path: root/sql/core/src/main/scala/org/apache/spark/sql/execution/window/WindowExec.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.window

import scala.collection.mutable
import scala.collection.mutable.ArrayBuffer

import org.apache.spark.{SparkEnv, TaskContext}
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.catalyst.InternalRow
import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.catalyst.expressions.aggregate._
import org.apache.spark.sql.catalyst.plans.physical._
import org.apache.spark.sql.execution.{SparkPlan, UnaryExecNode}
import org.apache.spark.sql.types.IntegerType
import org.apache.spark.util.collection.unsafe.sort.UnsafeExternalSorter

/**
 * This class calculates and outputs (windowed) aggregates over the rows in a single (sorted)
 * partition. The aggregates are calculated for each row in the group. Special processing
 * instructions, frames, are used to calculate these aggregates. Frames are processed in the order
 * specified in the window specification (the ORDER BY ... clause). There are four different frame
 * types:
 * - Entire partition: The frame is the entire partition, i.e.
 *   UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING. For this case, window function will take all
 *   rows as inputs and be evaluated once.
 * - Growing frame: We only add new rows into the frame, i.e. UNBOUNDED PRECEDING AND ....
 *   Every time we move to a new row to process, we add some rows to the frame. We do not remove
 *   rows from this frame.
 * - Shrinking frame: We only remove rows from the frame, i.e. ... AND UNBOUNDED FOLLOWING.
 *   Every time we move to a new row to process, we remove some rows from the frame. We do not add
 *   rows to this frame.
 * - Moving frame: Every time we move to a new row to process, we remove some rows from the frame
 *   and we add some rows to the frame. Examples are:
 *     1 PRECEDING AND CURRENT ROW and 1 FOLLOWING AND 2 FOLLOWING.
 * - Offset frame: The frame consist of one row, which is an offset number of rows away from the
 *   current row. Only [[OffsetWindowFunction]]s can be processed in an offset frame.
 *
 * Different frame boundaries can be used in Growing, Shrinking and Moving frames. A frame
 * boundary can be either Row or Range based:
 * - Row Based: A row based boundary is based on the position of the row within the partition.
 *   An offset indicates the number of rows above or below the current row, the frame for the
 *   current row starts or ends. For instance, given a row based sliding frame with a lower bound
 *   offset of -1 and a upper bound offset of +2. The frame for row with index 5 would range from
 *   index 4 to index 6.
 * - Range based: A range based boundary is based on the actual value of the ORDER BY
 *   expression(s). An offset is used to alter the value of the ORDER BY expression, for
 *   instance if the current order by expression has a value of 10 and the lower bound offset
 *   is -3, the resulting lower bound for the current row will be 10 - 3 = 7. This however puts a
 *   number of constraints on the ORDER BY expressions: there can be only one expression and this
 *   expression must have a numerical data type. An exception can be made when the offset is 0,
 *   because no value modification is needed, in this case multiple and non-numeric ORDER BY
 *   expression are allowed.
 *
 * This is quite an expensive operator because every row for a single group must be in the same
 * partition and partitions must be sorted according to the grouping and sort order. The operator
 * requires the planner to take care of the partitioning and sorting.
 *
 * The operator is semi-blocking. The window functions and aggregates are calculated one group at
 * a time, the result will only be made available after the processing for the entire group has
 * finished. The operator is able to process different frame configurations at the same time. This
 * is done by delegating the actual frame processing (i.e. calculation of the window functions) to
 * specialized classes, see [[WindowFunctionFrame]], which take care of their own frame type:
 * Entire Partition, Sliding, Growing & Shrinking. Boundary evaluation is also delegated to a pair
 * of specialized classes: [[RowBoundOrdering]] & [[RangeBoundOrdering]].
 */
case class WindowExec(
    windowExpression: Seq[NamedExpression],
    partitionSpec: Seq[Expression],
    orderSpec: Seq[SortOrder],
    child: SparkPlan)
  extends UnaryExecNode {

  override def output: Seq[Attribute] =
    child.output ++ windowExpression.map(_.toAttribute)

  override def requiredChildDistribution: Seq[Distribution] = {
    if (partitionSpec.isEmpty) {
      // Only show warning when the number of bytes is larger than 100 MB?
      logWarning("No Partition Defined for Window operation! Moving all data to a single "
        + "partition, this can cause serious performance degradation.")
      AllTuples :: Nil
    } else ClusteredDistribution(partitionSpec) :: Nil
  }

  override def requiredChildOrdering: Seq[Seq[SortOrder]] =
    Seq(partitionSpec.map(SortOrder(_, Ascending)) ++ orderSpec)

  override def outputOrdering: Seq[SortOrder] = child.outputOrdering

  /**
   * Create a bound ordering object for a given frame type and offset. A bound ordering object is
   * used to determine which input row lies within the frame boundaries of an output row.
   *
   * This method uses Code Generation. It can only be used on the executor side.
   *
   * @param frameType to evaluate. This can either be Row or Range based.
   * @param offset with respect to the row.
   * @return a bound ordering object.
   */
  private[this] def createBoundOrdering(frameType: FrameType, offset: Int): BoundOrdering = {
    frameType match {
      case RangeFrame =>
        val (exprs, current, bound) = if (offset == 0) {
          // Use the entire order expression when the offset is 0.
          val exprs = orderSpec.map(_.child)
          val buildProjection = () => newMutableProjection(exprs, child.output)
          (orderSpec, buildProjection(), buildProjection())
        } else if (orderSpec.size == 1) {
          // Use only the first order expression when the offset is non-null.
          val sortExpr = orderSpec.head
          val expr = sortExpr.child
          // Create the projection which returns the current 'value'.
          val current = newMutableProjection(expr :: Nil, child.output)
          // Flip the sign of the offset when processing the order is descending
          val boundOffset = sortExpr.direction match {
            case Descending => -offset
            case Ascending => offset
          }
          // Create the projection which returns the current 'value' modified by adding the offset.
          val boundExpr = Add(expr, Cast(Literal.create(boundOffset, IntegerType), expr.dataType))
          val bound = newMutableProjection(boundExpr :: Nil, child.output)
          (sortExpr :: Nil, current, bound)
        } else {
          sys.error("Non-Zero range offsets are not supported for windows " +
            "with multiple order expressions.")
        }
        // Construct the ordering. This is used to compare the result of current value projection
        // to the result of bound value projection. This is done manually because we want to use
        // Code Generation (if it is enabled).
        val sortExprs = exprs.zipWithIndex.map { case (e, i) =>
          SortOrder(BoundReference(i, e.dataType, e.nullable), e.direction)
        }
        val ordering = newOrdering(sortExprs, Nil)
        RangeBoundOrdering(ordering, current, bound)
      case RowFrame => RowBoundOrdering(offset)
    }
  }

  /**
   * Collection containing an entry for each window frame to process. Each entry contains a frames'
   * WindowExpressions and factory function for the WindowFrameFunction.
   */
  private[this] lazy val windowFrameExpressionFactoryPairs = {
    type FrameKey = (String, FrameType, Option[Int], Option[Int])
    type ExpressionBuffer = mutable.Buffer[Expression]
    val framedFunctions = mutable.Map.empty[FrameKey, (ExpressionBuffer, ExpressionBuffer)]

    // Add a function and its function to the map for a given frame.
    def collect(tpe: String, fr: SpecifiedWindowFrame, e: Expression, fn: Expression): Unit = {
      val key = (tpe, fr.frameType, FrameBoundary(fr.frameStart), FrameBoundary(fr.frameEnd))
      val (es, fns) = framedFunctions.getOrElseUpdate(
        key, (ArrayBuffer.empty[Expression], ArrayBuffer.empty[Expression]))
      es += e
      fns += fn
    }

    // Collect all valid window functions and group them by their frame.
    windowExpression.foreach { x =>
      x.foreach {
        case e @ WindowExpression(function, spec) =>
          val frame = spec.frameSpecification.asInstanceOf[SpecifiedWindowFrame]
          function match {
            case AggregateExpression(f, _, _, _) => collect("AGGREGATE", frame, e, f)
            case f: AggregateWindowFunction => collect("AGGREGATE", frame, e, f)
            case f: OffsetWindowFunction => collect("OFFSET", frame, e, f)
            case f => sys.error(s"Unsupported window function: $f")
          }
        case _ =>
      }
    }

    // Map the groups to a (unbound) expression and frame factory pair.
    var numExpressions = 0
    framedFunctions.toSeq.map {
      case (key, (expressions, functionSeq)) =>
        val ordinal = numExpressions
        val functions = functionSeq.toArray

        // Construct an aggregate processor if we need one.
        def processor = AggregateProcessor(
          functions,
          ordinal,
          child.output,
          (expressions, schema) =>
            newMutableProjection(expressions, schema, subexpressionEliminationEnabled))

        // Create the factory
        val factory = key match {
          // Offset Frame
          case ("OFFSET", RowFrame, Some(offset), Some(h)) if offset == h =>
            target: MutableRow =>
              new OffsetWindowFunctionFrame(
                target,
                ordinal,
                // OFFSET frame functions are guaranteed be OffsetWindowFunctions.
                functions.map(_.asInstanceOf[OffsetWindowFunction]),
                child.output,
                (expressions, schema) =>
                  newMutableProjection(expressions, schema, subexpressionEliminationEnabled),
                offset)

          // Growing Frame.
          case ("AGGREGATE", frameType, None, Some(high)) =>
            target: MutableRow => {
              new UnboundedPrecedingWindowFunctionFrame(
                target,
                processor,
                createBoundOrdering(frameType, high))
            }

          // Shrinking Frame.
          case ("AGGREGATE", frameType, Some(low), None) =>
            target: MutableRow => {
              new UnboundedFollowingWindowFunctionFrame(
                target,
                processor,
                createBoundOrdering(frameType, low))
            }

          // Moving Frame.
          case ("AGGREGATE", frameType, Some(low), Some(high)) =>
            target: MutableRow => {
              new SlidingWindowFunctionFrame(
                target,
                processor,
                createBoundOrdering(frameType, low),
                createBoundOrdering(frameType, high))
            }

          // Entire Partition Frame.
          case ("AGGREGATE", frameType, None, None) =>
            target: MutableRow => {
              new UnboundedWindowFunctionFrame(target, processor)
            }
        }

        // Keep track of the number of expressions. This is a side-effect in a map...
        numExpressions += expressions.size

        // Create the Frame Expression - Factory pair.
        (expressions, factory)
    }
  }

  /**
   * Create the resulting projection.
   *
   * This method uses Code Generation. It can only be used on the executor side.
   *
   * @param expressions unbound ordered function expressions.
   * @return the final resulting projection.
   */
  private[this] def createResultProjection(expressions: Seq[Expression]): UnsafeProjection = {
    val references = expressions.zipWithIndex.map{ case (e, i) =>
      // Results of window expressions will be on the right side of child's output
      BoundReference(child.output.size + i, e.dataType, e.nullable)
    }
    val unboundToRefMap = expressions.zip(references).toMap
    val patchedWindowExpression = windowExpression.map(_.transform(unboundToRefMap))
    UnsafeProjection.create(
      child.output ++ patchedWindowExpression,
      child.output)
  }

  protected override def doExecute(): RDD[InternalRow] = {
    // Unwrap the expressions and factories from the map.
    val expressions = windowFrameExpressionFactoryPairs.flatMap(_._1)
    val factories = windowFrameExpressionFactoryPairs.map(_._2).toArray

    // Start processing.
    child.execute().mapPartitions { stream =>
      new Iterator[InternalRow] {

        // Get all relevant projections.
        val result = createResultProjection(expressions)
        val grouping = UnsafeProjection.create(partitionSpec, child.output)

        // Manage the stream and the grouping.
        var nextRow: UnsafeRow = null
        var nextGroup: UnsafeRow = null
        var nextRowAvailable: Boolean = false
        private[this] def fetchNextRow() {
          nextRowAvailable = stream.hasNext
          if (nextRowAvailable) {
            nextRow = stream.next().asInstanceOf[UnsafeRow]
            nextGroup = grouping(nextRow)
          } else {
            nextRow = null
            nextGroup = null
          }
        }
        fetchNextRow()

        // Manage the current partition.
        val rows = ArrayBuffer.empty[UnsafeRow]
        val inputFields = child.output.length
        var sorter: UnsafeExternalSorter = null
        var rowBuffer: RowBuffer = null
        val windowFunctionResult = new SpecificMutableRow(expressions.map(_.dataType))
        val frames = factories.map(_(windowFunctionResult))
        val numFrames = frames.length
        private[this] def fetchNextPartition() {
          // Collect all the rows in the current partition.
          // Before we start to fetch new input rows, make a copy of nextGroup.
          val currentGroup = nextGroup.copy()

          // clear last partition
          if (sorter != null) {
            // the last sorter of this task will be cleaned up via task completion listener
            sorter.cleanupResources()
            sorter = null
          } else {
            rows.clear()
          }

          while (nextRowAvailable && nextGroup == currentGroup) {
            if (sorter == null) {
              rows += nextRow.copy()

              if (rows.length >= 4096) {
                // We will not sort the rows, so prefixComparator and recordComparator are null.
                sorter = UnsafeExternalSorter.create(
                  TaskContext.get().taskMemoryManager(),
                  SparkEnv.get.blockManager,
                  SparkEnv.get.serializerManager,
                  TaskContext.get(),
                  null,
                  null,
                  1024,
                  SparkEnv.get.memoryManager.pageSizeBytes,
                  SparkEnv.get.conf.getLong("spark.shuffle.spill.numElementsForceSpillThreshold",
                    UnsafeExternalSorter.DEFAULT_NUM_ELEMENTS_FOR_SPILL_THRESHOLD),
                  false)
                rows.foreach { r =>
                  sorter.insertRecord(r.getBaseObject, r.getBaseOffset, r.getSizeInBytes, 0, false)
                }
                rows.clear()
              }
            } else {
              sorter.insertRecord(nextRow.getBaseObject, nextRow.getBaseOffset,
                nextRow.getSizeInBytes, 0, false)
            }
            fetchNextRow()
          }
          if (sorter != null) {
            rowBuffer = new ExternalRowBuffer(sorter, inputFields)
          } else {
            rowBuffer = new ArrayRowBuffer(rows)
          }

          // Setup the frames.
          var i = 0
          while (i < numFrames) {
            frames(i).prepare(rowBuffer.copy())
            i += 1
          }

          // Setup iteration
          rowIndex = 0
          rowsSize = rowBuffer.size
        }

        // Iteration
        var rowIndex = 0
        var rowsSize = 0L

        override final def hasNext: Boolean = rowIndex < rowsSize || nextRowAvailable

        val join = new JoinedRow
        override final def next(): InternalRow = {
          // Load the next partition if we need to.
          if (rowIndex >= rowsSize && nextRowAvailable) {
            fetchNextPartition()
          }

          if (rowIndex < rowsSize) {
            // Get the results for the window frames.
            var i = 0
            val current = rowBuffer.next()
            while (i < numFrames) {
              frames(i).write(rowIndex, current)
              i += 1
            }

            // 'Merge' the input row with the window function result
            join(current, windowFunctionResult)
            rowIndex += 1

            // Return the projection.
            result(join)
          } else throw new NoSuchElementException
        }
      }
    }
  }
}