在Spark Scala中的列上运行累积/迭代Costum方法

Question

嗨，我是Spark/Scala的新手，我一直在尝试-AKA失败，根据特定的递归公式在火花数据帧中创建一列：

这里是伪代码。

someDf.col2[0] = 0

for i > 0
someDf.col2[i] = x * someDf.col1[i-1] + (1-x) * someDf.col2[i-1]

为了深入了解更多细节，这里是我的出发点：这个数据帧是在日期和个人id级别上聚合的结果。

所有进一步的计算都必须针对特定的id，并且必须考虑到前一周发生的事情。

为了说明这一点，我将这些值简化为0和1，删除了乘法器x和1-x，并将col2初始化为零。

var someDf = Seq(("2016-01-10 00:00:00.0","385608",0,0), 
         ("2016-01-17 00:00:00.0","385608",0,0),
         ("2016-01-24 00:00:00.0","385608",1,0),
         ("2016-01-31 00:00:00.0","385608",1,0),
         ("2016-02-07 00:00:00.0","385608",1,0),
         ("2016-02-14 00:00:00.0","385608",1,0),
         ("2016-01-17 00:00:00.0","105010",0,0),
         ("2016-01-24 00:00:00.0","105010",1,0),
         ("2016-01-31 00:00:00.0","105010",0,0),
         ("2016-02-07 00:00:00.0","105010",1,0)
        ).toDF("dates", "id", "col1","col2" )

someDf.show()

+--------------------+------+----+----+
|               dates|    id|col1|col2|
+--------------------+------+----+----+
|2016-01-10 00:00:...|385608|   0|   0|
|2016-01-17 00:00:...|385608|   0|   0|
|2016-01-24 00:00:...|385608|   1|   0|
|2016-01-31 00:00:...|385608|   1|   0|
|2016-02-07 00:00:...|385608|   1|   0|
|2016-02-14 00:00:...|385608|   1|   0|
+--------------------+------+----+----+
|2016-01-17 00:00:...|105010|   0|   0|
|2016-01-24 00:00:...|105010|   1|   0|
|2016-01-31 00:00:...|105010|   0|   0|
|2016-02-07 00:00:...|105010|   1|   0|
+--------------------+------+----+----+

到目前为止我所尝试的与所期望的

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

val date_id_window = Window.partitionBy("id").orderBy(asc("dates")) 

someDf.withColumn("col2", lag($"col1",1 ).over(date_id_window) + 
lag($"col2",1 ).over(date_id_window) ).show() 

+--------------------+------+----+----+ / +--------------------+
|               dates|    id|col1|col2| / | what_col2_should_be|
+--------------------+------+----+----+ / +--------------------+
|2016-01-17 00:00:...|105010|   0|null| / |                   0| 
|2016-01-24 00:00:...|105010|   1|   0| / |                   0|
|2016-01-31 00:00:...|105010|   0|   1| / |                   1|
|2016-02-07 00:00:...|105010|   1|   0| / |                   1|
+-------------------------------------+ / +--------------------+
|2016-01-10 00:00:...|385608|   0|null| / |                   0|
|2016-01-17 00:00:...|385608|   0|   0| / |                   0|
|2016-01-24 00:00:...|385608|   1|   0| / |                   0|
|2016-01-31 00:00:...|385608|   1|   1| / |                   1|
|2016-02-07 00:00:...|385608|   1|   1| / |                   2|
|2016-02-14 00:00:...|385608|   1|   1| / |                   3|
+--------------------+------+----+----+ / +--------------------+

有没有办法做到这一点与火花数据帧，我见过多个累积类型计算，但从来没有包括相同的列，我相信问题是，行i-1的新计算值没有被考虑，而是使用旧的i-1总是0。

任何帮助都将不胜感激。

Answer 1

您应该将转换应用于数据框，而不是将其视为var。获取所需内容的一种方法是使用Windows的row对每个窗口分区内的行通过前一行（即row-1）累积求和col1的值：

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

val window = Window.partitionBy("id").orderBy("dates").rowsBetween(Long.MinValue, -1)

val newDF = someDf.
  withColumn(
    "col2", sum($"col1").over(window)
  ).withColumn(
    "col2", when($"col2".isNull, 0).otherwise($"col2")
  ).orderBy("id", "dates")

newDF.show
+--------------------+------+----+----+
|               dates|    id|col1|col2|
+--------------------+------+----+----+
|2016-01-17 00:00:...|105010|   0|   0|
|2016-01-24 00:00:...|105010|   1|   0|
|2016-01-31 00:00:...|105010|   0|   1|
|2016-02-07 00:00:...|105010|   1|   1|
|2016-01-10 00:00:...|385608|   0|   0|
|2016-01-17 00:00:...|385608|   0|   0|
|2016-01-24 00:00:...|385608|   1|   0|
|2016-01-31 00:00:...|385608|   1|   1|
|2016-02-07 00:00:...|385608|   1|   2|
|2016-02-14 00:00:...|385608|   1|   3|
+--------------------+------+----+----+

Answer 2

你可以使用rowsapi和你正在使用的Window函数，你应该有想要的输出

val date_id_window = Window.partitionBy("id").orderBy(asc("dates"))
someDf.withColumn("col2", sum(lag($"col1", 1).over(date_id_window)).over(date_id_window.rowsBetween(Long.MinValue, 0)))
  .withColumn("col2", when($"col2".isNull, lit(0)).otherwise($"col2"))
  .show()

给定输入dataframeas

+--------------------+------+----+----+
|               dates|    id|col1|col2|
+--------------------+------+----+----+
|2016-01-10 00:00:...|385608|   0|   0|
|2016-01-17 00:00:...|385608|   0|   0|
|2016-01-24 00:00:...|385608|   1|   0|
|2016-01-31 00:00:...|385608|   1|   0|
|2016-02-07 00:00:...|385608|   1|   0|
|2016-02-14 00:00:...|385608|   1|   0|
|2016-01-17 00:00:...|105010|   0|   0|
|2016-01-24 00:00:...|105010|   1|   0|
|2016-01-31 00:00:...|105010|   0|   0|
|2016-02-07 00:00:...|105010|   1|   0|
+--------------------+------+----+----+

应用上述逻辑后，您应该有输出数据帧

+--------------------+------+----+----+
|               dates|    id|col1|col2|
+--------------------+------+----+----+
|2016-01-17 00:00:...|105010|   0|   0|
|2016-01-24 00:00:...|105010|   1|   0|
|2016-01-31 00:00:...|105010|   0|   1|
|2016-02-07 00:00:...|105010|   1|   1|
|2016-01-10 00:00:...|385608|   0|   0|
|2016-01-17 00:00:...|385608|   0|   0|
|2016-01-24 00:00:...|385608|   1|   0|
|2016-01-31 00:00:...|385608|   1|   1|
|2016-02-07 00:00:...|385608|   1|   2|
|2016-02-14 00:00:...|385608|   1|   3|
+--------------------+------+----+----+

我希望答案是有帮助的

Answer 3

数据集应该可以正常工作：

val x = 0.1

case class Record(dates: String, id: String, col1: Int)

someDf.drop("col2").as[Record].groupByKey(_.id).flatMapGroups((_,  records) => {
  val sorted = records.toSeq.sortBy(_.dates)
  sorted.scanLeft((null: Record, 0.0)){
    case ((_, col2), record) => (record, x * record.col1 + (1 - x) * col2)
  }.tail
}).select($"_1.*", $"_2".alias("col2"))