[jvm-packages] enable deterministic repartitioning when checkpoint is enabled (#4807)

* do reparititoning in DataUtil * keep previous behavior of partitioning without checkpoint * deterministic repartitioning * change
2019-09-19 15:21:05 -07:00 · 2019-09-19 15:21:05 -07:00 · fc8c9b0521
commit fc8c9b0521
parent 277e25797b
6 changed files with 204 additions and 28 deletions
--- a/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/DataUtils.scala
+++ b/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/DataUtils.scala
@ -18,6 +18,7 @@ package ml.dmlc.xgboost4j.scala.spark

 import ml.dmlc.xgboost4j.{LabeledPoint => XGBLabeledPoint}

+import org.apache.spark.HashPartitioner
 import org.apache.spark.ml.feature.{LabeledPoint => MLLabeledPoint}
 import org.apache.spark.ml.linalg.{DenseVector, SparseVector, Vector, Vectors}
 import org.apache.spark.ml.param.Param
@ -73,12 +74,81 @@ object DataUtils extends Serializable {
    }
  }

+  private def featureValueOfDenseVector(rowHashCode: Int, features: DenseVector): Float = {
+    val featureId = {
+      if (rowHashCode > 0) {
+        rowHashCode % features.size
+      } else {
+        // prevent overflow
+        math.abs(rowHashCode + 1) % features.size
+      }
+    }
+    features.values(featureId).toFloat
+  }
+
+  private def featureValueOfSparseVector(rowHashCode: Int, features: SparseVector): Float = {
+    val featureId = {
+      if (rowHashCode > 0) {
+        rowHashCode % features.indices.length
+      } else {
+        // prevent overflow
+        math.abs(rowHashCode + 1) % features.indices.length
+      }
+    }
+    features.values(featureId).toFloat
+  }
+
+  private def calculatePartitionKey(row: Row, numPartitions: Int): Int = {
+    val Row(_, features: Vector, _, _) = row
+    val rowHashCode = row.hashCode()
+    val featureValue = features match {
+      case denseVector: DenseVector =>
+        featureValueOfDenseVector(rowHashCode, denseVector)
+      case sparseVector: SparseVector =>
+        featureValueOfSparseVector(rowHashCode, sparseVector)
+    }
+    math.abs((rowHashCode.toLong + featureValue).toString.hashCode % numPartitions)
+  }
+
+  private def attachPartitionKey(
+      row: Row,
+      deterministicPartition: Boolean,
+      numWorkers: Int,
+      xgbLp: XGBLabeledPoint): (Int, XGBLabeledPoint) = {
+    if (deterministicPartition) {
+      (calculatePartitionKey(row, numWorkers), xgbLp)
+    } else {
+      (1, xgbLp)
+    }
+  }
+
+  private def repartitionRDDs(
+      deterministicPartition: Boolean,
+      numWorkers: Int,
+      arrayOfRDDs: Array[RDD[(Int, XGBLabeledPoint)]]): Array[RDD[XGBLabeledPoint]] = {
+    if (deterministicPartition) {
+      arrayOfRDDs.map {rdd => rdd.partitionBy(new HashPartitioner(numWorkers))}.map {
+        rdd => rdd.map(_._2)
+      }
+    } else {
+      arrayOfRDDs.map(rdd => {
+        if (rdd.getNumPartitions != numWorkers) {
+          rdd.map(_._2).repartition(numWorkers)
+        } else {
+          rdd.map(_._2)
+        }
+      })
+    }
+  }
+
  private[spark] def convertDataFrameToXGBLabeledPointRDDs(
      labelCol: Column,
      featuresCol: Column,
      weight: Column,
      baseMargin: Column,
      group: Option[Column],
+      numWorkers: Int,
+      deterministicPartition: Boolean,
      dataFrames: DataFrame*): Array[RDD[XGBLabeledPoint]] = {
    val selectedColumns = group.map(groupCol => Seq(labelCol.cast(FloatType),
      featuresCol,
@ -88,22 +158,26 @@ object DataUtils extends Serializable {
      featuresCol,
      weight.cast(FloatType),
      baseMargin.cast(FloatType)))
-    dataFrames.toArray.map {
+    val arrayOfRDDs = dataFrames.toArray.map {
      df => df.select(selectedColumns: _*).rdd.map {
-        case Row(label: Float, features: Vector, weight: Float, group: Int, baseMargin: Float) =>
+        case row @ Row(label: Float, features: Vector, weight: Float, group: Int,
+          baseMargin: Float) =>
          val (indices, values) = features match {
            case v: SparseVector => (v.indices, v.values.map(_.toFloat))
            case v: DenseVector => (null, v.values.map(_.toFloat))
          }
-          XGBLabeledPoint(label, indices, values, weight, group, baseMargin)
-        case Row(label: Float, features: Vector, weight: Float, baseMargin: Float) =>
+          val xgbLp = XGBLabeledPoint(label, indices, values, weight, group, baseMargin)
+          attachPartitionKey(row, deterministicPartition, numWorkers, xgbLp)
+        case row @ Row(label: Float, features: Vector, weight: Float, baseMargin: Float) =>
          val (indices, values) = features match {
            case v: SparseVector => (v.indices, v.values.map(_.toFloat))
            case v: DenseVector => (null, v.values.map(_.toFloat))
          }
-          XGBLabeledPoint(label, indices, values, weight, baseMargin = baseMargin)
+          val xgbLp = XGBLabeledPoint(label, indices, values, weight, baseMargin = baseMargin)
+          attachPartitionKey(row, deterministicPartition, numWorkers, xgbLp)
      }
    }
+    repartitionRDDs(deterministicPartition, numWorkers, arrayOfRDDs)
  }

 }
--- a/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoost.scala
+++ b/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoost.scala
@ -475,8 +475,7 @@ object XGBoost extends Serializable {
      Left(cacheData(ifCacheDataBoolean, repartitionedData).
        asInstanceOf[RDD[Array[XGBLabeledPoint]]])
    } else {
-      val repartitionedData = repartitionForTraining(trainingData, nWorkers)
-      Right(cacheData(ifCacheDataBoolean, repartitionedData).asInstanceOf[RDD[XGBLabeledPoint]])
+      Right(cacheData(ifCacheDataBoolean, trainingData).asInstanceOf[RDD[XGBLabeledPoint]])
    }
  }

@ -568,15 +567,6 @@ object XGBoost extends Serializable {
    }
  }

-  private[spark] def repartitionForTraining(trainingData: RDD[XGBLabeledPoint], nWorkers: Int) = {
-    if (trainingData.getNumPartitions != nWorkers) {
-      logger.info(s"repartitioning training set to $nWorkers partitions")
-      trainingData.repartition(nWorkers)
-    } else {
-      trainingData
-    }
-  }
-
  private def aggByGroupInfo(trainingData: RDD[XGBLabeledPoint]) = {
    val normalGroups: RDD[Array[XGBLabeledPoint]] = trainingData.mapPartitions(
      // LabeledPointGroupIterator returns (Boolean, Array[XGBLabeledPoint])
--- a/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostClassifier.scala
+++ b/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostClassifier.scala
@ -37,10 +37,6 @@ import org.json4s.DefaultFormats
 import scala.collection.JavaConverters._
 import scala.collection.{AbstractIterator, Iterator, mutable}

-private[spark] trait XGBoostClassifierParams extends GeneralParams with LearningTaskParams
-  with BoosterParams with HasWeightCol with HasBaseMarginCol with HasNumClass with ParamMapFuncs
-  with HasLeafPredictionCol with HasContribPredictionCol with NonParamVariables
-
 class XGBoostClassifier (
    override val uid: String,
    private val xgboostParams: Map[String, Any])
@ -182,11 +178,11 @@ class XGBoostClassifier (

    val trainingSet: RDD[XGBLabeledPoint] = DataUtils.convertDataFrameToXGBLabeledPointRDDs(
      col($(labelCol)), col($(featuresCol)), weight, baseMargin,
-      None, dataset.asInstanceOf[DataFrame]).head
+      None, $(numWorkers), needDeterministicRepartitioning, dataset.asInstanceOf[DataFrame]).head
    val evalRDDMap = getEvalSets(xgboostParams).map {
      case (name, dataFrame) => (name,
        DataUtils.convertDataFrameToXGBLabeledPointRDDs(col($(labelCol)), col($(featuresCol)),
-          weight, baseMargin, None, dataFrame).head)
+          weight, baseMargin, None, $(numWorkers), needDeterministicRepartitioning, dataFrame).head)
    }
    transformSchema(dataset.schema, logging = true)
    val derivedXGBParamMap = MLlib2XGBoostParams
--- a/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostEstimatorCommon.scala
+++ b/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostEstimatorCommon.scala
@ -0,0 +1,37 @@
+/*
+ Copyright (c) 2014 by Contributors
+
+ Licensed 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 ml.dmlc.xgboost4j.scala.spark
+
+import ml.dmlc.xgboost4j.scala.spark.params._
+
+import org.apache.spark.ml.param.shared.HasWeightCol
+
+private[spark] sealed trait XGBoostEstimatorCommon extends GeneralParams with LearningTaskParams
+  with BoosterParams with ParamMapFuncs with NonParamVariables {
+
+  def needDeterministicRepartitioning: Boolean = {
+    getCheckpointPath.nonEmpty && getCheckpointInterval > 0
+  }
+}
+
+private[spark] trait XGBoostClassifierParams extends HasWeightCol with HasBaseMarginCol
+  with HasNumClass with HasLeafPredictionCol with HasContribPredictionCol
+  with XGBoostEstimatorCommon
+
+private[spark] trait XGBoostRegressorParams extends HasBaseMarginCol with HasWeightCol
+  with HasGroupCol with HasLeafPredictionCol with HasContribPredictionCol
+  with XGBoostEstimatorCommon
--- a/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostRegressor.scala
+++ b/jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostRegressor.scala
@ -41,10 +41,6 @@ import scala.collection.mutable.ListBuffer

 import org.apache.spark.broadcast.Broadcast

-private[spark] trait XGBoostRegressorParams extends GeneralParams with BoosterParams
-  with LearningTaskParams with HasBaseMarginCol with HasWeightCol with HasGroupCol
-  with ParamMapFuncs with HasLeafPredictionCol with HasContribPredictionCol with NonParamVariables
-
 class XGBoostRegressor (
    override val uid: String,
    private val xgboostParams: Map[String, Any])
@ -178,11 +174,12 @@ class XGBoostRegressor (
    val group = if (!isDefined(groupCol) || $(groupCol).isEmpty) lit(-1) else col($(groupCol))
    val trainingSet: RDD[XGBLabeledPoint] = DataUtils.convertDataFrameToXGBLabeledPointRDDs(
      col($(labelCol)), col($(featuresCol)), weight, baseMargin, Some(group),
-      dataset.asInstanceOf[DataFrame]).head
+      $(numWorkers), needDeterministicRepartitioning, dataset.asInstanceOf[DataFrame]).head
    val evalRDDMap = getEvalSets(xgboostParams).map {
      case (name, dataFrame) => (name,
        DataUtils.convertDataFrameToXGBLabeledPointRDDs(col($(labelCol)), col($(featuresCol)),
-          weight, baseMargin, Some(group), dataFrame).head)
+          weight, baseMargin, Some(group), $(numWorkers), needDeterministicRepartitioning,
+          dataFrame).head)
    }
    transformSchema(dataset.schema, logging = true)
    val derivedXGBParamMap = MLlib2XGBoostParams
--- a/jvm-packages/xgboost4j-spark/src/test/scala/ml/dmlc/xgboost4j/scala/spark/DeterministicPartitioningSuite.scala
+++ b/jvm-packages/xgboost4j-spark/src/test/scala/ml/dmlc/xgboost4j/scala/spark/DeterministicPartitioningSuite.scala
@ -0,0 +1,82 @@
+/*
+ Copyright (c) 2014 by Contributors
+
+ Licensed 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 ml.dmlc.xgboost4j.scala.spark
+
+import org.scalatest.FunSuite
+
+import org.apache.spark.sql.functions._
+
+class DeterministicPartitioningSuite extends FunSuite with TmpFolderPerSuite with PerTest {
+
+  test("perform deterministic partitioning when checkpointInternal and" +
+    " checkpointPath is set (Classifier)") {
+    val tmpPath = createTmpFolder("model1").toAbsolutePath.toString
+    val paramMap = Map("eta" -> "1", "max_depth" -> 2,
+      "objective" -> "binary:logistic", "checkpoint_path" -> tmpPath,
+      "checkpoint_interval" -> 2, "num_workers" -> numWorkers)
+    val xgbClassifier = new XGBoostClassifier(paramMap)
+    assert(xgbClassifier.needDeterministicRepartitioning)
+  }
+
+  test("perform deterministic partitioning when checkpointInternal and" +
+    " checkpointPath is set (Regressor)") {
+    val tmpPath = createTmpFolder("model1").toAbsolutePath.toString
+    val paramMap = Map("eta" -> "1", "max_depth" -> 2,
+      "objective" -> "binary:logistic", "checkpoint_path" -> tmpPath,
+      "checkpoint_interval" -> 2, "num_workers" -> numWorkers)
+    val xgbRegressor = new XGBoostRegressor(paramMap)
+    assert(xgbRegressor.needDeterministicRepartitioning)
+  }
+
+  test("deterministic partitioning takes effect with various parts of data") {
+    val trainingDF = buildDataFrame(Classification.train)
+    // the test idea is that, we apply a chain of repartitions over trainingDFs but they
+    // have to produce the identical RDDs
+    val transformedDFs = (1 until 6).map(shuffleCount => {
+      var resultDF = trainingDF
+      for (i <- 0 until shuffleCount) {
+        resultDF = resultDF.repartition(numWorkers)
+      }
+      resultDF
+    })
+    val transformedRDDs = transformedDFs.map(df => DataUtils.convertDataFrameToXGBLabeledPointRDDs(
+      col("label"),
+      col("features"),
+      lit(1.0),
+      lit(Float.NaN),
+      None,
+      numWorkers,
+      deterministicPartition = true,
+      df
+    ).head)
+    val resultsMaps = transformedRDDs.map(rdd => rdd.mapPartitionsWithIndex {
+      case (partitionIndex, labelPoints) =>
+        Iterator((partitionIndex, labelPoints.toList))
+    }.collect().toMap)
+    resultsMaps.foldLeft(resultsMaps.head) { case (map1, map2) =>
+      assert(map1.keys.toSet === map2.keys.toSet)
+      for ((parIdx, labeledPoints) <- map1) {
+        val sortedA = labeledPoints.sortBy(_.hashCode())
+        val sortedB = map2(parIdx).sortBy(_.hashCode())
+        assert(sortedA.length === sortedB.length)
+        assert(sortedA.indices.forall(idx =>
+          sortedA(idx).values.toSet === sortedB(idx).values.toSet))
+      }
+      map2
+    }
+  }
+}