Hanumath Rao Maduri created DRILL-6997:
------------------------------------------
Summary: Semijoin is changing the join ordering for some tpcds
queries.
Key: DRILL-6997
URL: https://issues.apache.org/jira/browse/DRILL-6997
Project: Apache Drill
Issue Type: Bug
Components: Query Planning & Optimization
Affects Versions: 1.15.0
Reporter: Hanumath Rao Maduri
Assignee: Hanumath Rao Maduri
Fix For: 1.16.0
TPCDS query 95 runs 50% slower with semi-join enabled compared to semi-join
disabled at scale factor 100. It runs 100% slower at scale factor 1000. This
issue was introduced with commit 71809ca6216d95540b2a41ce1ab2ebb742888671.
DRILL-6798: Planner changes to support semi-join.
{code:java}
with ws_wh as
(select ws1.ws_order_number,ws1.ws_warehouse_sk wh1,ws2.ws_warehouse_sk wh2
from web_sales ws1,web_sales ws2
where ws1.ws_order_number = ws2.ws_order_number
and ws1.ws_warehouse_sk <> ws2.ws_warehouse_sk)
[_LIMITA] select [_LIMITB]
count(distinct ws_order_number) as "order count"
,sum(ws_ext_ship_cost) as "total shipping cost"
,sum(ws_net_profit) as "total net profit"
from
web_sales ws1
,date_dim
,customer_address
,web_site
where
d_date between '[YEAR]-[MONTH]-01' and
(cast('[YEAR]-[MONTH]-01' as date) + 60 days)
and ws1.ws_ship_date_sk = d_date_sk
and ws1.ws_ship_addr_sk = ca_address_sk
and ca_state = '[STATE]'
and ws1.ws_web_site_sk = web_site_sk
and web_company_name = 'pri'
and ws1.ws_order_number in (select ws_order_number
from ws_wh)
and ws1.ws_order_number in (select wr_order_number
from web_returns,ws_wh
where wr_order_number = ws_wh.ws_order_number)
order by count(distinct ws_order_number)
[_LIMITC];
{code}
I have attached two profiles. 240abc6d-b816-5320-93b1-2a07d850e734 has
semi-join enabled. 240aa5f8-24c4-e678-8d42-0fc06e5d2465 has semi-join disabled.
Both are executed with commit id 6267185823c4c50ab31c029ee5b4d9df2fc94d03 and
scale factor 100.
The plan with semi-join enabled has moved the first hash join:
and ws1.ws_order_number in (select ws_order_number
from ws_wh)
It used to be on the build side of the first HJ on the left hand side (04-05).
It is now on the build side of the fourth HJ on the left hand side (01-13).
The plan with semi-join enabled has a hash_partition_sender (operator 05-00)
that takes 10 seconds to execute. But all the fragments take about the same
amount of time.
The plan with semi-join enabled has two HJ that process 1B rows while the plan
with semi-joins disabled has one HJ that processes 1B rows.
The plan with semi-join enabled has several senders and receivers that wait
more than 10 seconds, (00-07, 01-07, 03-00, 04-00, 07-00, 08-00, 14-00, 17-00).
When disabled, there is no operator waiting more than 10 seconds.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
