[jira] [Created] (IMPALA-8263) Planner failed to flip build/probe side of join

["Paul Rogers (JIRA)" <ji...@apache.org>]

Paul Rogers created IMPALA-8263:
-----------------------------------

             Summary: Planner failed to flip build/probe side of join
                 Key: IMPALA-8263
                 URL: https://issues.apache.org/jira/browse/IMPALA-8263
             Project: IMPALA
          Issue Type: Bug
          Components: Frontend
    Affects Versions: Impala 3.1.0
            Reporter: Paul Rogers
            Assignee: Paul Rogers


TPC-H query 8 was reviewed after applying the changes proposed for DRILL-8014. (See {{tpch-all.test}}.) The revised plan has better cardinality number, but contains this odd structure:

{noformat}
11:HASH JOIN [INNER JOIN]
|  hash predicates: o_custkey = c_custkey
|  row-size=139B cardinality=39.66K
|
|--04:SCAN HDFS [tpch.customer]
|     row-size=10B cardinality=150.00K
|
10:HASH JOIN [INNER JOIN]
|  hash predicates: o_orderkey = l_orderkey
|  row-size=129B cardinality=39.66K
{noformat}

As I understand it, the planner should flip the left and right sides of an inner join if the right side (the 04 scan) has a larger cardinality than the left (join 10) side. That flip did not happen in this case, causing the join to build a hash table about 4 times larger than necessary.

Perhaps there is some other constraint. Investigate to determine if the behavior is correct (and if so why), or the source of incorrect behavior.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)