Jesus Camacho Rodriguez created HIVE-8975:
---------------------------------------------
Summary: Possible performance regression on bucket_map_join_tez2.q
Key: HIVE-8975
URL: https://issues.apache.org/jira/browse/HIVE-8975
Project: Hive
Issue Type: Bug
Components: Logical Optimizer, Statistics
Affects Versions: 0.15.0
Reporter: Jesus Camacho Rodriguez
After introducing the identity project removal optimization in HIVE-8435, plan
in bucket_map_join_tez2.q that runs on Tez changed to be sub-optimal. In
particular, earlier it was doing a map-join and after HIVE-8435 it changed to a
reduce-join.
The query is the following one:
{noformat}
select a.key, b.key from (select distinct key from tab) a join tab b on b.key =
a.key
{noformat}
The plan before removing the projections is:
{noformat}
TS[0]-FIL[16]-SEL[1]-GBY[2]-RS[3]-GBY[4]-SEL[5]-RS[8]-JOIN[11]-SEL[12]-FS[13]
TS[6]-FIL[17]-RS[10]-JOIN[11]
{noformat}
And after removing identity projections:
{noformat}
TS[0]-FIL[16]-GBY[2]-RS[3]-GBY[4]-RS[8]-JOIN[11]-SEL[12]-FS[13]
TS[6]-FIL[17]-RS[10]-JOIN[11]
{noformat}
After digging a bit, I realized it is not converting the reduce-join into a
map-join because stats for GBY\[4\] change if SEL\[5\] is removed; thus the
optimization does not kick in.
The reason for the stats change in the GroupBy operator is in [this
line|https://github.com/apache/hive/blob/6f4365e8a21e7b480bf595d079a71303a50bf1b2/ql/src/java/org/apache/hadoop/hive/ql/optimizer/stats/annotation/StatsRulesProcFactory.java#L633],
where it is checked whether the GBY is immediately followed by a RS operator
or not, and calculate stats differently depending on it.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
