[SPARK-26448][SQL] retain the difference between 0.0 and -0.0

## What changes were proposed in this pull request?

In #23043 , we introduced a behavior change: Spark users are not able to distinguish 0.0 and -0.0 anymore.

This PR proposes an alternative fix to the original bug, to retain the difference between 0.0 and -0.0 inside Spark.

The idea is, we can rewrite the window partition key, join key and grouping key during logical phase, to normalize the special floating numbers. Thus only operators care about special floating numbers need to pay the perf overhead, and end users can distinguish -0.0.

## How was this patch tested?

existing test

Closes #23388 from cloud-fan/minor.

Authored-by: Wenchen Fan <wenchen@databricks.com>
Signed-off-by: gatorsmile <gatorsmile@gmail.com>

Author: cloud-fan

docs/sql-migration-guide-upgrade.md

@@ -25,7 +25,7 @@ displayTitle: Spark SQL Upgrading Guide
 
   - In Spark version 2.4 and earlier, `Dataset.groupByKey` results to a grouped dataset with key attribute wrongly named as "value", if the key is non-struct type, e.g. int, string, array, etc. This is counterintuitive and makes the schema of aggregation queries weird. For example, the schema of `ds.groupByKey(...).count()` is `(value, count)`. Since Spark 3.0, we name the grouping attribute to "key". The old behaviour is preserved under a newly added configuration `spark.sql.legacy.dataset.nameNonStructGroupingKeyAsValue` with a default value of `false`.
 
-  - In Spark version 2.4 and earlier, float/double -0.0 is semantically equal to 0.0, but users can still distinguish them via `Dataset.show`, `Dataset.collect` etc. Since Spark 3.0, float/double -0.0 is replaced by 0.0 internally, and users can't distinguish them any more.
+  - In Spark version 2.4 and earlier, float/double -0.0 is semantically equal to 0.0, but -0.0 and 0.0 are considered as different values when used in aggregate grouping keys, window partition keys and join keys. Since Spark 3.0, this bug is fixed. For example, `Seq(-0.0, 0.0).toDF("d").groupBy("d").count()` returns `[(0.0, 2)]` in Spark 3.0, and `[(0.0, 1), (-0.0, 1)]` in Spark 2.4 and earlier.
 
   - In Spark version 2.4 and earlier, users can create a map with duplicated keys via built-in functions like `CreateMap`, `StringToMap`, etc. The behavior of map with duplicated keys is undefined, e.g. map look up respects the duplicated key appears first, `Dataset.collect` only keeps the duplicated key appears last, `MapKeys` returns duplicated keys, etc. Since Spark 3.0, these built-in functions will remove duplicated map keys with last wins policy. Users may still read map values with duplicated keys from data sources which do not enforce it (e.g. Parquet), the behavior will be undefined.
 

sql/catalyst/src/main/java/org/apache/spark/sql/catalyst/expressions/codegen/UnsafeWriter.java

@@ -198,46 +198,11 @@ protected final void writeLong(long offset, long value) {
     Platform.putLong(getBuffer(), offset, value);
   }
 
-  // We need to take care of NaN and -0.0 in several places:
-  //   1. When compare values, different NaNs should be treated as same, `-0.0` and `0.0` should be
-  //      treated as same.
-  //   2. In GROUP BY, different NaNs should belong to the same group, -0.0 and 0.0 should belong
-  //      to the same group.
-  //   3. As join keys, different NaNs should be treated as same, `-0.0` and `0.0` should be
-  //      treated as same.
-  //   4. As window partition keys, different NaNs should be treated as same, `-0.0` and `0.0`
-  //      should be treated as same.
-  //
-  // Case 1 is fine, as we handle NaN and -0.0 well during comparison. For complex types, we
-  // recursively compare the fields/elements, so it's also fine.
-  //
-  // Case 2, 3 and 4 are problematic, as they compare `UnsafeRow` binary directly, and different
-  // NaNs have different binary representation, and the same thing happens for -0.0 and 0.0.
-  //
-  // Here we normalize NaN and -0.0, so that `UnsafeProjection` will normalize them when writing
-  // float/double columns and nested fields to `UnsafeRow`.
-  //
-  // Note that, we must do this for all the `UnsafeProjection`s, not only the ones that extract
-  // join/grouping/window partition keys. `UnsafeProjection` copies unsafe data directly for complex
-  // types, so nested float/double may not be normalized. We need to make sure that all the unsafe
-  // data(`UnsafeRow`, `UnsafeArrayData`, `UnsafeMapData`) will have flat/double normalized during
-  // creation.
   protected final void writeFloat(long offset, float value) {
-    if (Float.isNaN(value)) {
-      value = Float.NaN;
-    } else if (value == -0.0f) {
-      value = 0.0f;
-    }
     Platform.putFloat(getBuffer(), offset, value);
   }
 
-  // See comments for `writeFloat`.
   protected final void writeDouble(long offset, double value) {
-    if (Double.isNaN(value)) {
-      value = Double.NaN;
-    } else if (value == -0.0d) {
-      value = 0.0d;
-    }
     Platform.putDouble(getBuffer(), offset, value);
   }
 }

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

@@ -0,0 +1,198 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.sql.catalyst.optimizer
+
+import org.apache.spark.sql.catalyst.expressions.{Alias, And, ArrayTransform, CreateArray, CreateMap, CreateNamedStruct, CreateNamedStructUnsafe, CreateStruct, EqualTo, ExpectsInputTypes, Expression, GetStructField, LambdaFunction, NamedLambdaVariable, UnaryExpression}
+import org.apache.spark.sql.catalyst.expressions.codegen.{CodegenContext, ExprCode}
+import org.apache.spark.sql.catalyst.planning.ExtractEquiJoinKeys
+import org.apache.spark.sql.catalyst.plans.logical.{LogicalPlan, Subquery, Window}
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.types._
+
+/**
+ * We need to take care of special floating numbers (NaN and -0.0) in several places:
+ *   1. When compare values, different NaNs should be treated as same, `-0.0` and `0.0` should be
+ *      treated as same.
+ *   2. In aggregate grouping keys, different NaNs should belong to the same group, -0.0 and 0.0
+ *      should belong to the same group.
+ *   3. In join keys, different NaNs should be treated as same, `-0.0` and `0.0` should be
+ *      treated as same.
+ *   4. In window partition keys, different NaNs should belong to the same partition, -0.0 and 0.0
+ *      should belong to the same partition.
+ *
+ * Case 1 is fine, as we handle NaN and -0.0 well during comparison. For complex types, we
+ * recursively compare the fields/elements, so it's also fine.
+ *
+ * Case 2, 3 and 4 are problematic, as Spark SQL turns grouping/join/window partition keys into
+ * binary `UnsafeRow` and compare the binary data directly. Different NaNs have different binary
+ * representation, and the same thing happens for -0.0 and 0.0.
+ *
+ * This rule normalizes NaN and -0.0 in window partition keys, join keys and aggregate grouping
+ * keys.
+ *
+ * Ideally we should do the normalization in the physical operators that compare the
+ * binary `UnsafeRow` directly. We don't need this normalization if the Spark SQL execution engine
+ * is not optimized to run on binary data. This rule is created to simplify the implementation, so
+ * that we have a single place to do normalization, which is more maintainable.
+ *
+ * Note that, this rule must be executed at the end of optimizer, because the optimizer may create
+ * new joins(the subquery rewrite) and new join conditions(the join reorder).
+ */
+object NormalizeFloatingNumbers extends Rule[LogicalPlan] {
+
+  def apply(plan: LogicalPlan): LogicalPlan = plan match {
+    // A subquery will be rewritten into join later, and will go through this rule
+    // eventually. Here we skip subquery, as we only need to run this rule once.
+    case _: Subquery => plan
+
+    case _ => plan transform {
+      case w: Window if w.partitionSpec.exists(p => needNormalize(p.dataType)) =>
+        // Although the `windowExpressions` may refer to `partitionSpec` expressions, we don't need
+        // to normalize the `windowExpressions`, as they are executed per input row and should take
+        // the input row as it is.
+        w.copy(partitionSpec = w.partitionSpec.map(normalize))
+
+      // Only hash join and sort merge join need the normalization. Here we catch all Joins with
+      // join keys, assuming Joins with join keys are always planned as hash join or sort merge
+      // join. It's very unlikely that we will break this assumption in the near future.
+      case j @ ExtractEquiJoinKeys(_, leftKeys, rightKeys, condition, _, _, _)
+          // The analyzer guarantees left and right joins keys are of the same data type. Here we
+          // only need to check join keys of one side.
+          if leftKeys.exists(k => needNormalize(k.dataType)) =>
+        val newLeftJoinKeys = leftKeys.map(normalize)
+        val newRightJoinKeys = rightKeys.map(normalize)
+        val newConditions = newLeftJoinKeys.zip(newRightJoinKeys).map {
+          case (l, r) => EqualTo(l, r)
+        } ++ condition
+        j.copy(condition = Some(newConditions.reduce(And)))
+
+      // TODO: ideally Aggregate should also be handled here, but its grouping expressions are
+      // mixed in its aggregate expressions. It's unreliable to change the grouping expressions
+      // here. For now we normalize grouping expressions in `AggUtils` during planning.
+    }
+  }
+
+  private def needNormalize(dt: DataType): Boolean = dt match {
+    case FloatType | DoubleType => true
+    case StructType(fields) => fields.exists(f => needNormalize(f.dataType))
+    case ArrayType(et, _) => needNormalize(et)
+    // Currently MapType is not comparable and analyzer should fail earlier if this case happens.
+    case _: MapType =>
+      throw new IllegalStateException("grouping/join/window partition keys cannot be map type.")
+    case _ => false
+  }
+
+  private[sql] def normalize(expr: Expression): Expression = expr match {
+    case _ if expr.dataType == FloatType || expr.dataType == DoubleType =>
+      NormalizeNaNAndZero(expr)
+
+    case CreateNamedStruct(children) =>
+      CreateNamedStruct(children.map(normalize))
+
+    case CreateNamedStructUnsafe(children) =>
+      CreateNamedStructUnsafe(children.map(normalize))
+
+    case CreateArray(children) =>
+      CreateArray(children.map(normalize))
+
+    case CreateMap(children) =>
+      CreateMap(children.map(normalize))
+
+    case a: Alias if needNormalize(a.dataType) =>
+      a.withNewChildren(Seq(normalize(a.child)))
+
+    case _ if expr.dataType.isInstanceOf[StructType] && needNormalize(expr.dataType) =>
+      val fields = expr.dataType.asInstanceOf[StructType].fields.indices.map { i =>
+        normalize(GetStructField(expr, i))
+      }
+      CreateStruct(fields)
+
+    case _ if expr.dataType.isInstanceOf[ArrayType] && needNormalize(expr.dataType) =>
+      val ArrayType(et, containsNull) = expr.dataType
+      val lv = NamedLambdaVariable("arg", et, containsNull)
+      val function = normalize(lv)
+      ArrayTransform(expr, LambdaFunction(function, Seq(lv)))
+
+    case _ => expr
+  }
+}
+
+case class NormalizeNaNAndZero(child: Expression) extends UnaryExpression with ExpectsInputTypes {
+
+  override def dataType: DataType = child.dataType
+
+  override def inputTypes: Seq[AbstractDataType] = Seq(TypeCollection(FloatType, DoubleType))
+
+  private lazy val normalizer: Any => Any = child.dataType match {
+    case FloatType => (input: Any) => {
+      val f = input.asInstanceOf[Float]
+      if (f.isNaN) {
+        Float.NaN
+      } else if (f == -0.0f) {
+        0.0f
+      } else {
+        f
+      }
+    }
+
+    case DoubleType => (input: Any) => {
+      val d = input.asInstanceOf[Double]
+      if (d.isNaN) {
+        Double.NaN
+      } else if (d == -0.0d) {
+        0.0d
+      } else {
+        d
+      }
+    }
+  }
+
+  override def nullSafeEval(input: Any): Any = {
+    normalizer(input)
+  }
+
+  override protected def doGenCode(ctx: CodegenContext, ev: ExprCode): ExprCode = {
+    val codeToNormalize = child.dataType match {
+      case FloatType => (f: String) => {
+        s"""
+           |if (Float.isNaN($f)) {
+           |  ${ev.value} = Float.NaN;
+           |} else if ($f == -0.0f) {
+           |  ${ev.value} = 0.0f;
+           |} else {
+           |  ${ev.value} = $f;
+           |}
+         """.stripMargin
+      }
+
+      case DoubleType => (d: String) => {
+        s"""
+           |if (Double.isNaN($d)) {
+           |  ${ev.value} = Double.NaN;
+           |} else if ($d == -0.0d) {
+           |  ${ev.value} = 0.0d;
+           |} else {
+           |  ${ev.value} = $d;
+           |}
+         """.stripMargin
+      }
+    }
+
+    nullSafeCodeGen(ctx, ev, codeToNormalize)
+  }
+}

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

@@ -180,7 +180,9 @@ abstract class Optimizer(sessionCatalog: SessionCatalog)
       CollapseProject,
       RemoveNoopOperators) :+
     Batch("UpdateAttributeReferences", Once,
-      UpdateNullabilityInAttributeReferences)
+      UpdateNullabilityInAttributeReferences) :+
+    // This batch must be executed after the `RewriteSubquery` batch, which creates joins.
+    Batch("NormalizeFloatingNumbers", Once, NormalizeFloatingNumbers)
   }
 
   /**
@@ -210,7 +212,8 @@ abstract class Optimizer(sessionCatalog: SessionCatalog)
       PullupCorrelatedPredicates.ruleName ::
       RewriteCorrelatedScalarSubquery.ruleName ::
       RewritePredicateSubquery.ruleName ::
-      PullOutPythonUDFInJoinCondition.ruleName :: Nil
+      PullOutPythonUDFInJoinCondition.ruleName ::
+      NormalizeFloatingNumbers.ruleName :: Nil
 
   /**
    * Optimize all the subqueries inside expression.

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/planning/patterns.scala

@@ -105,8 +105,8 @@ object ExtractEquiJoinKeys extends Logging with PredicateHelper {
     (JoinType, Seq[Expression], Seq[Expression],
       Option[Expression], LogicalPlan, LogicalPlan, JoinHint)
 
-  def unapply(plan: LogicalPlan): Option[ReturnType] = plan match {
-    case join @ Join(left, right, joinType, condition, hint) =>
+  def unapply(join: Join): Option[ReturnType] = join match {
+    case Join(left, right, joinType, condition, hint) =>
       logDebug(s"Considering join on: $condition")
       // Find equi-join predicates that can be evaluated before the join, and thus can be used
       // as join keys.
@@ -140,7 +140,6 @@ object ExtractEquiJoinKeys extends Logging with PredicateHelper {
       } else {
         None
       }
-    case _ => None
   }
 }
 

sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/expressions/UnsafeRowConverterSuite.scala

@@ -246,22 +246,6 @@ class UnsafeRowConverterSuite extends SparkFunSuite with Matchers with PlanTestB
     // assert(setToNullAfterCreation.get(11) === rowWithNoNullColumns.get(11))
   }
 
-  testBothCodegenAndInterpreted("NaN canonicalization") {
-    val factory = UnsafeProjection
-    val fieldTypes: Array[DataType] = Array(FloatType, DoubleType)
-
-    val row1 = new SpecificInternalRow(fieldTypes)
-    row1.setFloat(0, java.lang.Float.intBitsToFloat(0x7f800001))
-    row1.setDouble(1, java.lang.Double.longBitsToDouble(0x7ff0000000000001L))
-
-    val row2 = new SpecificInternalRow(fieldTypes)
-    row2.setFloat(0, java.lang.Float.intBitsToFloat(0x7fffffff))
-    row2.setDouble(1, java.lang.Double.longBitsToDouble(0x7fffffffffffffffL))
-
-    val converter = factory.create(fieldTypes)
-    assert(converter.apply(row1).getBytes === converter.apply(row2).getBytes)
-  }
-
   testBothCodegenAndInterpreted("basic conversion with struct type") {
     val factory = UnsafeProjection
     val fieldTypes: Array[DataType] = Array(

sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/expressions/codegen/UnsafeRowWriterSuite.scala

@@ -49,25 +49,4 @@ class UnsafeRowWriterSuite extends SparkFunSuite {
     // The two rows should be the equal
     assert(res1 == res2)
   }
-
-  test("SPARK-26021: normalize float/double NaN and -0.0") {
-    val unsafeRowWriter1 = new UnsafeRowWriter(4)
-    unsafeRowWriter1.resetRowWriter()
-    unsafeRowWriter1.write(0, Float.NaN)
-    unsafeRowWriter1.write(1, Double.NaN)
-    unsafeRowWriter1.write(2, 0.0f)
-    unsafeRowWriter1.write(3, 0.0)
-    val res1 = unsafeRowWriter1.getRow
-
-    val unsafeRowWriter2 = new UnsafeRowWriter(4)
-    unsafeRowWriter2.resetRowWriter()
-    unsafeRowWriter2.write(0, 0.0f/0.0f)
-    unsafeRowWriter2.write(1, 0.0/0.0)
-    unsafeRowWriter2.write(2, -0.0f)
-    unsafeRowWriter2.write(3, -0.0)
-    val res2 = unsafeRowWriter2.getRow
-
-    // The two rows should be the equal
-    assert(res1 == res2)
-  }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/AggUtils.scala

@@ -19,9 +19,9 @@ package org.apache.spark.sql.execution.aggregate
 
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.catalyst.expressions.aggregate._
+import org.apache.spark.sql.catalyst.optimizer.NormalizeFloatingNumbers
 import org.apache.spark.sql.execution.SparkPlan
 import org.apache.spark.sql.execution.streaming.{StateStoreRestoreExec, StateStoreSaveExec}
-import org.apache.spark.sql.internal.SQLConf
 
 /**
  * Utility functions used by the query planner to convert our plan to new aggregation code path.
@@ -35,12 +35,20 @@ object AggUtils {
       initialInputBufferOffset: Int = 0,
       resultExpressions: Seq[NamedExpression] = Nil,
       child: SparkPlan): SparkPlan = {
+    // Ideally this should be done in `NormalizeFloatingNumbers`, but we do it here because
+    // `groupingExpressions` is not extracted during logical phase.
+    val normalizedGroupingExpressions = groupingExpressions.map { e =>
+      NormalizeFloatingNumbers.normalize(e) match {
+        case n: NamedExpression => n
+        case other => Alias(other, e.name)(exprId = e.exprId)
+      }
+    }
     val useHash = HashAggregateExec.supportsAggregate(
       aggregateExpressions.flatMap(_.aggregateFunction.aggBufferAttributes))
     if (useHash) {
       HashAggregateExec(
         requiredChildDistributionExpressions = requiredChildDistributionExpressions,
-        groupingExpressions = groupingExpressions,
+        groupingExpressions = normalizedGroupingExpressions,
         aggregateExpressions = aggregateExpressions,
         aggregateAttributes = aggregateAttributes,
         initialInputBufferOffset = initialInputBufferOffset,
@@ -53,7 +61,7 @@ object AggUtils {
       if (objectHashEnabled && useObjectHash) {
         ObjectHashAggregateExec(
           requiredChildDistributionExpressions = requiredChildDistributionExpressions,
-          groupingExpressions = groupingExpressions,
+          groupingExpressions = normalizedGroupingExpressions,
           aggregateExpressions = aggregateExpressions,
           aggregateAttributes = aggregateAttributes,
           initialInputBufferOffset = initialInputBufferOffset,
@@ -62,7 +70,7 @@ object AggUtils {
       } else {
         SortAggregateExec(
           requiredChildDistributionExpressions = requiredChildDistributionExpressions,
-          groupingExpressions = groupingExpressions,
+          groupingExpressions = normalizedGroupingExpressions,
           aggregateExpressions = aggregateExpressions,
           aggregateAttributes = aggregateAttributes,
           initialInputBufferOffset = initialInputBufferOffset,