[SPARK-19981][SQL] Respect aliases in output partitioning of projects and aggregates by maropu · Pull Request #17400 · apache/spark

maropu · 2017-03-23T16:34:03Z

What changes were proposed in this pull request?

The current master might wrongly add shuffle operations when projects and aggregates in physical plans have aliases in output expressions. A concrete example is as follows;

scala> :paste
sql("SET spark.sql.autoBroadcastJoinThreshold=-1")
spark.range(10).selectExpr("id AS key", "0").repartition($"key").createOrReplaceTempView("df1")
spark.range(10).selectExpr("id AS key", "0").repartition($"key").createOrReplaceTempView("df2")
sql("""
SELECT * FROM
(SELECT key AS k from df1) t1
INNER JOIN
(SELECT key AS k from df2) t2
ON t1.k = t2.k
""").explain
== Physical Plan ==
*SortMergeJoin [k#56L], [k#57L], Inner
:- *Sort [k#56L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(k#56L, 200) // <--- Unnecessary ShuffleExchange
: +- *Project [key#39L AS k#56L]
: +- Exchange hashpartitioning(key#39L, 200)
: +- *Project [id#36L AS key#39L]
: +- *Range (0, 10, step=1, splits=Some(4))
+- *Sort [k#57L ASC NULLS FIRST], false, 0
+- ReusedExchange [k#57L], Exchange hashpartitioning(k#56L, 200)

In the query, the second Exchange is not necessary. The root cause is that the planner wrongly regards key and k as different attributes because they have different exprId. Then, it fails distribution requirement checks in EnsureRequirements. This pr proposes to handle these aliases in EnsureRequirements so as to check if the operators satisfy their output distribution requirements.

How was this patch tested?

Added tests in SQLQueryTestSuite.

SparkQA · 2017-03-23T18:32:35Z

Test build #75104 has finished for PR 17400 at commit b5d1038.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

MasterDDT · 2017-03-28T13:16:42Z

Just curious, how come the fix is not in this code?

https://github.com/maropu/spark/blob/b5d1038edffff5d65a6ddec20ea6eef186d25fc3/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/Canonicalize.scala#L41

So anywhere we compare expressions for semantic equality, we can say a#1 is the same as a#1 as b#2? (where 1 and 2 are the expressionIds)

MasterDDT · 2017-03-28T13:26:54Z

Here is a sort example with 1 partition. I believe the extra sort on newA is unnecessary.

scala> val df1 = Seq((1, 2), (3, 4)).toDF("a", "b").coalesce(1).sortWithinPartitions("a")
df1: org.apache.spark.sql.Dataset[org.apache.spark.sql.Row] = [a: int, b: int]
scala> val df2 = df1.selectExpr("a as newA", "b")
df2: org.apache.spark.sql.DataFrame = [newA: int, b: int]
scala> println(df1.join(df2, df1("a") === df2("newA")).queryExecution.executedPlan)
*SortMergeJoin [args=[a#37225], [newA#37232], Inner][outPart=PartitioningCollection(1, )][outOrder=List(a#37225 ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL, newA#37232:int%NONNULL, b#37243:int%NONNULL)]
:- *Sort [args=[a#37225 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37225 ASC%NONNULL)][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL)]
: +- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL)]
: +- LocalTableScan [args=[a#37225, b#37226]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37225:int%NONNULL, b#37226:int%NONNULL)]
+- *Sort [args=[newA#37232 ASC], false, 0][outPart=SinglePartition][outOrder=List(newA#37232 ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Project [args=[a#37242 AS newA#37232, b#37243]][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242 ASC%NONNULL)][output=ArrayBuffer(newA#37232:int%NONNULL, b#37243:int%NONNULL)]
+- *Sort [args=[a#37242 ASC], false, 0][outPart=SinglePartition][outOrder=ArrayBuffer(a#37242 ASC%NONNULL)][output=List(a#37242:int%NONNULL, b#37243:int%NONNULL)]
+- Coalesce [args=1][outPart=SinglePartition][outOrder=List()][output=List(a#37242:int%NONNULL, b#37243:int%NONNULL)]
+- LocalTableScan [args=[a#37242, b#37243]][outPart=UnknownPartitioning(0)][outOrder=List()][output=List(a#37242:int%NONNULL, b#37243:int%NONNULL)]```

maropu · 2017-03-28T15:57:35Z

ISTM the solution you suggested does not work because the planner actually compares references (that is, AttributeReferences output by child.outputPartioning) instead of Alias (See related code for details). So, IMO we cannot depend on semantically-equal checks you suggested.

allengeorge · 2017-03-29T22:17:22Z

I suggest the following code for outputOrdering:

override def outputOrdering: Seq[SortOrder] = child.outputOrdering.map {
case s @ SortOrder(e, _) =>
s.copy(child = maybeReplaceExpr(e))
case s =>
s
}```

maropu · 2017-04-17T08:14:06Z

@allengeorge yea, we could there. But, I think we should first make sure about how to fix this issue. I'm not sure that the approach of this pr is the best. cc: @gatorsmile

SparkQA · 2017-05-10T10:21:24Z

Test build #76738 has finished for PR 17400 at commit 91a412e.

This patch passes all tests.
This patch does not merge cleanly.
This patch adds no public classes.

SparkQA · 2017-05-10T10:28:50Z

Test build #76740 has finished for PR 17400 at commit 0492c0f.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

maropu · 2017-05-16T19:26:57Z

ping

gatorsmile · 2017-05-16T22:15:06Z

Could we fix the issue in EnsureRequirements? Aggregate operators can also introduce alias.

Aha, ok. I'll consider again.

SparkQA · 2017-05-17T21:44:04Z

Test build #77030 has finished for PR 17400 at commit ac3ffe4.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
abstract class AggregateExec extends UnaryExecNode
case class ProjectExec(

maropu · 2017-05-18T13:44:06Z

Jenkins, retest this please.

SparkQA · 2017-05-18T17:33:10Z

Test build #77053 has finished for PR 17400 at commit 49a1732.

This patch passes all tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
abstract class AggregateExec extends UnaryExecNode
case class ProjectExec(

maropu · 2017-05-20T14:47:08Z

@gatorsmile ping

maropu · 2017-05-23T15:43:05Z

ping

SparkQA · 2017-07-11T12:34:05Z

Test build #79520 has finished for PR 17400 at commit a7dd063.

This patch passes all tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
abstract class AggregateExec extends UnaryExecNode
case class ProjectExec(

SparkQA · 2017-12-18T01:04:18Z

Test build #85034 has finished for PR 17400 at commit e6ce117.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
abstract class AggregateExec extends UnaryExecNode
case class ProjectExec(

eyalfa · 2018-08-14T10:51:56Z

@maropu , any reason why this is on hold for so long?

eyalfa · 2018-08-14T10:56:16Z

this might do more than you'd like it to (at least if it behaves the way I understand collect first), i.e.
df.select($"x" as "x1, struct($"a" as "a1", $"b" as "b1") as "s1")
x1 and s1 are aliases, a1 and b1 are not. it could even get more complicated if there was an a1 alias in the top level projections list.

This pr only focuses on aliases, so the point you described above is out-of-scope in this pr. IMO more complicated cases should be fixed in follow-ups.

@maropu , I didn't aim for supporting complex partitioning expressions (which deserves its own separate PR), I meant that this code could introduce regressions by 'over-capturing' nested aliases.
my specific example is wrong since struct is transformed into a named struct (alias is replaced by an explicit name).

maropu · 2018-08-14T11:01:24Z

I think that's because the priority is not much high. This issue causes any problem in your query?

eyalfa · 2018-08-14T12:17:49Z

@maropu , yes it does :-)

maropu · 2018-08-14T12:45:53Z

If possible, could you describe that problem in your case to encourage this work?

eyalfa · 2018-08-14T13:02:18Z

in my use case, I aggregate a dataset, the use select to align columns with a case-class. I later try to join the resulting dataset based on the same columns used for aggregattion.
the join introduces shuffles (exchange nodes).

SparkQA · 2018-08-21T07:05:01Z

Test build #94999 has finished for PR 17400 at commit 1813738.

This patch fails due to an unknown error code, -9.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-08-21T10:09:52Z

Test build #95009 has finished for PR 17400 at commit 089f218.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-08-22T03:34:58Z

Test build #95079 has finished for PR 17400 at commit ec3e6d9.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-08-22T05:37:21Z

Test build #95078 has finished for PR 17400 at commit e288288.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-08-22T06:31:51Z

Test build #95080 has finished for PR 17400 at commit c67d11a.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2018-08-22T09:53:10Z

Test build #95097 has finished for PR 17400 at commit 91809e5.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

SparkQA · 2018-08-22T16:33:08Z

Test build #95102 has finished for PR 17400 at commit 5482b1b.

This patch passes all tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

maropu · 2018-09-14T05:57:22Z

retest this please

SparkQA · 2018-09-14T07:05:02Z

Test build #96061 has finished for PR 17400 at commit 5482b1b.

This patch fails due to an unknown error code, -9.
This patch merges cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

maropu · 2018-09-17T23:16:41Z

retest this please

gatorsmile · 2018-09-17T23:41:12Z

+import org.apache.spark.sql.catalyst.expressions.{Alias, Expression, NamedExpression}
+import org.apache.spark.sql.catalyst.plans.physical._
+
+trait AliasAwareOutputPartitioning extends UnaryExecNode {


We might need a general utility class for this. cc @maryannxue She did the similar things for the other projects in the past. Maybe @maryannxue can help deliver such a utility class?

ok, I'll wait for @maryannxue suggestion.

SparkQA · 2018-09-18T02:08:29Z

Test build #96161 has finished for PR 17400 at commit 5482b1b.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

maropu · 2018-09-18T02:10:15Z

retest this please

SparkQA · 2018-09-18T04:58:46Z

Test build #96164 has finished for PR 17400 at commit 5482b1b.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

SparkQA · 2018-10-22T12:20:31Z

Test build #97746 has finished for PR 17400 at commit 5482b1b.

This patch fails Spark unit tests.
This patch does not merge cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

SparkQA · 2018-10-22T13:09:09Z

Test build #97753 has finished for PR 17400 at commit 5482b1b.

This patch fails Spark unit tests.
This patch does not merge cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

SparkQA · 2018-10-22T13:39:20Z

Test build #97788 has finished for PR 17400 at commit 5482b1b.

This patch fails build dependency tests.
This patch does not merge cleanly.
This patch adds the following public classes (experimental):
trait AliasAwareOutputPartitioning extends UnaryExecNode

maropu force-pushed the SPARK-19981 branch 2 times, most recently from 91a412e to 0492c0fCompare May 10, 2017 08:03

gatorsmile reviewed May 16, 2017
View reviewed changes

maropu force-pushed the SPARK-19981 branch from 0492c0f to ac3ffe4Compare May 17, 2017 20:33

maropu changed the title ~~[SPARK-19981][SQL] Update output partitioning info. in ProjectExec when having aliases~~[SPARK-19981][SQL] Update output partitioning info. when children having aliasesMay 17, 2017

maropu force-pushed the SPARK-19981 branch from ac3ffe4 to 49a1732Compare May 18, 2017 12:02

maropu force-pushed the SPARK-19981 branch from 49a1732 to a7dd063Compare July 11, 2017 10:12

maropu force-pushed the SPARK-19981 branch from a7dd063 to e6ce117Compare December 18, 2017 00:57

eyalfa reviewed Aug 14, 2018
View reviewed changes

maropu force-pushed the SPARK-19981 branch from 7d1fbd2 to 1813738Compare August 21, 2018 06:50

maropu changed the title ~~[SPARK-19981][SQL] Update output partitioning info. when children having aliases~~[SPARK-19981][SQL] Respect aliases in output partitioning of projects and aggregatesAug 21, 2018

maropu force-pushed the SPARK-19981 branch 2 times, most recently from e288288 to ec3e6d9Compare August 22, 2018 03:31

maropu force-pushed the SPARK-19981 branch from ec3e6d9 to c67d11aCompare August 22, 2018 03:36

Another solution
ceb3806

maropu force-pushed the SPARK-19981 branch from c67d11a to 91809e5Compare August 22, 2018 07:45

Make AliasAwareOutputPartitioning trait
5482b1b

maropu force-pushed the SPARK-19981 branch from 91809e5 to 5482b1bCompare August 22, 2018 12:46

gatorsmile reviewed Sep 17, 2018
View reviewed changes

HeartSaVioR mentioned this pull request Feb 6, 2019
[SPARK-25951][SQL] Ignore aliases for distributions and orderings #22957
Closed

dongjoon-hyun added the SQL label Jun 14, 2019

maropu closed this Dec 20, 2019

imback82 mentioned this pull request Dec 20, 2019
[SPARK-30298][SQL] Respect aliases in output partitioning of projects and aggregates #26943
Closed

Uh oh!

Conversation

maropu commented Mar 23, 2017• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

SparkQA commented Mar 23, 2017

Uh oh!

MasterDDT commented Mar 28, 2017• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MasterDDT commented Mar 28, 2017• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

maropu commented Mar 28, 2017• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

allengeorge commented Mar 29, 2017

Uh oh!

maropu commented Apr 17, 2017

Uh oh!

SparkQA commented May 10, 2017

Uh oh!

SparkQA commented May 10, 2017

Uh oh!

maropu commented May 16, 2017

Uh oh!

gatorsmileMay 16, 2017

Choose a reason for hiding this comment

Uh oh!

maropuMay 17, 2017

Choose a reason for hiding this comment

Uh oh!

SparkQA commented May 17, 2017

Uh oh!

maropu commented May 18, 2017

Uh oh!

SparkQA commented May 18, 2017

Uh oh!

maropu commented May 20, 2017

Uh oh!

maropu commented May 23, 2017

Uh oh!

SparkQA commented Jul 11, 2017

Uh oh!

SparkQA commented Dec 18, 2017

Uh oh!

eyalfa commented Aug 14, 2018

Uh oh!

eyalfaAug 14, 2018

Choose a reason for hiding this comment

Uh oh!

maropuAug 21, 2018

Choose a reason for hiding this comment

Uh oh!

eyalfaAug 21, 2018

Choose a reason for hiding this comment

Uh oh!

maropu commented Aug 14, 2018

Uh oh!

eyalfa commented Aug 14, 2018

Uh oh!

maropu commented Aug 14, 2018

Uh oh!

eyalfa commented Aug 14, 2018

Uh oh!

SparkQA commented Aug 21, 2018

Uh oh!

SparkQA commented Aug 21, 2018

Uh oh!

SparkQA commented Aug 22, 2018

Uh oh!

SparkQA commented Aug 22, 2018

Uh oh!

SparkQA commented Aug 22, 2018

Uh oh!

SparkQA commented Aug 22, 2018

Uh oh!

SparkQA commented Aug 22, 2018

maropu commented Mar 23, 2017•
edited
Loading

MasterDDT commented Mar 28, 2017•
edited
Loading

MasterDDT commented Mar 28, 2017•
edited
Loading

maropu commented Mar 28, 2017•
edited
Loading