spark udaf 源码

spark udaf 代码

文件路径：/sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/udaf.scala

/*
 * 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.execution.aggregate

import org.apache.spark.internal.Logging
import org.apache.spark.sql.Row
import org.apache.spark.sql.catalyst.{CatalystTypeConverters, InternalRow}
import org.apache.spark.sql.catalyst.encoders.ExpressionEncoder
import org.apache.spark.sql.catalyst.expressions.{AttributeReference, Expression, _}
import org.apache.spark.sql.catalyst.expressions.aggregate.{ImperativeAggregate, TypedImperativeAggregate}
import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
import org.apache.spark.sql.catalyst.rules.Rule
import org.apache.spark.sql.expressions.{Aggregator, MutableAggregationBuffer, UserDefinedAggregateFunction}
import org.apache.spark.sql.types._

/**
 * A helper trait used to create specialized setter and getter for types supported by
 * [[org.apache.spark.sql.execution.UnsafeFixedWidthAggregationMap]]'s buffer.
 * (see UnsafeFixedWidthAggregationMap.supportsAggregationBufferSchema).
 */
sealed trait BufferSetterGetterUtils {

  def createGetters(schema: StructType): Array[(InternalRow, Int) => Any] = {
    val dataTypes = schema.fields.map(_.dataType)
    val getters = new Array[(InternalRow, Int) => Any](dataTypes.length)

    var i = 0
    while (i < getters.length) {
      getters(i) = dataTypes(i) match {
        case NullType =>
          (row: InternalRow, ordinal: Int) => null

        case BooleanType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getBoolean(ordinal)

        case ByteType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getByte(ordinal)

        case ShortType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getShort(ordinal)

        case IntegerType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getInt(ordinal)

        case LongType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getLong(ordinal)

        case FloatType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getFloat(ordinal)

        case DoubleType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getDouble(ordinal)

        case dt: DecimalType =>
          val precision = dt.precision
          val scale = dt.scale
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getDecimal(ordinal, precision, scale)

        case DateType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getInt(ordinal)

        case TimestampType | TimestampNTZType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getLong(ordinal)

        case _: YearMonthIntervalType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getInt(ordinal)

        case _: DayTimeIntervalType =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.getLong(ordinal)

        case other =>
          (row: InternalRow, ordinal: Int) =>
            if (row.isNullAt(ordinal)) null else row.get(ordinal, other)
      }

      i += 1
    }

    getters
  }

  def createSetters(schema: StructType): Array[((InternalRow, Int, Any) => Unit)] = {
    val dataTypes = schema.fields.map(_.dataType)
    val setters = new Array[(InternalRow, Int, Any) => Unit](dataTypes.length)

    var i = 0
    while (i < setters.length) {
      setters(i) = dataTypes(i) match {
        case NullType =>
          (row: InternalRow, ordinal: Int, value: Any) => row.setNullAt(ordinal)

        case b: BooleanType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setBoolean(ordinal, value.asInstanceOf[Boolean])
            } else {
              row.setNullAt(ordinal)
            }

        case ByteType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setByte(ordinal, value.asInstanceOf[Byte])
            } else {
              row.setNullAt(ordinal)
            }

        case ShortType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setShort(ordinal, value.asInstanceOf[Short])
            } else {
              row.setNullAt(ordinal)
            }

        case IntegerType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setInt(ordinal, value.asInstanceOf[Int])
            } else {
              row.setNullAt(ordinal)
            }

        case LongType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setLong(ordinal, value.asInstanceOf[Long])
            } else {
              row.setNullAt(ordinal)
            }

        case FloatType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setFloat(ordinal, value.asInstanceOf[Float])
            } else {
              row.setNullAt(ordinal)
            }

        case DoubleType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setDouble(ordinal, value.asInstanceOf[Double])
            } else {
              row.setNullAt(ordinal)
            }

        case dt: DecimalType =>
          val precision = dt.precision
          (row: InternalRow, ordinal: Int, value: Any) =>
            // To make it work with UnsafeRow, we cannot use setNullAt.
            // Please see the comment of UnsafeRow's setDecimal.
            row.setDecimal(ordinal, value.asInstanceOf[Decimal], precision)

        case DateType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setInt(ordinal, value.asInstanceOf[Int])
            } else {
              row.setNullAt(ordinal)
            }

        case TimestampType | TimestampNTZType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setLong(ordinal, value.asInstanceOf[Long])
            } else {
              row.setNullAt(ordinal)
            }

        case _: YearMonthIntervalType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setInt(ordinal, value.asInstanceOf[Int])
            } else {
              row.setNullAt(ordinal)
            }

        case _: DayTimeIntervalType =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.setLong(ordinal, value.asInstanceOf[Long])
            } else {
              row.setNullAt(ordinal)
            }

        case other =>
          (row: InternalRow, ordinal: Int, value: Any) =>
            if (value != null) {
              row.update(ordinal, value)
            } else {
              row.setNullAt(ordinal)
            }
      }

      i += 1
    }

    setters
  }
}

/**
 * A Mutable [[Row]] representing a mutable aggregation buffer.
 */
private[aggregate] class MutableAggregationBufferImpl(
    schema: StructType,
    toCatalystConverters: Array[Any => Any],
    toScalaConverters: Array[Any => Any],
    bufferOffset: Int,
    var underlyingBuffer: InternalRow)
  extends MutableAggregationBuffer with BufferSetterGetterUtils {

  private[this] val offsets: Array[Int] = {
    val newOffsets = new Array[Int](length)
    var i = 0
    while (i < newOffsets.length) {
      newOffsets(i) = bufferOffset + i
      i += 1
    }
    newOffsets
  }

  private[this] val bufferValueGetters = createGetters(schema)

  private[this] val bufferValueSetters = createSetters(schema)

  override def length: Int = toCatalystConverters.length

  override def get(i: Int): Any = {
    if (i >= length || i < 0) {
      throw new IllegalArgumentException(
        s"Could not access ${i}th value in this buffer because it only has $length values.")
    }

    toScalaConverters(i)(bufferValueGetters(i)(underlyingBuffer, offsets(i)))
  }

  def update(i: Int, value: Any): Unit = {
    if (i >= length || i < 0) {
      throw new IllegalArgumentException(
        s"Could not update ${i}th value in this buffer because it only has $length values.")
    }

    bufferValueSetters(i)(underlyingBuffer, offsets(i), toCatalystConverters(i)(value))
  }

  // Because get method call specialized getter based on the schema, we cannot use the
  // default implementation of the isNullAt (which is get(i) == null).
  // We have to override it to call isNullAt of the underlyingBuffer.
  override def isNullAt(i: Int): Boolean = {
    underlyingBuffer.isNullAt(offsets(i))
  }

  override def copy(): MutableAggregationBufferImpl = {
    new MutableAggregationBufferImpl(
      schema,
      toCatalystConverters,
      toScalaConverters,
      bufferOffset,
      underlyingBuffer)
  }
}

/**
 * A [[Row]] representing an immutable aggregation buffer.
 */
private[aggregate] class InputAggregationBuffer(
    schema: StructType,
    toCatalystConverters: Array[Any => Any],
    toScalaConverters: Array[Any => Any],
    bufferOffset: Int,
    var underlyingInputBuffer: InternalRow)
  extends Row with BufferSetterGetterUtils {

  private[this] val offsets: Array[Int] = {
    val newOffsets = new Array[Int](length)
    var i = 0
    while (i < newOffsets.length) {
      newOffsets(i) = bufferOffset + i
      i += 1
    }
    newOffsets
  }

  private[this] val bufferValueGetters = createGetters(schema)

  def getBufferOffset: Int = bufferOffset

  override def length: Int = toCatalystConverters.length

  override def get(i: Int): Any = {
    if (i >= length || i < 0) {
      throw new IllegalArgumentException(
        s"Could not access ${i}th value in this buffer because it only has $length values.")
    }
    toScalaConverters(i)(bufferValueGetters(i)(underlyingInputBuffer, offsets(i)))
  }

  // Because get method call specialized getter based on the schema, we cannot use the
  // default implementation of the isNullAt (which is get(i) == null).
  // We have to override it to call isNullAt of the underlyingInputBuffer.
  override def isNullAt(i: Int): Boolean = {
    underlyingInputBuffer.isNullAt(offsets(i))
  }

  override def copy(): InputAggregationBuffer = {
    new InputAggregationBuffer(
      schema,
      toCatalystConverters,
      toScalaConverters,
      bufferOffset,
      underlyingInputBuffer)
  }
}

/**
 * The internal wrapper used to hook a [[UserDefinedAggregateFunction]] `udaf` in the
 * internal aggregation code path.
 */
case class ScalaUDAF(
    children: Seq[Expression],
    udaf: UserDefinedAggregateFunction,
    mutableAggBufferOffset: Int = 0,
    inputAggBufferOffset: Int = 0,
    udafName: Option[String] = None)
  extends ImperativeAggregate
  with NonSQLExpression
  with Logging
  with ImplicitCastInputTypes
  with UserDefinedExpression {

  override def withNewMutableAggBufferOffset(newMutableAggBufferOffset: Int): ImperativeAggregate =
    copy(mutableAggBufferOffset = newMutableAggBufferOffset)

  override def withNewInputAggBufferOffset(newInputAggBufferOffset: Int): ImperativeAggregate =
    copy(inputAggBufferOffset = newInputAggBufferOffset)

  override def nullable: Boolean = true

  override def dataType: DataType = udaf.dataType

  override lazy val deterministic: Boolean = udaf.deterministic

  override val inputTypes: Seq[DataType] = udaf.inputSchema.map(_.dataType)

  override val aggBufferSchema: StructType = udaf.bufferSchema

  override val aggBufferAttributes: Seq[AttributeReference] = aggBufferSchema.toAttributes

  // 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] =
    aggBufferAttributes.map(_.newInstance())

  private[this] lazy val childrenSchema: StructType = {
    val inputFields = children.zipWithIndex.map {
      case (child, index) =>
        StructField(s"input$index", child.dataType, child.nullable, Metadata.empty)
    }
    StructType(inputFields)
  }

  private lazy val inputProjection = {
    val inputAttributes = childrenSchema.toAttributes
    log.debug(
      s"Creating MutableProj: $children, inputSchema: $inputAttributes.")
    MutableProjection.create(children, inputAttributes)
  }

  private[this] lazy val inputToScalaConverters: Any => Any =
    CatalystTypeConverters.createToScalaConverter(childrenSchema)

  private[this] lazy val bufferValuesToCatalystConverters: Array[Any => Any] = {
    aggBufferSchema.fields.map { field =>
      CatalystTypeConverters.createToCatalystConverter(field.dataType)
    }
  }

  private[this] lazy val bufferValuesToScalaConverters: Array[Any => Any] = {
    aggBufferSchema.fields.map { field =>
      CatalystTypeConverters.createToScalaConverter(field.dataType)
    }
  }

  private[this] lazy val outputToCatalystConverter: Any => Any = {
    CatalystTypeConverters.createToCatalystConverter(dataType)
  }

  // This buffer is only used at executor side.
  private[this] lazy val inputAggregateBuffer: InputAggregationBuffer = {
    new InputAggregationBuffer(
      aggBufferSchema,
      bufferValuesToCatalystConverters,
      bufferValuesToScalaConverters,
      inputAggBufferOffset,
      null)
  }

  // This buffer is only used at executor side.
  private[this] lazy val mutableAggregateBuffer: MutableAggregationBufferImpl = {
    new MutableAggregationBufferImpl(
      aggBufferSchema,
      bufferValuesToCatalystConverters,
      bufferValuesToScalaConverters,
      mutableAggBufferOffset,
      null)
  }

  // This buffer is only used at executor side.
  private[this] lazy val evalAggregateBuffer: InputAggregationBuffer = {
    new InputAggregationBuffer(
      aggBufferSchema,
      bufferValuesToCatalystConverters,
      bufferValuesToScalaConverters,
      mutableAggBufferOffset,
      null)
  }

  override def initialize(buffer: InternalRow): Unit = {
    mutableAggregateBuffer.underlyingBuffer = buffer

    udaf.initialize(mutableAggregateBuffer)
  }

  override def update(buffer: InternalRow, input: InternalRow): Unit = {
    mutableAggregateBuffer.underlyingBuffer = buffer

    udaf.update(
      mutableAggregateBuffer,
      inputToScalaConverters(inputProjection(input)).asInstanceOf[Row])
  }

  override def merge(buffer1: InternalRow, buffer2: InternalRow): Unit = {
    mutableAggregateBuffer.underlyingBuffer = buffer1
    inputAggregateBuffer.underlyingInputBuffer = buffer2

    udaf.merge(mutableAggregateBuffer, inputAggregateBuffer)
  }

  override def eval(buffer: InternalRow): Any = {
    evalAggregateBuffer.underlyingInputBuffer = buffer

    outputToCatalystConverter(udaf.evaluate(evalAggregateBuffer))
  }

  override def toString: String = {
    s"""$nodeName(${children.mkString(",")})"""
  }

  override def nodeName: String = name

  override def name: String = udafName.getOrElse(udaf.getClass.getSimpleName)

  override protected def withNewChildrenInternal(newChildren: IndexedSeq[Expression]): ScalaUDAF =
    copy(children = newChildren)
}

case class ScalaAggregator[IN, BUF, OUT](
    children: Seq[Expression],
    agg: Aggregator[IN, BUF, OUT],
    inputEncoder: ExpressionEncoder[IN],
    bufferEncoder: ExpressionEncoder[BUF],
    nullable: Boolean = true,
    isDeterministic: Boolean = true,
    mutableAggBufferOffset: Int = 0,
    inputAggBufferOffset: Int = 0,
    aggregatorName: Option[String] = None)
  extends TypedImperativeAggregate[BUF]
  with NonSQLExpression
  with UserDefinedExpression
  with ImplicitCastInputTypes
  with Logging {

  // input and buffer encoders are resolved by ResolveEncodersInScalaAgg
  private[this] lazy val inputDeserializer = inputEncoder.createDeserializer()
  private[this] lazy val bufferSerializer = bufferEncoder.createSerializer()
  private[this] lazy val bufferDeserializer = bufferEncoder.createDeserializer()
  private[this] lazy val outputEncoder = agg.outputEncoder.asInstanceOf[ExpressionEncoder[OUT]]
  private[this] lazy val outputSerializer = outputEncoder.createSerializer()

  def dataType: DataType = outputEncoder.objSerializer.dataType

  def inputTypes: Seq[DataType] = inputEncoder.schema.map(_.dataType)

  override lazy val deterministic: Boolean = isDeterministic

  def withNewMutableAggBufferOffset(newMutableAggBufferOffset: Int): ScalaAggregator[IN, BUF, OUT] =
    copy(mutableAggBufferOffset = newMutableAggBufferOffset)

  def withNewInputAggBufferOffset(newInputAggBufferOffset: Int): ScalaAggregator[IN, BUF, OUT] =
    copy(inputAggBufferOffset = newInputAggBufferOffset)

  private[this] lazy val inputProjection = UnsafeProjection.create(children)

  def createAggregationBuffer(): BUF = agg.zero

  def update(buffer: BUF, input: InternalRow): BUF =
    agg.reduce(buffer, inputDeserializer(inputProjection(input)))

  def merge(buffer: BUF, input: BUF): BUF = agg.merge(buffer, input)

  def eval(buffer: BUF): Any = {
    val row = outputSerializer(agg.finish(buffer))
    if (outputEncoder.isSerializedAsStruct) row else row.get(0, dataType)
  }

  private[this] lazy val bufferRow = new UnsafeRow(bufferEncoder.namedExpressions.length)

  def serialize(agg: BUF): Array[Byte] =
    bufferSerializer(agg).asInstanceOf[UnsafeRow].getBytes()

  def deserialize(storageFormat: Array[Byte]): BUF = {
    bufferRow.pointTo(storageFormat, storageFormat.length)
    bufferDeserializer(bufferRow)
  }

  override def toString: String = s"""${nodeName}(${children.mkString(",")})"""

  override def nodeName: String = name

  override def name: String = aggregatorName.getOrElse(agg.getClass.getSimpleName)

  override protected def withNewChildrenInternal(
      newChildren: IndexedSeq[Expression]): ScalaAggregator[IN, BUF, OUT] =
    copy(children = newChildren)
}

/**
 * An extension rule to resolve encoder expressions from a [[ScalaAggregator]]
 */
object ResolveEncodersInScalaAgg extends Rule[LogicalPlan] {
  override def apply(plan: LogicalPlan): LogicalPlan = plan.resolveOperatorsUp {
    case p if !p.resolved => p
    case p => p.transformExpressionsUp {
      case agg: ScalaAggregator[_, _, _] =>
        agg.copy(
          inputEncoder = agg.inputEncoder.resolveAndBind(),
          bufferEncoder = agg.bufferEncoder.resolveAndBind())
    }
  }
}

相关信息

spark 源码目录

相关文章

spark AggUtils 源码

spark AggregateCodegenSupport 源码

spark AggregationIterator 源码

spark BaseAggregateExec 源码

spark HashAggregateExec 源码

spark HashMapGenerator 源码

spark MergingSessionsExec 源码

spark MergingSessionsIterator 源码

spark ObjectAggregationIterator 源码

spark ObjectAggregationMap 源码