Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
-
Reviewed
Description
The predicate pushdown filter is not applying left associativity of joins correctly in determining possible aliases for pushing predicates.
In hive.ql.ppd.OpProcFactory.JoinPPD.getQualifiedAliases, the criteria for pushing aliases is specified as:
/**
* Figures out the aliases for whom it is safe to push predicates based on
* ANSI SQL semantics For inner join, all predicates for all aliases can be
* pushed For full outer join, none of the predicates can be pushed as that
* would limit the number of rows for join For left outer join, all the
* predicates on the left side aliases can be pushed up For right outer
* join, all the predicates on the right side aliases can be pushed up Joins
* chain containing both left and right outer joins are treated as full
* outer join. [...]
*
* @param op
* Join Operator
* @param rr
* Row resolver
* @return set of qualified aliases
*/
Since hive joins are left associative, something like "a RIGHT OUTER JOIN b LEFT OUTER JOIN c INNER JOIN d" should be interpreted as "((a RIGHT OUTER JOIN b) LEFT OUTER JOIN c) INNER JOIN d", so there would be cases where joins with both left and right outer joins can have aliases that can be pushed. Here, aliases b and d are eligible to be pushed up while the current criteria provide that none are eligible.
Using:
create table t1 (id int, key string, value string); create table t2 (id int, key string, value string); create table t3 (id int, key string, value string); create table t4 (id int, key string, value string);
For example, the query
explain select * from t1 full outer join t2 on t1.id=t2.id join t3 on t2.id=t3.id where t3.id=20;
currently gives
STAGE DEPENDENCIES:
Stage-1 is a root stage
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-1
Map Reduce
Alias -> Map Operator Tree:
t1
TableScan
alias: t1
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 0
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
t2
TableScan
alias: t2
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 1
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
t3
TableScan
alias: t3
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 2
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
Reduce Operator Tree:
Join Operator
condition map:
Outer Join 0 to 1
Inner Join 1 to 2
condition expressions:
0 {VALUE._col0} {VALUE._col1} {VALUE._col2}
1 {VALUE._col0} {VALUE._col1} {VALUE._col2}
2 {VALUE._col0} {VALUE._col1} {VALUE._col2}
handleSkewJoin: false
outputColumnNames: _col0, _col1, _col2, _col5, _col6, _col7, _col10, _col11, _col12
Filter Operator
predicate:
expr: (_col10 = 20)
type: boolean
Select Operator
expressions:
expr: _col0
type: int
expr: _col1
type: string
expr: _col2
type: string
expr: _col5
type: int
expr: _col6
type: string
expr: _col7
type: string
expr: _col10
type: int
expr: _col11
type: string
expr: _col12
type: string
outputColumnNames: _col0, _col1, _col2, _col3, _col4, _col5, _col6, _col7, _col8
File Output Operator
compressed: false
GlobalTableId: 0
table:
input format: org.apache.hadoop.mapred.TextInputFormat
output format: org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat
Stage: Stage-0
Fetch Operator
limit: -1
while the correct behavior is to push the filter "t3.id=20" down:
STAGE DEPENDENCIES:
Stage-1 is a root stage
Stage-0 is a root stage
STAGE PLANS:
Stage: Stage-1
Map Reduce
Alias -> Map Operator Tree:
t1
TableScan
alias: t1
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 0
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
t2
TableScan
alias: t2
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 1
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
t3
TableScan
alias: t3
Filter Operator
predicate:
expr: (id = 20)
type: boolean
Reduce Output Operator
key expressions:
expr: id
type: int
sort order: +
Map-reduce partition columns:
expr: id
type: int
tag: 2
value expressions:
expr: id
type: int
expr: key
type: string
expr: value
type: string
Reduce Operator Tree:
Join Operator
condition map:
Outer Join 0 to 1
Inner Join 1 to 2
condition expressions:
0 {VALUE._col0} {VALUE._col1} {VALUE._col2}
1 {VALUE._col0} {VALUE._col1} {VALUE._col2}
2 {VALUE._col0} {VALUE._col1} {VALUE._col2}
handleSkewJoin: false
outputColumnNames: _col0, _col1, _col2, _col5, _col6, _col7, _col10, _col11, _col12
Select Operator
expressions:
expr: _col0
type: int
expr: _col1
type: string
expr: _col2
type: string
expr: _col5
type: int
expr: _col6
type: string
expr: _col7
type: string
expr: _col10
type: int
expr: _col11
type: string
expr: _col12
type: string
outputColumnNames: _col0, _col1, _col2, _col3, _col4, _col5, _col6, _col7, _col8
File Output Operator
compressed: false
GlobalTableId: 0
table:
input format: org.apache.hadoop.mapred.TextInputFormat
output format: org.apache.hadoop.hive.ql.io.HiveIgnoreKeyTextOutputFormat
Stage: Stage-0
Fetch Operator
limit: -1
The current behavior is actually stranger than this: for a left outer join (similarly for a right outer join), hive finds the leftmost alias referred to in the predicates of left outer joins and rejects any alias to the right of it for pushdown. So in this query the filter "t2.id=20" pushed down:
explain select * from t1 join t2 on (t1.id=t2.id) left outer join t3 on (t2.id=t3.id) where t2.id=20;
while it isn't here:
explain select * from t1 join t2 on (t1.id=t2.id) left outer join t3 on (t1.id=t3.id) where t2.id=20;