[SPARK-16134][SQL] optimizer rules for typed filter

## What changes were proposed in this pull request?

This PR adds 3 optimizer rules for typed filter:

1. push typed filter down through `SerializeFromObject` and eliminate the deserialization in filter condition.
2. pull typed filter up through `SerializeFromObject` and eliminate the deserialization in filter condition.
3. combine adjacent typed filters and share the deserialized object among all the condition expressions.

This PR also adds `TypedFilter` logical plan, to separate it from normal filter, so that the concept is more clear and it's easier to write optimizer rules.

## How was this patch tested?

`TypedFilterOptimizationSuite`

Author: Wenchen Fan <wenchen@databricks.com>

Closes #13846 from cloud-fan/filter.

(cherry picked from commit d063898)
Signed-off-by: Cheng Lian <lian@databricks.com>

Author: cloud-fan

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/dsl/package.scala

@@ -293,11 +293,7 @@ package object dsl {
 
       def where(condition: Expression): LogicalPlan = Filter(condition, logicalPlan)
 
-      def filter[T : Encoder](func: T => Boolean): LogicalPlan = {
-        val deserialized = logicalPlan.deserialize[T]
-        val condition = expressions.callFunction(func, BooleanType, deserialized.output.head)
-        Filter(condition, deserialized).serialize[T]
-      }
+      def filter[T : Encoder](func: T => Boolean): LogicalPlan = TypedFilter(func, logicalPlan)
 
       def serialize[T : Encoder]: LogicalPlan = CatalystSerde.serialize[T](logicalPlan)
 

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/ReferenceToExpressions.scala

@@ -45,6 +45,7 @@ case class ReferenceToExpressions(result: Expression, children: Seq[Expression])
     var maxOrdinal = -1
     result foreach {
       case b: BoundReference if b.ordinal > maxOrdinal => maxOrdinal = b.ordinal
+      case _ =>
     }
     if (maxOrdinal > children.length) {
       return TypeCheckFailure(s"The result expression need $maxOrdinal input expressions, but " +

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/Optimizer.scala

@@ -21,6 +21,7 @@ import scala.annotation.tailrec
 import scala.collection.immutable.HashSet
 import scala.collection.mutable.ArrayBuffer
 
+import org.apache.spark.api.java.function.FilterFunction
 import org.apache.spark.sql.catalyst.{CatalystConf, SimpleCatalystConf}
 import org.apache.spark.sql.catalyst.analysis._
 import org.apache.spark.sql.catalyst.catalog.{InMemoryCatalog, SessionCatalog}
@@ -109,8 +110,7 @@ abstract class Optimizer(sessionCatalog: SessionCatalog, conf: CatalystConf)
     Batch("Decimal Optimizations", fixedPoint,
       DecimalAggregates) ::
     Batch("Typed Filter Optimization", fixedPoint,
-      EmbedSerializerInFilter,
-      RemoveAliasOnlyProject) ::
+      CombineTypedFilters) ::
     Batch("LocalRelation", fixedPoint,
       ConvertToLocalRelation) ::
     Batch("OptimizeCodegen", Once,
@@ -205,15 +205,33 @@ object RemoveAliasOnlyProject extends Rule[LogicalPlan] {
 object EliminateSerialization extends Rule[LogicalPlan] {
   def apply(plan: LogicalPlan): LogicalPlan = plan transform {
     case d @ DeserializeToObject(_, _, s: SerializeFromObject)
-        if d.outputObjectType == s.inputObjectType =>
+        if d.outputObjAttr.dataType == s.inputObjAttr.dataType =>
       // Adds an extra Project here, to preserve the output expr id of `DeserializeToObject`.
       // We will remove it later in RemoveAliasOnlyProject rule.
-      val objAttr =
-        Alias(s.child.output.head, s.child.output.head.name)(exprId = d.output.head.exprId)
+      val objAttr = Alias(s.inputObjAttr, s.inputObjAttr.name)(exprId = d.outputObjAttr.exprId)
       Project(objAttr :: Nil, s.child)
+
     case a @ AppendColumns(_, _, _, s: SerializeFromObject)
-        if a.deserializer.dataType == s.inputObjectType =>
+        if a.deserializer.dataType == s.inputObjAttr.dataType =>
       AppendColumnsWithObject(a.func, s.serializer, a.serializer, s.child)
+
+    // If there is a `SerializeFromObject` under typed filter and its input object type is same with
+    // the typed filter's deserializer, we can convert typed filter to normal filter without
+    // deserialization in condition, and push it down through `SerializeFromObject`.
+    // e.g. `ds.map(...).filter(...)` can be optimized by this rule to save extra deserialization,
+    // but `ds.map(...).as[AnotherType].filter(...)` can not be optimized.
+    case f @ TypedFilter(_, _, s: SerializeFromObject)
+        if f.deserializer.dataType == s.inputObjAttr.dataType =>
+      s.copy(child = f.withObjectProducerChild(s.child))
+
+    // If there is a `DeserializeToObject` upon typed filter and its output object type is same with
+    // the typed filter's deserializer, we can convert typed filter to normal filter without
+    // deserialization in condition, and pull it up through `DeserializeToObject`.
+    // e.g. `ds.filter(...).map(...)` can be optimized by this rule to save extra deserialization,
+    // but `ds.filter(...).as[AnotherType].map(...)` can not be optimized.
+    case d @ DeserializeToObject(_, _, f: TypedFilter)
+        if d.outputObjAttr.dataType == f.deserializer.dataType =>
+      f.withObjectProducerChild(d.copy(child = f.child))
   }
 }
 
@@ -1606,54 +1624,30 @@ case class GetCurrentDatabase(sessionCatalog: SessionCatalog) extends Rule[Logic
 }
 
 /**
- * Typed [[Filter]] is by default surrounded by a [[DeserializeToObject]] beneath it and a
- * [[SerializeFromObject]] above it.  If these serializations can't be eliminated, we should embed
- * the deserializer in filter condition to save the extra serialization at last.
+ * Combines two adjacent [[TypedFilter]]s, which operate on same type object in condition, into one,
+ * mering the filter functions into one conjunctive function.
  */
-object EmbedSerializerInFilter extends Rule[LogicalPlan] {
+object CombineTypedFilters extends Rule[LogicalPlan] {
   def apply(plan: LogicalPlan): LogicalPlan = plan transform {
-    case s @ SerializeFromObject(_, Filter(condition, d: DeserializeToObject))
-      // SPARK-15632: Conceptually, filter operator should never introduce schema change. This
-      // optimization rule also relies on this assumption. However, Dataset typed filter operator
-      // does introduce schema changes in some cases. Thus, we only enable this optimization when
-      //
-      //  1. either input and output schemata are exactly the same, or
-      //  2. both input and output schemata are single-field schema and share the same type.
-      //
-      // The 2nd case is included because encoders for primitive types always have only a single
-      // field with hard-coded field name "value".
-      // TODO Cleans this up after fixing SPARK-15632.
-      if s.schema == d.child.schema || samePrimitiveType(s.schema, d.child.schema) =>
-
-      val numObjects = condition.collect {
-        case a: Attribute if a == d.output.head => a
-      }.length
-
-      if (numObjects > 1) {
-        // If the filter condition references the object more than one times, we should not embed
-        // deserializer in it as the deserialization will happen many times and slow down the
-        // execution.
-        // TODO: we can still embed it if we can make sure subexpression elimination works here.
-        s
-      } else {
-        val newCondition = condition transform {
-          case a: Attribute if a == d.output.head => d.deserializer
-        }
-        val filter = Filter(newCondition, d.child)
-
-        // Adds an extra Project here, to preserve the output expr id of `SerializeFromObject`.
-        // We will remove it later in RemoveAliasOnlyProject rule.
-        val objAttrs = filter.output.zip(s.output).map { case (fout, sout) =>
-          Alias(fout, fout.name)(exprId = sout.exprId)
-        }
-        Project(objAttrs, filter)
-      }
-  }
-
-  def samePrimitiveType(lhs: StructType, rhs: StructType): Boolean = {
-    (lhs, rhs) match {
-      case (StructType(Array(f1)), StructType(Array(f2))) => f1.dataType == f2.dataType
-      case _ => false
+    case t1 @ TypedFilter(_, _, t2 @ TypedFilter(_, _, child))
+        if t1.deserializer.dataType == t2.deserializer.dataType =>
+      TypedFilter(combineFilterFunction(t2.func, t1.func), t1.deserializer, child)
+  }
+
+  private def combineFilterFunction(func1: AnyRef, func2: AnyRef): Any => Boolean = {
+    (func1, func2) match {
+      case (f1: FilterFunction[_], f2: FilterFunction[_]) =>
+        input => f1.asInstanceOf[FilterFunction[Any]].call(input) &&
+          f2.asInstanceOf[FilterFunction[Any]].call(input)
+      case (f1: FilterFunction[_], f2) =>
+        input => f1.asInstanceOf[FilterFunction[Any]].call(input) &&
+          f2.asInstanceOf[Any => Boolean](input)
+      case (f1, f2: FilterFunction[_]) =>
+        input => f1.asInstanceOf[Any => Boolean].apply(input) &&
+          f2.asInstanceOf[FilterFunction[Any]].call(input)
+      case (f1, f2) =>
+        input => f1.asInstanceOf[Any => Boolean].apply(input) &&
+          f2.asInstanceOf[Any => Boolean].apply(input)
     }
   }
 }

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/object.scala

@@ -17,11 +17,15 @@
 
 package org.apache.spark.sql.catalyst.plans.logical
 
+import scala.language.existentials
+
+import org.apache.spark.api.java.function.FilterFunction
 import org.apache.spark.broadcast.Broadcast
 import org.apache.spark.sql.{Encoder, Row}
 import org.apache.spark.sql.catalyst.analysis.UnresolvedDeserializer
 import org.apache.spark.sql.catalyst.encoders._
 import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.expressions.objects.Invoke
 import org.apache.spark.sql.types._
 
 object CatalystSerde {
@@ -45,13 +49,11 @@ object CatalystSerde {
  */
 trait ObjectProducer extends LogicalPlan {
   // The attribute that reference to the single object field this operator outputs.
-  protected def outputObjAttr: Attribute
+  def outputObjAttr: Attribute
 
   override def output: Seq[Attribute] = outputObjAttr :: Nil
 
   override def producedAttributes: AttributeSet = AttributeSet(outputObjAttr)
-
-  def outputObjectType: DataType = outputObjAttr.dataType
 }
 
 /**
@@ -64,7 +66,7 @@ trait ObjectConsumer extends UnaryNode {
   // This operator always need all columns of its child, even it doesn't reference to.
   override def references: AttributeSet = child.outputSet
 
-  def inputObjectType: DataType = child.output.head.dataType
+  def inputObjAttr: Attribute = child.output.head
 }
 
 /**
@@ -167,6 +169,43 @@ case class MapElements(
     outputObjAttr: Attribute,
     child: LogicalPlan) extends ObjectConsumer with ObjectProducer
 
+object TypedFilter {
+  def apply[T : Encoder](func: AnyRef, child: LogicalPlan): TypedFilter = {
+    TypedFilter(func, UnresolvedDeserializer(encoderFor[T].deserializer), child)
+  }
+}
+
+/**
+ * A relation produced by applying `func` to each element of the `child` and filter them by the
+ * resulting boolean value.
+ *
+ * This is logically equal to a normal [[Filter]] operator whose condition expression is decoding
+ * the input row to object and apply the given function with decoded object. However we need the
+ * encapsulation of [[TypedFilter]] to make the concept more clear and make it easier to write
+ * optimizer rules.
+ */
+case class TypedFilter(
+    func: AnyRef,
+    deserializer: Expression,
+    child: LogicalPlan) extends UnaryNode {
+
+  override def output: Seq[Attribute] = child.output
+
+  def withObjectProducerChild(obj: LogicalPlan): Filter = {
+    assert(obj.output.length == 1)
+    Filter(typedCondition(obj.output.head), obj)
+  }
+
+  def typedCondition(input: Expression): Expression = {
+    val (funcClass, methodName) = func match {
+      case m: FilterFunction[_] => classOf[FilterFunction[_]] -> "call"
+      case _ => classOf[Any => Boolean] -> "apply"
+    }
+    val funcObj = Literal.create(func, ObjectType(funcClass))
+    Invoke(funcObj, methodName, BooleanType, input :: Nil)
+  }
+}
+
 /** Factory for constructing new `AppendColumn` nodes. */
 object AppendColumns {
   def apply[T : Encoder, U : Encoder](

sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/optimizer/TypedFilterOptimizationSuite.scala

@@ -19,12 +19,11 @@ package org.apache.spark.sql.catalyst.optimizer
 
 import scala.reflect.runtime.universe.TypeTag
 
-import org.apache.spark.sql.catalyst.analysis.UnresolvedDeserializer
 import org.apache.spark.sql.catalyst.dsl.expressions._
 import org.apache.spark.sql.catalyst.dsl.plans._
-import org.apache.spark.sql.catalyst.encoders.{encoderFor, ExpressionEncoder}
+import org.apache.spark.sql.catalyst.encoders.ExpressionEncoder
 import org.apache.spark.sql.catalyst.plans.PlanTest
-import org.apache.spark.sql.catalyst.plans.logical.{LocalRelation, LogicalPlan}
+import org.apache.spark.sql.catalyst.plans.logical.{LocalRelation, LogicalPlan, TypedFilter}
 import org.apache.spark.sql.catalyst.rules.RuleExecutor
 import org.apache.spark.sql.types.BooleanType
 
@@ -33,44 +32,91 @@ class TypedFilterOptimizationSuite extends PlanTest {
     val batches =
       Batch("EliminateSerialization", FixedPoint(50),
         EliminateSerialization) ::
-      Batch("EmbedSerializerInFilter", FixedPoint(50),
-        EmbedSerializerInFilter) :: Nil
+      Batch("CombineTypedFilters", FixedPoint(50),
+        CombineTypedFilters) :: Nil
   }
 
   implicit private def productEncoder[T <: Product : TypeTag] = ExpressionEncoder[T]()
 
-  test("back to back filter") {
+  test("filter after serialize with the same object type") {
     val input = LocalRelation('_1.int, '_2.int)
-    val f1 = (i: (Int, Int)) => i._1 > 0
-    val f2 = (i: (Int, Int)) => i._2 > 0
+    val f = (i: (Int, Int)) => i._1 > 0
 
-    val query = input.filter(f1).filter(f2).analyze
+    val query = input
+      .deserialize[(Int, Int)]
+      .serialize[(Int, Int)]
+      .filter(f).analyze
 
     val optimized = Optimize.execute(query)
 
-    val expected = input.deserialize[(Int, Int)]
-      .where(callFunction(f1, BooleanType, 'obj))
-      .select('obj.as("obj"))
-      .where(callFunction(f2, BooleanType, 'obj))
+    val expected = input
+      .deserialize[(Int, Int)]
+      .where(callFunction(f, BooleanType, 'obj))
       .serialize[(Int, Int)].analyze
 
     comparePlans(optimized, expected)
   }
 
-  // TODO: Remove this after we completely fix SPARK-15632 by adding optimization rules
-  // for typed filters.
-  ignore("embed deserializer in typed filter condition if there is only one filter") {
+  test("filter after serialize with different object types") {
+    val input = LocalRelation('_1.int, '_2.int)
+    val f = (i: OtherTuple) => i._1 > 0
+
+    val query = input
+      .deserialize[(Int, Int)]
+      .serialize[(Int, Int)]
+      .filter(f).analyze
+    val optimized = Optimize.execute(query)
+    comparePlans(optimized, query)
+  }
+
+  test("filter before deserialize with the same object type") {
     val input = LocalRelation('_1.int, '_2.int)
     val f = (i: (Int, Int)) => i._1 > 0
 
-    val query = input.filter(f).analyze
+    val query = input
+      .filter(f)
+      .deserialize[(Int, Int)]
+      .serialize[(Int, Int)].analyze
 
     val optimized = Optimize.execute(query)
 
-    val deserializer = UnresolvedDeserializer(encoderFor[(Int, Int)].deserializer)
-    val condition = callFunction(f, BooleanType, deserializer)
-    val expected = input.where(condition).select('_1.as("_1"), '_2.as("_2")).analyze
+    val expected = input
+      .deserialize[(Int, Int)]
+      .where(callFunction(f, BooleanType, 'obj))
+      .serialize[(Int, Int)].analyze
 
     comparePlans(optimized, expected)
   }
+
+  test("filter before deserialize with different object types") {
+    val input = LocalRelation('_1.int, '_2.int)
+    val f = (i: OtherTuple) => i._1 > 0
+
+    val query = input
+      .filter(f)
+      .deserialize[(Int, Int)]
+      .serialize[(Int, Int)].analyze
+    val optimized = Optimize.execute(query)
+    comparePlans(optimized, query)
+  }
+
+  test("back to back filter with the same object type") {
+    val input = LocalRelation('_1.int, '_2.int)
+    val f1 = (i: (Int, Int)) => i._1 > 0
+    val f2 = (i: (Int, Int)) => i._2 > 0
+
+    val query = input.filter(f1).filter(f2).analyze
+    val optimized = Optimize.execute(query)
+    assert(optimized.collect { case t: TypedFilter => t }.length == 1)
+  }
+
+  test("back to back filter with different object types") {
+    val input = LocalRelation('_1.int, '_2.int)
+    val f1 = (i: (Int, Int)) => i._1 > 0
+    val f2 = (i: OtherTuple) => i._2 > 0
+
+    val query = input.filter(f1).filter(f2).analyze
+    val optimized = Optimize.execute(query)
+    assert(optimized.collect { case t: TypedFilter => t }.length == 2)
+  }
 }

sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala

@@ -1964,11 +1964,7 @@ class Dataset[T] private[sql](
    */
   @Experimental
   def filter(func: T => Boolean): Dataset[T] = {
-    val deserializer = UnresolvedDeserializer(encoderFor[T].deserializer)
-    val function = Literal.create(func, ObjectType(classOf[T => Boolean]))
-    val condition = Invoke(function, "apply", BooleanType, deserializer :: Nil)
-    val filter = Filter(condition, logicalPlan)
-    withTypedPlan(filter)
+    withTypedPlan(TypedFilter(func, logicalPlan))
   }
 
   /**
@@ -1981,11 +1977,7 @@ class Dataset[T] private[sql](
    */
   @Experimental
   def filter(func: FilterFunction[T]): Dataset[T] = {
-    val deserializer = UnresolvedDeserializer(encoderFor[T].deserializer)
-    val function = Literal.create(func, ObjectType(classOf[FilterFunction[T]]))
-    val condition = Invoke(function, "call", BooleanType, deserializer :: Nil)
-    val filter = Filter(condition, logicalPlan)
-    withTypedPlan(filter)
+    withTypedPlan(TypedFilter(func, logicalPlan))
   }
 
   /**

sql/core/src/main/scala/org/apache/spark/sql/execution/SparkStrategies.scala

@@ -385,6 +385,8 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
         execution.ProjectExec(projectList, planLater(child)) :: Nil
       case logical.Filter(condition, child) =>
         execution.FilterExec(condition, planLater(child)) :: Nil
+      case f: logical.TypedFilter =>
+        execution.FilterExec(f.typedCondition(f.deserializer), planLater(f.child)) :: Nil
       case e @ logical.Expand(_, _, child) =>
         execution.ExpandExec(e.projections, e.output, planLater(child)) :: Nil
       case logical.Window(windowExprs, partitionSpec, orderSpec, child) =>