Hyunsik Choi created TAJO-926:
---------------------------------
Summary: Join condition including column references of a
row-preserving table in left outer join causes incorrect result
Key: TAJO-926
URL: https://issues.apache.org/jira/browse/TAJO-926
Project: Tajo
Issue Type: Bug
Components: physical operator, planner/optimizer
Reporter: Hyunsik Choi
Assignee: Hyunsik Choi
Fix For: 0.9.0
This patch fixes two bugs.
One is wrong projection push down (PPD). See the example, reproducing the bug:
{noformat}
select
r_name,
r_regionkey,
n_name,
n_regionkey
from
region left outer join nation on n_regionkey = r_regionkey and r_name in
('AMERICA', 'ASIA')
order by r_name;
{noformat}
The above query includes one left outer join (LOJ) and one join filter. Since
this join filter {{R_NAME in ('AMERICA', 'ASIA')}} includes column references
corresponding to the row preserved table {{region}}, the join filter is placed
on the LOJ operator. It only results in the sub expression push down of
RowConstantEval and replaces right expression of IN predicate by FieldEval.
But, we assume that the RHS of InEval is always RowConstantEval. This is the
main clause of this bug.
{noformat}
2014-07-09 16:39:37,527 ERROR: org.apache.tajo.worker.Task (run(395)) -
org.apache.tajo.engine.eval.FieldEval cannot be cast to
org.apache.tajo.engine.eval.RowConstantEval
java.lang.ClassCastException: org.apache.tajo.engine.eval.FieldEval cannot be
cast to org.apache.tajo.engine.eval.RowConstantEval
at org.apache.tajo.engine.eval.InEval.eval(InEval.java:62)
at org.apache.tajo.engine.eval.BinaryEval.eval(BinaryEval.java:104)
at
org.apache.tajo.engine.planner.physical.NLLeftOuterJoinExec.next(NLLeftOuterJoinExec.java:109)
at
org.apache.tajo.engine.planner.physical.ExternalSortExec.sortAndStoreAllChunks(ExternalSortExec.java:201)
at
org.apache.tajo.engine.planner.physical.ExternalSortExec.next(ExternalSortExec.java:278)
at
org.apache.tajo.engine.planner.physical.RangeShuffleFileWriteExec.next(RangeShuffleFileWriteExec.java:99)
at org.apache.tajo.worker.Task.run(Task.java:388)
at org.apache.tajo.worker.TaskRunner$1.run(TaskRunner.java:406)
at java.lang.Thread.run(Thread.java:744)
2014-07-09 16:39:37,528 INFO: org.apache.tajo.worker.TaskAttemptContext
(setState(115)) - Query status of ta_1404891573341_0004_000003_000000_02 is
changed to TA_FAILED
2014-07-09 16:39:37,529 INFO: org.apache.tajo.worker.Task (run(452)) - Worker's
task counter - total:3, succeeded: 0, killed: 3, failed: 3
{noformat}
The second bug is that HashLeftOuterJoin results in wrong result when it has
join filter corresponding to row preserved table like the above example query.
In order to fix this bug, we have to skip the right iterator of hash table when
if the joined tuple is filtered.
Expected:
{noformat}
r_name,r_regionkey,n_name,n_regionkey
-------------------------------
AFRICA,0,null,null
AMERICA,1,ARGENTINA,1
AMERICA,1,BRAZIL,1
AMERICA,1,CANADA,1
AMERICA,1,PERU,1
AMERICA,1,UNITED STATES,1
ASIA,2,INDIA,2
ASIA,2,INDONESIA,2
ASIA,2,JAPAN,2
ASIA,2,CHINA,2
ASIA,2,VIETNAM,2
EUROPE,3,null,null
MIDDLE EAST,4,null,null
{noformat}
Actual result:
{noformat}
r_name,r_regionkey,n_name,n_regionkey
-------------------------------
AFRICA,0,null,null
AFRICA,0,null,null
AFRICA,0,null,null
AFRICA,0,null,null
AFRICA,0,null,null
AMERICA,1,ARGENTINA,1
AMERICA,1,BRAZIL,1
AMERICA,1,CANADA,1
AMERICA,1,PERU,1
AMERICA,1,UNITED STATES,1
ASIA,2,INDIA,2
ASIA,2,INDONESIA,2
ASIA,2,JAPAN,2
ASIA,2,CHINA,2
ASIA,2,VIETNAM,2
EUROPE,3,null,null
EUROPE,3,null,null
EUROPE,3,null,null
EUROPE,3,null,null
EUROPE,3,null,null
MIDDLE EAST,4,null,null
MIDDLE EAST,4,null,null
MIDDLE EAST,4,null,null
MIDDLE EAST,4,null,null
MIDDLE EAST,4,null,null
{noformat}
--
This message was sent by Atlassian JIRA
(v6.2#6252)
