在基于列值进行分区后，将Bucketizer应用于Spark数据帧

Question

我需要在下面的数据框df上应用火花划水器。这是模型数据。原始数据帧大约有10k记录。

 instance   name                 value    percentage
 A37        Histogram.ratio      1            0.20
 A37        Histogram.ratio      20           0.34           
 A37        Histogram.ratio      50           0.04           
 A37        Histogram.ratio      500          0.13           
 A37        Histogram.ratio      2000         0.05           
 A37        Histogram.ratio      9000         0.32           
 A49        Histogram.ratio      1            0.50
 A49        Histogram.ratio      20           0.24           
 A49        Histogram.ratio      25           0.09           
 A49        Histogram.ratio      55           0.12           
 A49        Histogram.ratio      120          0.06           
 A49        Histogram.ratio      300          0.08

在按列实例对数据帧进行分区之后，我需要应用bucketizer。实例中的每个值都有下面定义的不同分割数组

val splits_map =  Map("A37" -> Array(0,30,1000,5000,9000), "A49" -> Array(0,10,30,80,998))

我将使用下面的代码在单列上执行bucketing。但是需要通过实例列对数据帧进行分区，然后应用bucketizer。使改变

val bucketizer = new Bucketizer().setInputCol("value").setOutputCol("value_range").setSplits(splits)
val df2 = bucketizer.transform(df)

df2.groupBy("value_range").sum("percentage").show()

是否可以使用列值实例将数据帧拆分为多个数据帧，然后对值列进行压缩，然后使用groupBy（）。sum（）计算百分比之和。

预期产出：

instance   name                 bucket    percentage
A37        Histogram.ratio      0            0.54                
A37        Histogram.ratio      1            0.17           
A37        Histogram.ratio      3            0.05           
A37        Histogram.ratio      4            0.32           
A49        Histogram.ratio      0            0.50
A49        Histogram.ratio      1            0.33                     
A49        Histogram.ratio      2            0.12           
A49        Histogram.ratio      3            0.14   




Answer 1

分区内数据备份的另一种方法是：

import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._
import spark.implicits._

def bucketizeWithinPartition(df: DataFrame, splits: Map[String, Array[Int]], partitionCol: String, featureCol: String): DataFrame = {
  val window = Window.partitionBy(partitionCol).orderBy($"bucket_start")

  val splitsDf = splits.toList.toDF(partitionCol, "splits")
    .withColumn("bucket_start", explode($"splits"))
    .withColumn("bucket_end", coalesce(lead($"bucket_start", 1).over(window), lit(Int.MaxValue)))
    .withColumn("bucket", row_number().over(window))

  val joinCond = "d.%s = s.%s AND d.%s >= s.bucket_start AND d.%s < bucket_end".format(partitionCol, partitionCol, featureCol, featureCol)
  df.as("d")
    .join(splitsDf.as("s"), expr(joinCond), "inner")
    .select($"d.*", $"s.bucket")
}


val data =
  List(
    ("A37", "Histogram.ratio", 1, 0.20),
    ("A37", "Histogram.ratio", 20, 0.34),
    ("A37", "Histogram.ratio", 9000, 0.32),
    ("A49", "Histogram.ratio", 1, 0.50),
    ("A49", "Histogram.ratio", 20, 0.24)
  ).toDF("instance", "name", "value", "percentage")

val splits_map =  Map("A37" -> Array(0,30,1000,5000,9000), "A49" -> Array(0,10,30,80,998))
val bucketedData = bucketizeWithinPartition(data, splits_map, "instance", "value")