Jihoon Son created TAJO-1342:
--------------------------------
Summary: The complex join query runs forever
Key: TAJO-1342
URL: https://issues.apache.org/jira/browse/TAJO-1342
Project: Tajo
Issue Type: Bug
Reporter: Jihoon Son
TAJO-1277 fixes a bug of join ordering which ignores the different
associativity according to the join types.
After TAJO-1277, I found an additional hidden bug. Please see the following
example.
{noformat}
default> select * from customer right outer join supplier on c_nationkey =
s_nationkey inner join nation on n_nationkey = c_nationkey full outer join
region on n_regionkey = r_regionkey;
Progress: 0%, response time: 1.632 sec
Progress: 0%, response time: 1.633 sec
Progress: 0%, response time: 2.038 sec
Progress: 0%, response time: 2.843 sec
Progress: 98%, response time: 3.848 sec
Progress: 98%, response time: 4.854 sec
...
(Infinitely repeated)
{noformat}
The join order of this query is:
{noformat}
(((default.customer ⟖ default.supplier) ⋈θ default.nation) ⟗ default.region)
{noformat}
Here is the query master log
{noformat}
...
2015-02-11 12:30:33,987 INFO
org.apache.tajo.engine.planner.PhysicalPlannerImpl: Right Outer Join (9)
chooses [Hash Join].
2015-02-11 12:30:33,990 INFO
org.apache.tajo.engine.planner.PhysicalPlannerImpl:
[ta_1423625266462_0001_000003_000000_00] the volume of Left relations
(eb_1423625266462_0001_000001, eb_1423625266462_0001_000002) is 24.8 MiB and is
fit to main maemory.
2015-02-11 12:30:33,990 INFO
org.apache.tajo.engine.planner.PhysicalPlannerImpl: Join (10) chooses
[In-memory Hash Join]
2015-02-11 12:30:33,990 INFO
org.apache.tajo.engine.planner.PhysicalPlannerImpl:
[ta_1423625266462_0001_000003_000000_00] Right relations default.nation (2.2
KiB) is smaller than Left relations eb_1423625266462_0001_000001,
eb_1423625266462_0001_000002 (24.8 MiB).
2015-02-11 12:30:33,991 INFO
org.apache.tajo.engine.planner.PhysicalPlannerImpl: Full Outer Join (11)
chooses [Hash Join]
{noformat}
As you can see, the query master seems to fall into an infinite loop during
executing the hash outer join.
This bug is not found in the current master branch. The difference is found in
the join order as follows:
{noformat}
(((default.nation ⟗ default.region) ⋈θ default.customer) ⟖ default.supplier)
{noformat}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
