out
diff -a -b /data/4/hive-local/a2307.halxg.cloudera.com-hiveptest-2/cdh-source/build/ql/test/logs/positive/case_sensitivity.q.xml /data/4/hive-local/a2307.halxg.cloudera.com-hiveptest-2/cdh-source/ql/src/test/results/compiler/plan/case_sensitivity.q.xml
3c3
<  <object class="org.apache.hadoop.hive.ql.exec.MapRedTask" id="MapRedTask0">
---
>  <object id="MapRedTask0" class="org.apache.hadoop.hive.ql.exec.MapRedTask"> 
12c12
<        <object class="java.util.ArrayList" id="ArrayList0">
---
>        <object id="ArrayList0" class="java.util.ArrayList"> 
14c14
<          <object class="org.apache.hadoop.hive.ql.exec.MoveTask" id="MoveTask0">
---
>          <object id="MoveTask0" class="org.apache.hadoop.hive.ql.exec.MoveTask"> 
18c18
<              <object class="org.apache.hadoop.hive.ql.exec.MoveTask" id="MoveTask1">
---
>              <object id="MoveTask1" class="org.apache.hadoop.hive.ql.exec.MoveTask"> 
22c22
<                  <object class="org.apache.hadoop.hive.ql.exec.StatsTask" id="StatsTask0">
---
>                  <object id="StatsTask0" class="org.apache.hadoop.hive.ql.exec.StatsTask"> 
60c60
<                  <object class="org.apache.hadoop.hive.ql.exec.MapRedTask" id="MapRedTask1">
---
>                  <object id="MapRedTask1" class="org.apache.hadoop.hive.ql.exec.MapRedTask"> 

As it can be seen, the only difference is the order of the attributes in the serialized XML doc, yet it brings 50+ test failures in Hive.

We need to have a better plan comparison, or object serialization to improve the situation.

Issue Links

• is depended upon by: HIVE-4583 Make Hive compile and run with JDK7
• is related to: HIVE-5064 TestParse fails on JDK7
• is related to: HIVE-1511 Hive plan serialization is slow