ericl
diff --git a/‎sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala‎
Lines changed: 149 additions & 50 deletions b/‎sql/hive/src/main/scala/org/apache/spark/sql/hive/hiveUDFs.scala‎
Lines changed: 149 additions & 50 deletions
@@ -17,16 +17,18 @@
 
 package org.apache.spark.sql.hive
 
+import java.nio.ByteBuffer
+
 import scala.collection.JavaConverters._
 import scala.collection.mutable.ArrayBuffer
 
 import org.apache.hadoop.hive.ql.exec._
 import org.apache.hadoop.hive.ql.udf.{UDFType => HiveUDFType}
 import org.apache.hadoop.hive.ql.udf.generic._
+import org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator.AggregationBuffer
 import org.apache.hadoop.hive.ql.udf.generic.GenericUDF._
 import org.apache.hadoop.hive.ql.udf.generic.GenericUDFUtils.ConversionHelper
-import org.apache.hadoop.hive.serde2.objectinspector.{ConstantObjectInspector, ObjectInspector,
-  ObjectInspectorFactory}
+import org.apache.hadoop.hive.serde2.objectinspector.{ConstantObjectInspector, ObjectInspector, ObjectInspectorFactory}
 import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorFactory.ObjectInspectorOptions
 
 import org.apache.spark.internal.Logging
@@ -58,7 +60,7 @@ case class HiveSimpleUDF(
 
   @transient
   private lazy val isUDFDeterministic = {
-    val udfType = function.getClass().getAnnotation(classOf[HiveUDFType])
+    val udfType = function.getClass.getAnnotation(classOf[HiveUDFType])
     udfType != null && udfType.deterministic() && !udfType.stateful()
   }
 
@@ -75,7 +77,7 @@ case class HiveSimpleUDF(
 
   @transient
   lazy val unwrapper = unwrapperFor(ObjectInspectorFactory.getReflectionObjectInspector(
-    method.getGenericReturnType(), ObjectInspectorOptions.JAVA))
+    method.getGenericReturnType, ObjectInspectorOptions.JAVA))
 
   @transient
   private lazy val cached: Array[AnyRef] = new Array[AnyRef](children.length)
@@ -263,8 +265,35 @@ case class HiveGenericUDTF(
 }
 
 /**
- * Currently we don't support partial aggregation for queries using Hive UDAF, which may hurt
- * performance a lot.
+ * While being evaluated by Spark SQL, the aggregation state of a Hive UDAF may be in the following
+ * three formats:
+ *
+ *  1. An instance of some concrete `GenericUDAFEvaluator.AggregationBuffer` class
+ *
+ *     This is the native Hive representation of an aggregation state. Hive `GenericUDAFEvaluator`
+ *     methods like `iterate()`, `merge()`, `terminatePartial()`, and `terminate()` use this format.
+ *     We call these methods to evaluate Hive UDAFs.
+ *
+ *  2. A Java object that can be inspected using the `ObjectInspector` returned by the
+ *     `GenericUDAFEvaluator.init()` method.
+ *
+ *     Hive uses this format to produce a serializable aggregation state so that it can shuffle
+ *     partial aggregation results. Whenever we need to convert a Hive `AggregationBuffer` instance
+ *     into a Spark SQL value, we have to convert it to this format first and then do the conversion
+ *     with the help of `ObjectInspector`s.
+ *
+ *  3. A Spark SQL value
+ *
+ *     We use this format for serializing Hive UDAF aggregation states on Spark side. To be more
+ *     specific, we convert `AggregationBuffer`s into equivalent Spark SQL values, write them into
+ *     `UnsafeRow`s, and then retrieve the byte array behind those `UnsafeRow`s as serialization
+ *     results.
+ *
+ * We may use the following methods to convert the aggregation state back and forth:
+ *
+ *  - `wrap()`/`wrapperFor()`: from 3 to 1
+ *  - `unwrap()`/`unwrapperFor()`: from 1 to 3
+ *  - `GenericUDAFEvaluator.terminatePartial()`: from 2 to 3
  */
 case class HiveUDAFFunction(
     name: String,
@@ -273,89 +302,89 @@ case class HiveUDAFFunction(
     isUDAFBridgeRequired: Boolean = false,
     mutableAggBufferOffset: Int = 0,
     inputAggBufferOffset: Int = 0)
-  extends ImperativeAggregate with HiveInspectors {
+  extends TypedImperativeAggregate[GenericUDAFEvaluator.AggregationBuffer] with HiveInspectors {
 
   override def withNewMutableAggBufferOffset(newMutableAggBufferOffset: Int): ImperativeAggregate =
     copy(mutableAggBufferOffset = newMutableAggBufferOffset)
 
   override def withNewInputAggBufferOffset(newInputAggBufferOffset: Int): ImperativeAggregate =
     copy(inputAggBufferOffset = newInputAggBufferOffset)
 
+  // Hive `ObjectInspector`s for all child expressions (input parameters of the function).
   @transient
-  private lazy val resolver =
-    if (isUDAFBridgeRequired) {
+  private lazy val inputInspectors = children.map(toInspector).toArray
+
+  // Spark SQL data types of input parameters.
+  @transient
+  private lazy val inputDataTypes: Array[DataType] = children.map(_.dataType).toArray
+
+  private def newEvaluator(): GenericUDAFEvaluator = {
+    val resolver = if (isUDAFBridgeRequired) {
       new GenericUDAFBridge(funcWrapper.createFunction[UDAF]())
     } else {
       funcWrapper.createFunction[AbstractGenericUDAFResolver]()
     }
 
-  @transient
-  private lazy val inspectors = children.map(toInspector).toArray
-
-  @transient
-  private lazy val functionAndInspector = {
-    val parameterInfo = new SimpleGenericUDAFParameterInfo(inspectors, false, false)
-    val f = resolver.getEvaluator(parameterInfo)
-    f -> f.init(GenericUDAFEvaluator.Mode.COMPLETE, inspectors)
+    val parameterInfo = new SimpleGenericUDAFParameterInfo(inputInspectors, false, false)
+    resolver.getEvaluator(parameterInfo)
   }
 
+  // The UDAF evaluator used to consume raw input rows and produce partial aggregation results.
   @transient
-  private lazy val function = functionAndInspector._1
+  private lazy val partial1ModeEvaluator = newEvaluator()
 
+  // Hive `ObjectInspector` used to inspect partial aggregation results.
   @transient
-  private lazy val wrappers = children.map(x => wrapperFor(toInspector(x), x.dataType)).toArray
+  private val partialResultInspector = partial1ModeEvaluator.init(
+    GenericUDAFEvaluator.Mode.PARTIAL1,
+    inputInspectors
+  )
 
+  // The UDAF evaluator used to merge partial aggregation results.
   @transient
-  private lazy val returnInspector = functionAndInspector._2
+  private lazy val partial2ModeEvaluator = {
+    val evaluator = newEvaluator()
+    evaluator.init(GenericUDAFEvaluator.Mode.PARTIAL2, Array(partialResultInspector))
+    evaluator
+  }
 
+  // Spark SQL data type of partial aggregation results
   @transient
-  private lazy val unwrapper = unwrapperFor(returnInspector)
+  private lazy val partialResultDataType = inspectorToDataType(partialResultInspector)
 
+  // The UDAF evaluator used to compute the final result from a partial aggregation result objects.
   @transient
-  private[this] var buffer: GenericUDAFEvaluator.AggregationBuffer = _
-
-  override def eval(input: InternalRow): Any = unwrapper(function.evaluate(buffer))
+  private lazy val finalModeEvaluator = newEvaluator()
 
+  // Hive `ObjectInspector` used to inspect the final aggregation result object.
   @transient
-  private lazy val inputProjection = new InterpretedProjection(children)
+  private val returnInspector = finalModeEvaluator.init(
+    GenericUDAFEvaluator.Mode.FINAL,
+    Array(partialResultInspector)
+  )
 
+  // Wrapper functions used to wrap Spark SQL input arguments into Hive specific format.
   @transient
-  private lazy val cached = new Array[AnyRef](children.length)
+  private lazy val inputWrappers = children.map(x => wrapperFor(toInspector(x), x.dataType)).toArray
 
+  // Unwrapper function used to unwrap final aggregation result objects returned by Hive UDAFs into
+  // Spark SQL specific format.
   @transient
-  private lazy val inputDataTypes: Array[DataType] = children.map(_.dataType).toArray
-
-  // Hive UDAF has its own buffer, so we don't need to occupy a slot in the aggregation
-  // buffer for it.
-  override def aggBufferSchema: StructType = StructType(Nil)
-
-  override def update(_buffer: InternalRow, input: InternalRow): Unit = {
-    val inputs = inputProjection(input)
-    function.iterate(buffer, wrap(inputs, wrappers, cached, inputDataTypes))
-  }
-
-  override def merge(buffer1: InternalRow, buffer2: InternalRow): Unit = {
-    throw new UnsupportedOperationException(
-      "Hive UDAF doesn't support partial aggregate")
-  }
+  private lazy val resultUnwrapper = unwrapperFor(returnInspector)
 
-  override def initialize(_buffer: InternalRow): Unit = {
-    buffer = function.getNewAggregationBuffer
-  }
-
-  override val aggBufferAttributes: Seq[AttributeReference] = Nil
+  @transient
+  private lazy val cached: Array[AnyRef] = new Array[AnyRef](children.length)
 
-  // Note: although this simply copies aggBufferAttributes, this common code can not be placed
-  // in the superclass because that will lead to initialization ordering issues.
-  override val inputAggBufferAttributes: Seq[AttributeReference] = Nil
+  @transient
+  private lazy val aggBufferSerDe: AggregationBufferSerDe = new AggregationBufferSerDe
 
   // We rely on Hive to check the input data types, so use `AnyDataType` here to bypass our
   // catalyst type checking framework.
   override def inputTypes: Seq[AbstractDataType] = children.map(_ => AnyDataType)
 
   override def nullable: Boolean = true
 
-  override def supportsPartial: Boolean = false
+  override def supportsPartial: Boolean = true
 
   override lazy val dataType: DataType = inspectorToDataType(returnInspector)
 
@@ -365,4 +394,74 @@ case class HiveUDAFFunction(
     val distinct = if (isDistinct) "DISTINCT " else " "
     s"$name($distinct${children.map(_.sql).mkString(", ")})"
   }
+
+  override def createAggregationBuffer(): AggregationBuffer =
+    partial1ModeEvaluator.getNewAggregationBuffer
+
+  @transient
+  private lazy val inputProjection = UnsafeProjection.create(children)
+
+  override def update(buffer: AggregationBuffer, input: InternalRow): Unit = {
+    partial1ModeEvaluator.iterate(
+      buffer, wrap(inputProjection(input), inputWrappers, cached, inputDataTypes))
+  }
+
+  override def merge(buffer: AggregationBuffer, input: AggregationBuffer): Unit = {
+    // The 2nd argument of the Hive `GenericUDAFEvaluator.merge()` method is an input aggregation
+    // buffer in the 3rd format mentioned in the ScalaDoc of this class. Originally, Hive converts
+    // this `AggregationBuffer`s into this format before shuffling partial aggregation results, and
+    // calls `GenericUDAFEvaluator.terminatePartial()` to do the conversion.
+    partial2ModeEvaluator.merge(buffer, partial1ModeEvaluator.terminatePartial(input))
+  }
+
+  override def eval(buffer: AggregationBuffer): Any = {
+    resultUnwrapper(finalModeEvaluator.terminate(buffer))
+  }
+
+  override def serialize(buffer: AggregationBuffer): Array[Byte] = {
+    // Serializes an `AggregationBuffer` that holds partial aggregation results so that we can
+    // shuffle it for global aggregation later.
+    aggBufferSerDe.serialize(buffer)
+  }
+
+  override def deserialize(bytes: Array[Byte]): AggregationBuffer = {
+    // Deserializes an `AggregationBuffer` from the shuffled partial aggregation phase to prepare
+    // for global aggregation by merging multiple partial aggregation results within a single group.
+    aggBufferSerDe.deserialize(bytes)
+  }
+
+  // Helper class used to de/serialize Hive UDAF `AggregationBuffer` objects
+  private class AggregationBufferSerDe {
+    private val partialResultUnwrapper = unwrapperFor(partialResultInspector)
+
+    private val partialResultWrapper = wrapperFor(partialResultInspector, partialResultDataType)
+
+    private val projection = UnsafeProjection.create(Array(partialResultDataType))
+
+    private val mutableRow = new GenericInternalRow(1)
+
+    def serialize(buffer: AggregationBuffer): Array[Byte] = {
+      // `GenericUDAFEvaluator.terminatePartial()` converts an `AggregationBuffer` into an object
+      // that can be inspected by the `ObjectInspector` returned by `GenericUDAFEvaluator.init()`.
+      // Then we can unwrap it to a Spark SQL value.
+      mutableRow.update(0, partialResultUnwrapper(partial1ModeEvaluator.terminatePartial(buffer)))
+      val unsafeRow = projection(mutableRow)
+      val bytes = ByteBuffer.allocate(unsafeRow.getSizeInBytes)
+      unsafeRow.writeTo(bytes)
+      bytes.array()
+    }
+
+    def deserialize(bytes: Array[Byte]): AggregationBuffer = {
+      // `GenericUDAFEvaluator` doesn't provide any method that is capable to convert an object
+      // returned by `GenericUDAFEvaluator.terminatePartial()` back to an `AggregationBuffer`. The
+      // workaround here is creating an initial `AggregationBuffer` first and then merge the
+      // deserialized object into the buffer.
+      val buffer = partial2ModeEvaluator.getNewAggregationBuffer
+      val unsafeRow = new UnsafeRow(1)
+      unsafeRow.pointTo(bytes, bytes.length)
+      val partialResult = unsafeRow.get(0, partialResultDataType)
+      partial2ModeEvaluator.merge(buffer, partialResultWrapper(partialResult))
+      buffer
+    }
+  }
 }