[
https://issues.apache.org/jira/browse/SPARK-24556?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
yucai updated SPARK-24556:
--------------------------
Description:
Currently, ReusedExchange would rewrite output partitioning if child's
partitioning is HashPartitioning, but it does not do the same when child's
partitioning is RangePartitioning, sometimes, it could introduce extra shuffle,
see:
{code:java}
val df = Seq(1 -> "a", 3 -> "c", 2 -> "b").toDF("i", "j")
val df1 = df.as("t1")
val df2 = df.as("t2")
val t = df1.orderBy("j").join(df2.orderBy("j"), $"t1.i" === $"t2.i", "right")
t.cache.orderBy($"t2.j").explain
{code}
Before fix:
{code:sql}
== Physical Plan ==
*(1) Sort [j#14 ASC NULLS FIRST], true, 0
+- Exchange rangepartitioning(j#14 ASC NULLS FIRST, 200)
+- InMemoryTableScan [i#5, j#6, i#13, j#14]
+- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
+- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
:- BroadcastExchange
HashedRelationBroadcastMode(List(cast(input[0, int, false] as...
: +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(j#6 ASC NULLS FIRST, 200)
: +- LocalTableScan [i#5, j#6]
+- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
+- ReusedExchange [i#13, j#14], Exchange
rangepartitioning(j#6 ASC NULLS FIRST, 200)
{code}
Better plan should avoid "Exchange rangepartitioning(j#14 ASC NULLS FIRST,
200)", like:
{code:sql}
== Physical Plan ==
*(1) Sort [j#14 ASC NULLS FIRST], true, 0
+- InMemoryTableScan [i#5, j#6, i#13, j#14]
+- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
+- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
:- BroadcastExchange
HashedRelationBroadcastMode(List(cast(input[0, int, false] as bigint)))
: +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(j#6 ASC NULLS FIRST, 200)
: +- LocalTableScan [i#5, j#6]
+- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
+- ReusedExchange [i#13, j#14], Exchange
rangepartitioning(j#6 ASC NULLS FIRST, 200)
{code}
was:
Currently, ReusedExchange would rewrite output partitioning if child's
partitioning is HashPartitioning, but it does not do the same when child's
partitioning is RangePartitioning, sometimes, it could introduce extra shuffle,
see:
{code}
val df = Seq(1 -> "a", 3 -> "c", 2 -> "b").toDF("i", "j")
val df1 = df.as("t1")
val df2 = df.as("t2")
val t = df1.orderBy("j").join(df2.orderBy("j"), $"t1.i" === $"t2.i", "right")
t.cache.orderBy($"t2.j").explain
{code}
Before fix:
{code:sql}
== Physical Plan ==
*(1) Sort [j#14 ASC NULLS FIRST], true, 0
+- Exchange rangepartitioning(j#14 ASC NULLS FIRST, 200)
+- InMemoryTableScan [i#5, j#6, i#13, j#14]
+- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
+- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
:- BroadcastExchange
HashedRelationBroadcastMode(List(cast(input[0, int, false] as bigint)))
: +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(j#6 ASC NULLS FIRST, 200)
: +- LocalTableScan [i#5, j#6]
+- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
+- ReusedExchange [i#13, j#14], Exchange
rangepartitioning(j#6 ASC NULLS FIRST, 200)
{code}
Better plan should avoid "Exchange rangepartitioning(j#14 ASC NULLS FIRST,
200)", like:
{code:sql}
== Physical Plan ==
*(1) Sort [j#14 ASC NULLS FIRST], true, 0
+- InMemoryTableScan [i#5, j#6, i#13, j#14]
+- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
+- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
:- BroadcastExchange
HashedRelationBroadcastMode(List(cast(input[0, int, false] as bigint)))
: +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
: +- Exchange rangepartitioning(j#6 ASC NULLS FIRST, 200)
: +- LocalTableScan [i#5, j#6]
+- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
+- ReusedExchange [i#13, j#14], Exchange
rangepartitioning(j#6 ASC NULLS FIRST, 200)
{code}
> ReusedExchange should rewrite output partitioning also when child's
> partitioning is RangePartitioning
> -----------------------------------------------------------------------------------------------------
>
> Key: SPARK-24556
> URL: https://issues.apache.org/jira/browse/SPARK-24556
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.3.2
> Reporter: yucai
> Priority: Major
>
> Currently, ReusedExchange would rewrite output partitioning if child's
> partitioning is HashPartitioning, but it does not do the same when child's
> partitioning is RangePartitioning, sometimes, it could introduce extra
> shuffle, see:
> {code:java}
> val df = Seq(1 -> "a", 3 -> "c", 2 -> "b").toDF("i", "j")
> val df1 = df.as("t1")
> val df2 = df.as("t2")
> val t = df1.orderBy("j").join(df2.orderBy("j"), $"t1.i" === $"t2.i", "right")
> t.cache.orderBy($"t2.j").explain
> {code}
> Before fix:
> {code:sql}
> == Physical Plan ==
> *(1) Sort [j#14 ASC NULLS FIRST], true, 0
> +- Exchange rangepartitioning(j#14 ASC NULLS FIRST, 200)
> +- InMemoryTableScan [i#5, j#6, i#13, j#14]
> +- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
> +- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
> :- BroadcastExchange
> HashedRelationBroadcastMode(List(cast(input[0, int, false] as...
> : +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
> : +- Exchange rangepartitioning(j#6 ASC NULLS FIRST,
> 200)
> : +- LocalTableScan [i#5, j#6]
> +- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
> +- ReusedExchange [i#13, j#14], Exchange
> rangepartitioning(j#6 ASC NULLS FIRST, 200)
> {code}
> Better plan should avoid "Exchange rangepartitioning(j#14 ASC NULLS FIRST,
> 200)", like:
> {code:sql}
> == Physical Plan ==
> *(1) Sort [j#14 ASC NULLS FIRST], true, 0
> +- InMemoryTableScan [i#5, j#6, i#13, j#14]
> +- InMemoryRelation [i#5, j#6, i#13, j#14], CachedRDDBuilder...
> +- *(2) BroadcastHashJoin [i#5], [i#13], RightOuter, BuildLeft
> :- BroadcastExchange
> HashedRelationBroadcastMode(List(cast(input[0, int, false] as bigint)))
> : +- *(1) Sort [j#6 ASC NULLS FIRST], true, 0
> : +- Exchange rangepartitioning(j#6 ASC NULLS FIRST, 200)
> : +- LocalTableScan [i#5, j#6]
> +- *(2) Sort [j#14 ASC NULLS FIRST], true, 0
> +- ReusedExchange [i#13, j#14], Exchange
> rangepartitioning(j#6 ASC NULLS FIRST, 200)
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
