liyunzhang_intel created HIVE-17407:
---------------------------------------
Summary: TPC-DS/query65 hangs on HoS in certain settings
Key: HIVE-17407
URL: https://issues.apache.org/jira/browse/HIVE-17407
Project: Hive
Issue Type: Bug
Reporter: liyunzhang_intel
[TPC-DS/query65.sql|https://github.com/kellyzly/hive-testbench/blob/hive14/sample-queries-tpcds/query65.sql]
hangs when using following settings on 3TB scale.
{code}
set hive.auto.convert.join.noconditionaltask.size=3000000;
{code}
the explain is attached in explain65. The screenshot shows that it hanged in
the Stage5.
Let's explain why hang.
{code}
Reducer 10 <- Map 9 (GROUP, 1009)
Reducer 2 <- Map 1 (PARTITION-LEVEL SORT, 1), Map 5 (PARTITION-LEVEL
SORT, 1), Reducer 7 (PARTITION-LEVEL SORT, 1)
Reducer 3 <- Reducer 10 (PARTITION-LEVEL SORT, 1009), Reducer 2
(PARTITION-LEVEL SORT, 1009)
Reducer 4 <- Reducer 3 (SORT, 1)
Reducer 7 <- Map 6 (GROUP PARTITION-LEVEL SORT, 1009)
{code}
The numPartitions of SparkEdgeProperty which connects Reducer 2 and Reducer 3
is 1. This is because
org.apache.hadoop.hive.ql.parse.spark.GenSparkUtils#createReduceWork
{code}
public ReduceWork createReduceWork(GenSparkProcContext context, Operator<?>
root,
SparkWork sparkWork) throws SemanticException {
for (Operator<? extends OperatorDesc> parentOfRoot :
root.getParentOperators()) {
Preconditions.checkArgument(parentOfRoot instanceof ReduceSinkOperator,
"AssertionError: expected parentOfRoot to be an "
+ "instance of ReduceSinkOperator, but was "
+ parentOfRoot.getClass().getName());
ReduceSinkOperator reduceSink = (ReduceSinkOperator) parentOfRoot;
maxExecutors = Math.max(maxExecutors,
reduceSink.getConf().getNumReducers());
}
reduceWork.setNumReduceTasks(maxExecutors);
{code}
here the numReducers of all parentOfRoot is 1( in the explain, the parallelism
of Map 1, Map 5, Reducer 7 is 1), so the numPartitions of SparkEdgeProperty
which connects Reducer 2 and Reducer 3 is 1.
More explain why the parallelism of Map 1, Map 5,Reducer 7 are 1. The physical
plan of the query is
{code}
TS[0]-FIL[50]-RS[2]-JOIN[5]-FIL[49]-SEL[7]-GBY[8]-RS[9]-GBY[10]-SEL[11]-GBY[15]-SEL[16]-RS[33]-JOIN[34]-RS[36]-JOIN[39]-FIL[48]-SEL[41]-RS[42]-SEL[43]-LIM[44]-FS[45]
TS[1]-FIL[51]-RS[4]-JOIN[5]
TS[17]-FIL[53]-RS[19]-JOIN[22]-FIL[52]-SEL[24]-GBY[25]-RS[26]-GBY[27]-RS[38]-JOIN[39]
TS[18]-FIL[54]-RS[21]-JOIN[22]
TS[29]-FIL[55]-RS[31]-JOIN[34]
TS[30]-FIL[56]-RS[32]-JOIN[34]
{code}
The related RS of Map1, Map5, Reducer 7 is RS\[31\], RS\[32\], RS\[33\]. The
parallelism is set by
[SemanticAnalyzer#genJoinReduceSinkChild|https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/parse/SemanticAnalyzer.java#L8267]
It seems that there is no logical error in the code. But it is not reasonable
to use 1 task to execute to deal with so big data(more than 30GB). Is there any
way to pass the query in this situation( the reason why i set
hive.auto.convert.join.noconditionaltask.size as 3000000, if the join is
converted to the map join, it will throw disk error).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
