Robert Hou created DRILL-6897:
---------------------------------
Summary: TPCH 13 has regressed
Key: DRILL-6897
URL: https://issues.apache.org/jira/browse/DRILL-6897
Project: Apache Drill
Issue Type: Bug
Components: Execution - Relational Operators
Affects Versions: 1.15.0
Reporter: Robert Hou
Assignee: Karthikeyan Manivannan
Attachments: 240099ed-ef2a-a23a-4559-f1b2e0809e72.sys.drill,
2400be84-c024-cb92-8743-3211589e0247.sys.drill
I ran TPCH query 13 with both scale factor 100 and 1000, and ran them 3x to get
a warm start, and ran them twice to verify the regression. It is regressing
between 26 and 33%.
Here is the query:
{noformat}
select
c_count,
count(*) as custdist
from
(
select
c.c_custkey,
count(o.o_orderkey)
from
customer c
left outer join orders o
on c.c_custkey = o.o_custkey
and o.o_comment not like '%special%requests%'
group by
c.c_custkey
) as orders (c_custkey, c_count)
group by
c_count
order by
custdist desc,
c_count desc;
{noformat}
I have attached two profiles. 240099ed-ef2a-a23a-4559-f1b2e0809e72 is for Drill
1.15. 2400be84-c024-cb92-8743-3211589e0247 is for Drill 1.14. The commit for
Drill 1.15 is 596227bbbecfb19bdb55dd8ea58159890f83bc9c. The commit for Drill
1.14 is 0508a128853ce796ca7e99e13008e49442f83147.
The two plans nearly the same. One difference is that Drill 1.15 is using four
times more memory in operator 07-01 Unordered Mux Exchange. I think the problem
may be in operator 09-01 Project. Drill 1.15 is projecting the comment field
while Drill 1.14 does not project the comment field.
Another issue is that the Drill 1.15 takes more processing time to filter the
order table. Filter operator 09-03 takes an average of 19.3s. For Drill 1.14,
filter operator 09-04 takes an average of 15.6s. They process the same number
of rows, and have the same number of minor fragments.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
