Sun Rui created HIVE-5256:
-----------------------------

             Summary: A map join operator may have in-consistent output row 
schema with the common join operator which it will replace
                 Key: HIVE-5256
                 URL: https://issues.apache.org/jira/browse/HIVE-5256
             Project: Hive
          Issue Type: Bug
          Components: Query Processor
    Affects Versions: 0.11.0
            Reporter: Sun Rui


When generating a common join operator, Semantic Analyzer gets the input 
RowResolver of each parent operators. It then uses the following piece of code 
to interate the tables from the RowResolver (refer to 
genJoinOperatorChildren()):
        RowResolver inputRS = opParseCtx.get(input).getRowResolver();
        Iterator<String> keysIter = inputRS.getTableNames().iterator();
                ...
        while (keysIter.hasNext()) {
          String key = keysIter.next();

Note that the interation order is not deterministic because of the current 
RowResolver implementation:
  private  HashMap<String, LinkedHashMap<String, ColumnInfo>> rslvMap;

  ...

  public Set<String> getTableNames() {
    return rslvMap.keySet();
  }

Generally, the interation order has no problem. However, it may be problematic 
when a common join operator is being converted to a map join operator.

MapJoinProcessor.convertMapJoin():

      RowResolver inputRS = 
opParseCtxMap.get(newParentOps.get(pos)).getRowResolver();
      List<ExprNodeDesc> values = new ArrayList<ExprNodeDesc>();

      Iterator<String> keysIter = inputRS.getTableNames().iterator();
      while (keysIter.hasNext()) {

The problem is that the table iteration order for a input RowResolver may be 
different from that in the generation of the common join operator, which result 
in an in-consistent output row schema. Thus wrong row schema may be input to 
child operators and will cause problems.

I found this issue when running a TPC-DS query. And this issue happens to be 
exposed due to HIVE-4078.

The proposed fix is to change RowResolver to define rslvMap as LinkedHashMap 
instead of HashMap. Thus the table iteration order of a RowResolver is fixed.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to