[
https://issues.apache.org/jira/browse/SPARK-21479?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16100691#comment-16100691
]
Anton Okolnychyi commented on SPARK-21479:
------------------------------------------
I used the following code to investigate:
{code}
val inputSchema1 = StructType(
StructField("col1", StringType) ::
StructField("col2", IntegerType) ::
Nil)
val inputSchema2 = StructType(
StructField("col1", StringType) ::
StructField("col3", StringType) ::
Nil)
val rdd1 = sc.parallelize(1 to 3).map(v => Row(s"value $v", v))
val df1 = spark.createDataFrame(rdd1, inputSchema1)
val rdd2 = sc.parallelize(1 to 3).map(v => Row(s"value $v", "some value"))
val df2 = spark.createDataFrame(rdd2, inputSchema2)
// 1st use case
df1.join(df2, Seq("col1"), "right_outer").where("col2 = 2").explain(true)
// 2nd use case
df1.join(df2, Seq("col1"), "right_outer").where("col1 = 'value
2'").explain(true)
{code}
It is important to notice that the actual join type in the first case is
`inner` and not `rigth_outer`. This happens due to the `EliminateOuterJoin`
rule, which sees that `col2 = 2` filters out non-matching rows on the left side
of the join. Once the join type is changed, the `PushPredicateThroughJoin` rule
pushes `col2 = 2` to the left relation. The analyzed and optimized logical
plans are:
{noformat}
== Analyzed Logical Plan ==
col1: string, col2: int, col3: string
Filter (col2#3 = 2)
+- Project [col1#9, col2#3, col3#10]
+- Join RightOuter, (col1#2 = col1#9)
:- LogicalRDD [col1#2, col2#3]
+- LogicalRDD [col1#9, col3#10]
== Optimized Logical Plan ==
Project [col1#9, col2#3, col3#10]
+- Join Inner, (col1#2 = col1#9)
:- Filter ((isnotnull(col2#3) && (col2#3 = 2)) && isnotnull(col1#2))
: +- LogicalRDD [col1#2, col2#3]
+- Filter isnotnull(col1#9)
+- LogicalRDD [col1#9, col3#10]
{noformat}
The second case is different. The join type stays the same (i.e.,
`right_outer`) and the analyzed logical plan looks like:
{noformat}
== Analyzed Logical Plan ==
col1: string, col2: int, col3: string
Filter (col1#9 = value 2)
+- Project [col1#9, col2#3, col3#10]
+- Join RightOuter, (col1#2 = col1#9)
:- LogicalRDD [col1#2, col2#3]
+- LogicalRDD [col1#9, col3#10]
{noformat}
`col1#9` from the Filter belongs to the right relation. After
`PushPredicateThroughJoin` we have:
{noformat}
Join RightOuter, (col1#2 = col1#9)
:- LogicalRDD [col1#2, col2#3]
+- Filter (isnotnull(col1#9) && (col1#9 = value 2))
+- LogicalRDD [col1#9, col3#10]
{noformat}
In theory, `InferFiltersFromConstraints` is capable of inferring `(col1#2 =
value 2)` from `(col1#9 = value 2, col1#2 = col1#9)`. However, not in this case
since the join type is `right_outer` and `InferFiltersFromConstraints` will
process only constraints from the right relation (i.e., `(isnotnull(col1#9) &&
(col1#9 = value 2))`), which is not enough to infer `(col1#2 = value 2)`.
It seems like this is done on purpose and it is expected behavior even though
additional `(col1#2 = value 2)` on the left relation would be logically correct
here (as far as I understand).
> Outer join filter pushdown in null supplying table when condition is on one
> of the joined columns
> -------------------------------------------------------------------------------------------------
>
> Key: SPARK-21479
> URL: https://issues.apache.org/jira/browse/SPARK-21479
> Project: Spark
> Issue Type: Bug
> Components: Optimizer, SQL
> Affects Versions: 2.1.0, 2.1.1, 2.2.0
> Reporter: Abhijit Bhole
>
> Here are two different query plans -
> {code:java}
> df1 = spark.createDataFrame([{ "a": 1, "b" : 2}, { "a": 3, "b" : 4}])
> df2 = spark.createDataFrame([{ "a": 1, "c" : 5}, { "a": 3, "c" : 6}, { "a":
> 5, "c" : 8}])
> df1.join(df2, ['a'], 'right_outer').where("b = 2").explain()
> == Physical Plan ==
> *Project [a#16299L, b#16295L, c#16300L]
> +- *SortMergeJoin [a#16294L], [a#16299L], Inner
> :- *Sort [a#16294L ASC NULLS FIRST], false, 0
> : +- Exchange hashpartitioning(a#16294L, 4)
> : +- *Filter ((isnotnull(b#16295L) && (b#16295L = 2)) &&
> isnotnull(a#16294L))
> : +- Scan ExistingRDD[a#16294L,b#16295L]
> +- *Sort [a#16299L ASC NULLS FIRST], false, 0
> +- Exchange hashpartitioning(a#16299L, 4)
> +- *Filter isnotnull(a#16299L)
> +- Scan ExistingRDD[a#16299L,c#16300L]
> df1 = spark.createDataFrame([{ "a": 1, "b" : 2}, { "a": 3, "b" : 4}])
> df2 = spark.createDataFrame([{ "a": 1, "c" : 5}, { "a": 3, "c" : 6}, { "a":
> 5, "c" : 8}])
> df1.join(df2, ['a'], 'right_outer').where("a = 1").explain()
> == Physical Plan ==
> *Project [a#16314L, b#16310L, c#16315L]
> +- SortMergeJoin [a#16309L], [a#16314L], RightOuter
> :- *Sort [a#16309L ASC NULLS FIRST], false, 0
> : +- Exchange hashpartitioning(a#16309L, 4)
> : +- Scan ExistingRDD[a#16309L,b#16310L]
> +- *Sort [a#16314L ASC NULLS FIRST], false, 0
> +- Exchange hashpartitioning(a#16314L, 4)
> +- *Filter (isnotnull(a#16314L) && (a#16314L = 1))
> +- Scan ExistingRDD[a#16314L,c#16315L]
> {code}
> If condition on b can be pushed down on df1 then why not condition on a?
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
