[SPARK-53809][SQL] Add canonicalization for DataSourceV2ScanRelation

yhuang-db · peter-toth · commit 0732e44977ea · 2025-11-01T10:09:39.000+01:00
### What changes were proposed in this pull request? This PR proposes to add `doCanonicalize` function for DataSourceV2ScanRelation. The implementation is similar to [the one in BatchScanExec](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/v2/BatchScanExec.scala#L150), as well as the [the one in LogicalRelation](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/LogicalRelation.scala#L52). ### Why are the changes needed? Query optimization rules such as MergeScalarSubqueries check if two plans are identical by [comparing their canonicalized form](https://github.com/apache/spark/blob/master/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/MergeScalarSubqueries.scala#L219). For DSv2, for physical plan, the canonicalization goes down in the child hierarchy to the BatchScanExec, which [has a doCanonicalize function](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/v2/BatchScanExec.scala#L150); for logical plan, the canonicalization goes down to the DataSourceV2ScanRelation, which, however, does not have a doCanonicalize function. As a result, two logical plans who are semantically identical are not identified. Moreover, for reference, [DSv1 LogicalRelation](https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/LogicalRelation.scala#L52) also has `doCanonicalize()`. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? A new unit test is added to show that `MergeScalarSubqueries` is working for DataSourceV2ScanRelation. For a query ```sql select (select max(i) from df) as max_i, (select min(i) from df) as min_i ``` Before introducing the canonicalization, the plan is ``` == Parsed Logical Plan == 'Project [scalar-subquery#2 [] AS max_i#3, scalar-subquery#4 [] AS min_i#5] : :- 'Project [unresolvedalias('max('i))] : : +- 'UnresolvedRelation [df], [], false : +- 'Project [unresolvedalias('min('i))] : +- 'UnresolvedRelation [df], [], false +- OneRowRelation == Analyzed Logical Plan == max_i: int, min_i: int Project [scalar-subquery#2 [] AS max_i#3, scalar-subquery#4 [] AS min_i#5] : :- Aggregate [max(i#0) AS max(i)#7] : : +- SubqueryAlias df : : +- View (`df`, [i#0, j#1]) : : +- RelationV2[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 : +- Aggregate [min(i#10) AS min(i)#9] : +- SubqueryAlias df : +- View (`df`, [i#10, j#11]) : +- RelationV2[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 +- OneRowRelation == Optimized Logical Plan == Project [scalar-subquery#2 [] AS max_i#3, scalar-subquery#4 [] AS min_i#5] : :- Aggregate [max(i#0) AS max(i)#7] : : +- Project [i#0] : : +- RelationV2[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 : +- Aggregate [min(i#10) AS min(i)#9] : +- Project [i#10] : +- RelationV2[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 +- OneRowRelation == Physical Plan == AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 0 +- *(1) Project [Subquery subquery#2, [id=#32] AS max_i#3, Subquery subquery#4, [id=#33] AS min_i#5] : :- Subquery subquery#2, [id=#32] : : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) HashAggregate(keys=[], functions=[max(i#0)], output=[max(i)#7]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=58] +- *(1) HashAggregate(keys=[], functions=[partial_max(i#0)], output=[max#14]) +- *(1) Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == HashAggregate(keys=[], functions=[max(i#0)], output=[max(i)#7]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=19] +- HashAggregate(keys=[], functions=[partial_max(i#0)], output=[max#14]) +- Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] : +- Subquery subquery#4, [id=#33] : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) HashAggregate(keys=[], functions=[min(i#10)], output=[min(i)#9]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=63] +- *(1) HashAggregate(keys=[], functions=[partial_min(i#10)], output=[min#15]) +- *(1) Project [i#10] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == HashAggregate(keys=[], functions=[min(i#10)], output=[min(i)#9]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=30] +- HashAggregate(keys=[], functions=[partial_min(i#10)], output=[min#15]) +- Project [i#10] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- *(1) Scan OneRowRelation[] +- == Initial Plan == Project [Subquery subquery#2, [id=#32] AS max_i#3, Subquery subquery#4, [id=#33] AS min_i#5] : :- Subquery subquery#2, [id=#32] : : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) HashAggregate(keys=[], functions=[max(i#0)], output=[max(i)#7]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=58] +- *(1) HashAggregate(keys=[], functions=[partial_max(i#0)], output=[max#14]) +- *(1) Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == HashAggregate(keys=[], functions=[max(i#0)], output=[max(i)#7]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=19] +- HashAggregate(keys=[], functions=[partial_max(i#0)], output=[max#14]) +- Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] : +- Subquery subquery#4, [id=#33] : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) HashAggregate(keys=[], functions=[min(i#10)], output=[min(i)#9]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=63] +- *(1) HashAggregate(keys=[], functions=[partial_min(i#10)], output=[min#15]) +- *(1) Project [i#10] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == HashAggregate(keys=[], functions=[min(i#10)], output=[min(i)#9]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=30] +- HashAggregate(keys=[], functions=[partial_min(i#10)], output=[min#15]) +- Project [i#10] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- Scan OneRowRelation[] ``` After introducing the canonicalization, the plan is as following, where you can see **ReusedSubquery** ``` == Parsed Logical Plan == 'Project [scalar-subquery#2 [] AS max_i#3, scalar-subquery#4 [] AS min_i#5] : :- 'Project [unresolvedalias('max('i))] : : +- 'UnresolvedRelation [df], [], false : +- 'Project [unresolvedalias('min('i))] : +- 'UnresolvedRelation [df], [], false +- OneRowRelation == Analyzed Logical Plan == max_i: int, min_i: int Project [scalar-subquery#2 [] AS max_i#3, scalar-subquery#4 [] AS min_i#5] : :- Aggregate [max(i#0) AS max(i)#7] : : +- SubqueryAlias df : : +- View (`df`, [i#0, j#1]) : : +- RelationV2[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 : +- Aggregate [min(i#10) AS min(i)#9] : +- SubqueryAlias df : +- View (`df`, [i#10, j#11]) : +- RelationV2[i#10, j#11] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 +- OneRowRelation == Optimized Logical Plan == Project [scalar-subquery#2 [].max(i) AS max_i#3, scalar-subquery#4 [].min(i) AS min_i#5] : :- Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] : : +- Aggregate [max(i#0) AS max(i)#7, min(i#0) AS min(i)#9] : : +- Project [i#0] : : +- RelationV2[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 : +- Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] : +- Aggregate [max(i#0) AS max(i)#7, min(i#0) AS min(i)#9] : +- Project [i#0] : +- RelationV2[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5 +- OneRowRelation == Physical Plan == AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 0 +- *(1) Project [Subquery subquery#2, [id=#40].max(i) AS max_i#3, ReusedSubquery Subquery subquery#2, [id=#40].min(i) AS min_i#5] : :- Subquery subquery#2, [id=#40] : : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] +- *(2) HashAggregate(keys=[], functions=[max(i#0), min(i#0)], output=[max(i)#7, min(i)#9]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=71] +- *(1) HashAggregate(keys=[], functions=[partial_max(i#0), partial_min(i#0)], output=[max#16, min#17]) +- *(1) Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] +- HashAggregate(keys=[], functions=[max(i#0), min(i#0)], output=[max(i)#7, min(i)#9]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=22] +- HashAggregate(keys=[], functions=[partial_max(i#0), partial_min(i#0)], output=[max#16, min#17]) +- Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] : +- ReusedSubquery Subquery subquery#2, [id=#40] +- *(1) Scan OneRowRelation[] +- == Initial Plan == Project [Subquery subquery#2, [id=#40].max(i) AS max_i#3, Subquery subquery#4, [id=#41].min(i) AS min_i#5] : :- Subquery subquery#2, [id=#40] : : +- AdaptiveSparkPlan isFinalPlan=true +- == Final Plan == ResultQueryStage 1 +- *(2) Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] +- *(2) HashAggregate(keys=[], functions=[max(i#0), min(i#0)], output=[max(i)#7, min(i)#9]) +- ShuffleQueryStage 0 +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=71] +- *(1) HashAggregate(keys=[], functions=[partial_max(i#0), partial_min(i#0)], output=[max#16, min#17]) +- *(1) Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- == Initial Plan == Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] +- HashAggregate(keys=[], functions=[max(i#0), min(i#0)], output=[max(i)#7, min(i)#9]) +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=22] +- HashAggregate(keys=[], functions=[partial_max(i#0), partial_min(i#0)], output=[max#16, min#17]) +- Project [i#0] +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] : +- Subquery subquery#4, [id=#41] : +- AdaptiveSparkPlan isFinalPlan=false : +- Project [named_struct(max(i), max(i)#7, min(i), min(i)#9) AS mergedValue#14] : +- HashAggregate(keys=[], functions=[max(i#0), min(i#0)], output=[max(i)#7, min(i)#9]) : +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=37] : +- HashAggregate(keys=[], functions=[partial_max(i#0), partial_min(i#0)], output=[max#16, min#17]) : +- Project [i#0] : +- BatchScan class org.apache.spark.sql.connector.SimpleDataSourceV2$$anon$5[i#0, j#1] class org.apache.spark.sql.connector.SimpleDataSourceV2$MyScanBuilder RuntimeFilters: [] +- Scan OneRowRelation[] ``` ### Was this patch authored or co-authored using generative AI tooling? No Closes apache#52529 from yhuang-db/scan-canonicalization. Authored-by: yhuang-db <itisyuchuan@gmail.com> Signed-off-by: Peter Toth <peter.toth@gmail.com>
diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/execution/datasources/v2/DataSourceV2Relation.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/execution/datasources/v2/DataSourceV2Relation.scala
@@ -20,6 +20,7 @@ package org.apache.spark.sql.execution.datasources.v2
 import org.apache.spark.SparkException
 import org.apache.spark.sql.catalyst.analysis.{MultiInstanceRelation, NamedRelation, TimeTravelSpec}
 import org.apache.spark.sql.catalyst.expressions.{Attribute, AttributeMap, AttributeReference, Expression, SortOrder}
+import org.apache.spark.sql.catalyst.plans.QueryPlan
 import org.apache.spark.sql.catalyst.plans.logical.{ColumnStat, ExposesMetadataColumns, Histogram, HistogramBin, LeafNode, LogicalPlan, Statistics}
 import org.apache.spark.sql.catalyst.types.DataTypeUtils.toAttributes
 import org.apache.spark.sql.catalyst.util.{quoteIfNeeded, truncatedString, CharVarcharUtils}
@@ -175,6 +176,15 @@ case class DataSourceV2ScanRelation(
         Statistics(sizeInBytes = conf.defaultSizeInBytes)
     }
   }
+
+  override def doCanonicalize(): DataSourceV2ScanRelation = {
+    this.copy(
+      relation = this.relation.copy(
+        output = this.relation.output.map(QueryPlan.normalizeExpressions(_, this.relation.output))
+      ),
+      output = this.output.map(QueryPlan.normalizeExpressions(_, this.output))
+    )
+  }
 }
 
 /**
diff --git a/sql/core/src/test/scala/org/apache/spark/sql/connector/DataSourceV2Suite.scala b/sql/core/src/test/scala/org/apache/spark/sql/connector/DataSourceV2Suite.scala
@@ -26,6 +26,8 @@ import test.org.apache.spark.sql.connector._
 import org.apache.spark.SparkUnsupportedOperationException
 import org.apache.spark.sql.{AnalysisException, DataFrame, QueryTest, Row}
 import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.catalyst.expressions.ScalarSubquery
+import org.apache.spark.sql.catalyst.plans.logical.{Aggregate, Project}
 import org.apache.spark.sql.connector.catalog.{PartitionInternalRow, SupportsRead, Table, TableCapability, TableProvider}
 import org.apache.spark.sql.connector.catalog.TableCapability._
 import org.apache.spark.sql.connector.expressions.{Expression, FieldReference, Literal, NamedReference, NullOrdering, SortDirection, SortOrder, Transform}
@@ -36,6 +38,7 @@ import org.apache.spark.sql.connector.read.partitioning.{KeyGroupedPartitioning,
 import org.apache.spark.sql.execution.SortExec
 import org.apache.spark.sql.execution.adaptive.AdaptiveSparkPlanHelper
 import org.apache.spark.sql.execution.datasources.v2.{BatchScanExec, DataSourceV2Relation, DataSourceV2ScanRelation}
+import org.apache.spark.sql.execution.datasources.v2.DataSourceV2Implicits._
 import org.apache.spark.sql.execution.exchange.{Exchange, ShuffleExchangeExec}
 import org.apache.spark.sql.execution.vectorized.OnHeapColumnVector
 import org.apache.spark.sql.expressions.Window
@@ -976,6 +979,79 @@ class DataSourceV2Suite extends QueryTest with SharedSparkSession with AdaptiveS
       assert(result.length == 1)
     }
   }
+
+  test("SPARK-53809: scan canonicalization") {
+    val table = new SimpleDataSourceV2().getTable(CaseInsensitiveStringMap.empty())
+
+    def createDsv2ScanRelation(): DataSourceV2ScanRelation = {
+      val relation = DataSourceV2Relation.create(
+        table, None, None, CaseInsensitiveStringMap.empty())
+      val scan = relation.table.asReadable.newScanBuilder(relation.options).build()
+      DataSourceV2ScanRelation(relation, scan, relation.output)
+    }
+
+    // Create two DataSourceV2ScanRelation instances, representing the scan of the same table
+    val scanRelation1 = createDsv2ScanRelation()
+    val scanRelation2 = createDsv2ScanRelation()
+
+    // the two instances should not be the same, as they should have different attribute IDs
+    assert(scanRelation1 != scanRelation2,
+      "Two created DataSourceV2ScanRelation instances should not be the same")
+    assert(scanRelation1.output.map(_.exprId).toSet != scanRelation2.output.map(_.exprId).toSet,
+      "Output attributes should have different expression IDs before canonicalization")
+    assert(scanRelation1.relation.output.map(_.exprId).toSet !=
+      scanRelation2.relation.output.map(_.exprId).toSet,
+      "Relation output attributes should have different expression IDs before canonicalization")
+
+    // After canonicalization, the two instances should be equal
+    assert(scanRelation1.canonicalized == scanRelation2.canonicalized,
+      "Canonicalized DataSourceV2ScanRelation instances should be equal")
+  }
+
+  test("SPARK-53809: check mergeScalarSubqueries is effective for DataSourceV2ScanRelation") {
+    val df = spark.read.format(classOf[SimpleDataSourceV2].getName).load()
+    df.createOrReplaceTempView("df")
+
+    val query = sql("select (select max(i) from df) as max_i, (select min(i) from df) as min_i")
+    val optimizedPlan = query.queryExecution.optimizedPlan
+
+    // check optimizedPlan merged scalar subqueries `select max(i), min(i) from df`
+    val sub1 = optimizedPlan.asInstanceOf[Project].projectList.head.collect {
+      case s: ScalarSubquery => s
+    }
+    val sub2 = optimizedPlan.asInstanceOf[Project].projectList(1).collect {
+      case s: ScalarSubquery => s
+    }
+
+    // Both subqueries should reference the same merged plan `select max(i), min(i) from df`
+    assert(sub1.nonEmpty && sub2.nonEmpty, "Both scalar subqueries should exist")
+    assert(sub1.head.plan == sub2.head.plan,
+      "Both subqueries should reference the same merged plan")
+
+    // Extract the aggregate from the merged plan sub1
+    val agg = sub1.head.plan.collect {
+      case a: Aggregate => a
+    }.head
+
+    // Check that the aggregate contains both max(i) and min(i)
+    val aggFunctionSet = agg.aggregateExpressions.flatMap { expr =>
+      expr.collect {
+        case ae: org.apache.spark.sql.catalyst.expressions.aggregate.AggregateExpression =>
+          ae.aggregateFunction
+      }
+    }.toSet
+
+    assert(aggFunctionSet.size == 2, "Aggregate should contain exactly two aggregate functions")
+    assert(aggFunctionSet
+      .exists(_.isInstanceOf[org.apache.spark.sql.catalyst.expressions.aggregate.Max]),
+      "Aggregate should contain max(i)")
+    assert(aggFunctionSet
+      .exists(_.isInstanceOf[org.apache.spark.sql.catalyst.expressions.aggregate.Min]),
+      "Aggregate should contain min(i)")
+
+    // Verify the query produces correct results
+    checkAnswer(query, Row(9, 0))
+  }
 }
 
 case class RangeInputPartition(start: Int, end: Int) extends InputPartition
@@ -1081,6 +1157,18 @@ class SimpleDataSourceV2 extends TestingV2Source {
     override def planInputPartitions(): Array[InputPartition] = {
       Array(RangeInputPartition(0, 5), RangeInputPartition(5, 10))
     }
+
+    override def equals(obj: Any): Boolean = {
+      obj match {
+        case s: Scan =>
+          this.readSchema() == s.readSchema()
+        case _ => false
+      }
+    }
+
+    override def hashCode(): Int = {
+      this.readSchema().hashCode()
+    }
   }
 
   override def getTable(options: CaseInsensitiveStringMap): Table = new SimpleBatchTable {
