Incorrect query results in hive when hive.convert.join.bucket.mapjoin.tez is set to true

When hive.convert.join.bucket.mapjoin.tez=true and bucketed column is in select clause but not in where clause, hive is performing a bucket map join and returning incorrect results. When the bucketed column is removed from select clause or  hive.convert.join.bucket.mapjoin.tez=false, returned query results are correct.

 

create table my_fact(AMT decimal(20,3),bucket_col string ,join_col string )

PARTITIONED BY (FISCAL_YEAR string ,ACCOUNTING_PERIOD string )
CLUSTERED BY (bucket_col) INTO 10
BUCKETS
stored as ORC
;
create table my_dim(join_col string,filter_col string) stored as orc;

After populating and analyzing above tables, explain  plan looks as below when  hive.convert.join.bucket.mapjoin.tez=TRUE:

 

explain  select T4.join_col as account1,my_fact.accounting_period
FROM my_fact JOIN my_dim T4 ON my_fact.join_col = T4.join_col
WHERE my_fact.fiscal_year = '2015'
AND T4.filter_col IN ( 'VAL1', 'VAL2' )
and my_fact.accounting_period in (10);

Vertex dependency in root stage
Map 1 <- Map 2 (CUSTOM_EDGE)

Stage-0
Fetch Operator
limit:-1
Stage-1
Map 1 vectorized, llap
File Output Operator [FS_24]
Select Operator [SEL_23] (rows=15282589 width=291)
Output:["_col0","_col1","_col2"]
Map Join Operator [MAPJOIN_22] (rows=15282589 width=291)
BucketMapJoin:true,Conds:SEL_21._col1=RS_19._col0(Inner),Output:["_col0","_col3","_col4"]
<-Map 2 [CUSTOM_EDGE] vectorized, llap
MULTICAST [RS_19]
PartitionCols:_col0
Select Operator [SEL_18] (rows=818 width=186)
Output:["_col0"]
Filter Operator [FIL_17] (rows=818 width=186)
predicate(filter_col) IN ('VAL1', 'VAL2') and join_col is not null)
TableScan [TS_3] (rows=1635 width=186)
default@my_dim,t4,Tbl:COMPLETE,Col:NONE,Output:["join_col","filter_col"]
<-Select Operator [SEL_21] (rows=13893263 width=291)
Output:["_col0","_col1","_col3"]
Filter Operator [FIL_20] (rows=13893263 width=291)
predicate:join_col is not null
TableScan [TS_0] (rows=13893263 width=291)
default@my_fact,my_fact,Tbl:COMPLETE,Col:NONE,Output:["bucket_col","join_col"]

extended_explain.txt has more detailed plan.

When  hive.convert.join.bucket.mapjoin.tez=false,  plan no longer has bucketjoin and query results are correct.

Vertex dependency in root stage
Map 1 <- Map 2 (BROADCAST_EDGE)

Stage-0
Fetch Operator
limit:-1
Stage-1
Map 1 vectorized, llap
File Output Operator [FS_24]
Select Operator [SEL_23] (rows=15282589 width=291)
Output:["_col0","_col1","_col2"]
Map Join Operator [MAPJOIN_22] (rows=15282589 width=291)
Conds:SEL_21._col1=RS_19._col0(Inner),Output:["_col0","_col3","_col4"]
<-Map 2 [BROADCAST_EDGE] vectorized, llap
BROADCAST [RS_19]
PartitionCols:_col0
Select Operator [SEL_18] (rows=818 width=186)
Output:["_col0"]
Filter Operator [FIL_17] (rows=818 width=186)
predicate(filter_col) IN ('VAL1', 'VAL2') and join_col is not null)
TableScan [TS_3] (rows=1635 width=186)
default@my_dim,t4,Tbl:COMPLETE,Col:NONE,Output:["join_col","filter_col"]
<-Select Operator [SEL_21] (rows=13893263 width=291)
Output:["_col0","_col1","_col3"]
Filter Operator [FIL_20] (rows=13893263 width=291)
predicate:join_col is not null
TableScan [TS_0] (rows=13893263 width=291)
default@my_fact,my_fact,Tbl:COMPLETE,Col:NONE,Output:

 

Per suggestion in bug# TEZ-3971, creating this case under Hive project.