Correlated scalar subqueries in Apache Spark SQL

on waitingforcode.com

Correlated scalar subqueries in Apache Spark SQL

Some weeks ago I presented correlated scalar subqueries in the example of PostgreSQL. However they can also be found in the Big Data processing systems, as for instance BigQuery or Apache Spark SQL.

This post focuses on correlated scalar subqueries in Apache Spark SQL and presents them in 3 sections. The first two show some sample queries with correlated subquery. Both contain the code and the execution plan output. The final section summarizes them by explaining applied rules.

Correlated aggregation query

The first executed query will retrieve all articles with the number of views greater than the average number of view in their category. The query is defined in the following test case:

import SparkLocalSession.implicits._

private val SportId = 1
private val GlobalNewsId = 2
private val ArticlesDataFrame = Seq(
  Article(1, SportId, "News about football", 30), Article(2, SportId, "News about tennis", 10),
  Article(3, GlobalNewsId, "Global news 1", 20), Article(4, GlobalNewsId, "Global news 2", 40)
).toDF
private val CategoriesDataFrame = Seq(
  Category(SportId, 2), Category(GlobalNewsId, 2)
).toDF
ArticlesDataFrame.createOrReplaceTempView("articles")
CategoriesDataFrame.createOrReplaceTempView("categories")

it should "should execute aggregation" in {
  val sqlResult = SparkLocalSession.sql(
    """
      |SELECT title, views FROM articles a WHERE a.views > (
      | SELECT AVG(views) AS avg_views_category FROM articles WHERE categoryId = a.categoryId
      | GROUP BY categoryId
      |)
    """.stripMargin)

  val mappedResults = sqlResult.collect().map(row => s"${row.getString(0)};${row.getInt(1)}")

  mappedResults should have size 2
  mappedResults should contain allOf("News about football;30", "Global news 2;40")
}

The execution plan looks like that:

== Parsed Logical Plan ==
'Project ['title, 'views]
+- 'Filter ('a.views > scalar-subquery#31 [])
   :  +- 'Aggregate ['categoryId], ['AVG('views) AS avg_views_category#30]
   :     +- 'Filter ('categoryId = 'a.categoryId)
   :        +- 'UnresolvedRelation `articles`
   +- 'SubqueryAlias a
      +- 'UnresolvedRelation `articles`

== Analyzed Logical Plan ==
title: string, views: int
Project [title#6, views#7]
+- Filter (cast(views#7 as double) > scalar-subquery#31 [categoryId#5])
   :  +- Aggregate [categoryId#5], [avg(cast(views#7 as bigint)) AS avg_views_category#30]
   :     +- Filter (categoryId#5 = outer(categoryId#5))
   :        +- SubqueryAlias articles
   :           +- LocalRelation [articleId#4, categoryId#5, title#6, views#7]
   +- SubqueryAlias a
      +- SubqueryAlias articles
         +- LocalRelation [articleId#4, categoryId#5, title#6, views#7]

== Optimized Logical Plan ==
Project [title#6, views#7]
+- Join Inner, ((cast(views#7 as double) > avg_views_category#30) && (categoryId#5#43 = categoryId#5))
   :- LocalRelation [categoryId#5, title#6, views#7]
   +- Filter isnotnull(avg_views_category#30)
      +- Aggregate [categoryId#5], [avg(cast(views#7 as bigint)) AS avg_views_category#30, categoryId#5 AS categoryId#5#43]
         +- LocalRelation [categoryId#5, views#7]

== Physical Plan ==
*(3) Project [title#6, views#7]
+- *(3) BroadcastHashJoin [categoryId#5], [categoryId#5#43], Inner, BuildRight, (cast(views#7 as double) > avg_views_category#30)
   :- LocalTableScan [categoryId#5, title#6, views#7]
   +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[1, int, false] as bigint)))
      +- *(2) Filter isnotnull(avg_views_category#30)
         +- *(2) HashAggregate(keys=[categoryId#5], functions=[avg(cast(views#7 as bigint))], output=[avg_views_category#30, categoryId#5#43])
            +- Exchange hashpartitioning(categoryId#5, 200)
               +- *(1) HashAggregate(keys=[categoryId#5], functions=[partial_avg(cast(views#7 as bigint))], output=[categoryId#5, sum#46, count#47L])
                  +- LocalTableScan [categoryId#5, views#7]

After reading the execution plan we can see that Spark transforms the correlated subquery to an inner hash join where it broadcasts one part of the join side to executors. It happens in this specific case because one side of the relation is small enough to be send throughout the network. In other situations this plan could of course be different.

Correlated projection query

For the second example we'll see what happens with the correlated query used directly in the SELECT statement:

it should "be used in the projection" in {
  val sqlResult = SparkLocalSession.sql(
    """
      |SELECT a.title, a.categoryId, (
      | SELECT FIRST(c.allArticles) FROM categories c WHERE c.categoryId = a.categoryId
      |) AS allArticlesInCategory FROM articles a
    """.stripMargin)

  val mappedResults = sqlResult.collect().map(row => s"${row.getString(0)};${row.getInt(1)};${row.getInt(2)}")

  mappedResults should have size 4
  mappedResults should contain allOf("News about football;1;2", "News about tennis;1;2",
    "Global news 1;2;2", "Global news 2;2;2")
}

The execution plan looks like:

== Parsed Logical Plan ==
'Project ['a.title, 'a.categoryId, scalar-subquery#38 [] AS allArticlesInCategory#39]
:  +- 'Project [unresolvedalias(first('c.allArticles, false), None)]
:     +- 'Filter ('c.categoryId = 'a.categoryId)
:        +- 'SubqueryAlias c
:           +- 'UnresolvedRelation `categories`
+- 'SubqueryAlias a
   +- 'UnresolvedRelation `articles`

== Analyzed Logical Plan ==
title: string, categoryId: int, allArticlesInCategory: int
Project [title#6, categoryId#5, scalar-subquery#38 [categoryId#5] AS allArticlesInCategory#39]
:  +- Aggregate [first(allArticles#16, false) AS first(allArticles, false)#50]
:     +- Filter (categoryId#15 = outer(categoryId#5))
:        +- SubqueryAlias c
:           +- SubqueryAlias categories
:              +- LocalRelation [categoryId#15, allArticles#16]
+- SubqueryAlias a
   +- SubqueryAlias articles
      +- LocalRelation [articleId#4, categoryId#5, title#6, views#7]

== Optimized Logical Plan ==
Project [title#6, categoryId#5, first(allArticles, false)#50 AS allArticlesInCategory#39]
+- Join LeftOuter, (categoryId#15 = categoryId#5)
   :- LocalRelation [categoryId#5, title#6]
   +- Aggregate [categoryId#15], [first(allArticles#16, false) AS first(allArticles, false)#50, categoryId#15]
      +- LocalRelation [categoryId#15, allArticles#16]

== Physical Plan ==
*(3) Project [title#6, categoryId#5, first(allArticles, false)#50 AS allArticlesInCategory#39]
+- *(3) BroadcastHashJoin [categoryId#5], [categoryId#15], LeftOuter, BuildRight
   :- LocalTableScan [categoryId#5, title#6]
   +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[1, int, false] as bigint)))
      +- *(2) HashAggregate(keys=[categoryId#15], functions=[first(allArticles#16, false)], output=[first(allArticles, false)#50, categoryId#15])
         +- Exchange hashpartitioning(categoryId#15, 200)
            +- *(1) HashAggregate(keys=[categoryId#15], functions=[partial_first(allArticles#16, false)], output=[categoryId#15, first#53, valueSet#54])
               +- LocalTableScan [categoryId#15, allArticles#16]

Here too the execution plan was transformed to a broadcast join and that for the same reasons as previously.

Correlated subqueries rules

Two rules related to the correlated subqueries are used in the code executed above. First of them concerns the rewriting. Both queries are rewritten to the joins because of the RewriteCorrelatedScalarSubquery rule:

=== Applying Rule org.apache.spark.sql.catalyst.optimizer.RewriteCorrelatedScalarSubquery ===
!Project [title#6, categoryId#5, scalar-subquery#38 [(categoryId#15 = categoryId#5)] AS allArticlesInCategory#39]   Project [title#6, categoryId#5, first(allArticles, fa
lse)#40 AS allArticlesInCategory#39]
!:  +- Aggregate [categoryId#15], [first(allArticles#16, false) AS first(allArticles, false)#40, categoryId#15]     +- Project [articleId#4, categoryId#5, title#6, views
#7, first(allArticles, false)#40]
!:     +- LocalRelation [categoryId#15, allArticles#16]                                                                +- Join LeftOuter, (categoryId#15 = categoryId#5)
!+- LocalRelation [articleId#4, categoryId#5, title#6, views#7]                                                           :- LocalRelation [articleId#4, categoryId#5, ti
tle#6, views#7]
!                                                                                                                         +- Aggregate [categoryId#15], [first(allArticle
s#16, false) AS first(allArticles, false)#40, categoryId#15]
!                                                                                                                            +- LocalRelation [categoryId#15, allArticles
#16]

The rewriting is based on the constructLeftJoins(child: LogicalPlan, subqueries: ArrayBuffer[ScalarSubquery]) method. The transformation logic is quite simple. The subquery is splitted in 2 parts: the query and the conditions joining inner and outer queries. From that it's quite easy to construct a left outer join:

subqueries.foldLeft(child) {
  case (currentChild, ScalarSubquery(query, conditions, _)) =>
    val origOutput = query.output.head

    val resultWithZeroTups = evalSubqueryOnZeroTups(query)
    if (resultWithZeroTups.isEmpty) {
      // CASE 1: Subquery guaranteed not to have the COUNT bug
      Project(
        currentChild.output :+ origOutput,
        Join(currentChild, query, LeftOuter, conditions.reduceOption(And)))
    // ...

But apart this simple transformation the method handles also some bugs discovered 2 years ago that makes it more complicated than that.

Correlated subqueries fixes

Three bugs were discovered after adding the correlated subqueries support based on the mismatch of the tuples from the outer query with the subquery:

  • filter
        spark-sql> create table R(a integer) as values (1);
        spark-sql> create table S(b integer);
        spark-sql> select R.a from R 
                >     where (select count(*) from S where R.a = S.b) = 0;
        Time taken: 2.139 seconds                                                       
        spark-sql> 
        (returns zero rows; the answer should be one row of '1')
        
  • select
        spark-sql> select R.a, 
             >     (select count(*) from S where R.a = S.b) as cnt 
             > from R;
        1	NULL
        (the answer should be "1 0")
        
  • expression with null checks
        spark-sql> select R.a from R 
                > where (select sum(S.b) is null from S where R.a = S.b);
        (returns zero rows, should return one row)
        

All of them were fixed in SPARK-15370 and the fix obviously impacted the simplicity of the method transforming the correlated subquery to the left outer join.

Another applied rule is PullupCorrelatedPredicates. It moves the correlated predicates from the subquery to the outer query. We say then that the parts of the correlated subqueries are pulled up to the query above them. It's also known as unnesting or decorrelation. We can observe the use of this rule in the final physical plans where the filter clause was transformed to the ON condition. It's also visible earlier in the query optimization, as shown in this log fragment:

=== Applying Rule org.apache.spark.sql.catalyst.optimizer.PullupCorrelatedPredicates ===
!Project [title#6, categoryId#5, scalar-subquery#38 [categoryId#5] AS allArticlesInCategory#39]   Project [title#6, categoryId#5, scalar-subquery#38 [(categoryId#15 = ca
tegoryId#5)] AS allArticlesInCategory#39]
!:  +- Aggregate [first(allArticles#16, false) AS first(allArticles, false)#40]                   :  +- Aggregate [categoryId#15], [first(allArticles#16, false) AS first
(allArticles, false)#40, categoryId#15]
!:     +- Filter (categoryId#15 = outer(categoryId#5))                                            :     +- SubqueryAlias c
!:        +- SubqueryAlias c                                                                      :        +- SubqueryAlias categories
!:           +- SubqueryAlias categories                                                          :           +- LocalRelation [categoryId#15, allArticles#16]
!:              +- LocalRelation [categoryId#15, allArticles#16]                                  +- LocalRelation [articleId#4, categoryId#5, title#6, views#7]
!+- LocalRelation [articleId#4, categoryId#5, title#6, views#7]                                   

The correlated subqueries in Spark SQL are rewritten to the queries where the subquery is joined to the outer one with the left outer join. It happens independently on the subquery location. As we could see in this post, the rewriting is done as well for the subqueries defined in the projection part as well for the ones from the filtering part. Two rules are involved in this transformation: RewriteCorrelatedScalarSubquery and PullupCorrelatedPredicate. The former one converts the correlated subquery to the left outer join while the latter one extracts the filtering clauses and moves them to the upper level.

Share, like or comment this post on Twitter: