Spark Scala：使用Spark订购不同日期后，需要获取空日期排在最前的记录

问题描述

我有以下数据：

+-----------+-----------+-----------+-----+-----------+
| Env1_date | Env2_date | Env3_date | Pid | orderDate |
+-----------+-----------+-----------+-----+-----------+
| Null      | Null      | 1/9/2020  | abc | 10/6/2020 |
| Null      | 1/9/2020  | 1/8/2020  | pqr | 10/4/2020 |
| 1/9/2020  | Null      | Null      | xyz | 10/2/2020 |
| 1/8/2020  | 1/7/2020  | Null      | uvw | 10/1/2020 |
+-----------+-----------+-----------+-----+-----------+

我正在尝试创建3个新列，这些新列基本上说明Pid是否对env1，env2和env3有效。为此，我首先按降序对orderDate列上的记录进行排序（已在上表中进行排序）。

如果对于Env1_date，Env2_date，Env3_date，最上面的记录是Null，则它们被认为是有效的。在Null记录之后，如果日期小于特定日期（在此示例中为1/9/2020），则认为该日期有效。其他任何记录都标记为无效。
如果排名靠前的记录不是NULL，则需要检查日期是否等于1/9/2020。如果是这样，它们也被标记为有效

我的输出应如下所示：

+-----------+-----------+-----------+-----+-----------+-----------+-----------+-----------+
| Env1_date | Env2_date | Env3_date | Pid | orderDate | Env1_Flag | Env2_Flag | Env3_Flag |
+-----------+-----------+-----------+-----+-----------+-----------+-----------+-----------+
| Null      | Null      | 1/9/2020  | abc | 10/6/2020 | Valid     | Valid     | Valid     |
| Null      | 1/9/2020  | 1/8/2020  | pqr | 10/4/2020 | Valid     | Valid     | Invalid   |
| 1/9/2020  | Null      | Null      | xyz | 10/2/2020 | Valid     | Invalid   | Invalid   |
| 1/8/2020  | 1/7/2020  | Null      | uvw | 10/1/2020 | Invalid   | Invalid   | Invalid   |
+-----------+-----------+-----------+-----+-----------+-----------+-----------+-----------+

我正在尝试使用Spark 1.5和scala来实现这一目标。

我尝试使用lag函数。但不能包含所有方案。不确定如何解决此问题。

有人可以在这里帮我吗？

注意：Windows函数toDf（）和createDataFrame（）函数在我正在使用的spark中不起作用。它是一个自定义的Spark环境，并且没有什么限制

解决方法

import spark.implicits._

case class Source(
                 Env1_date: Option[String],Env2_date: Option[String],Env3_date: Option[String],Pid: String,orderDate: String
               )
case class Source1(
                   Env1_date: Option[String],orderDate: String,Env1_Flag: String,Env2_Flag: String,Env3_Flag: String
                 )

val source = Seq(
  Source(None,None,Some("1/9/2020"),"abc","10/6/2020"),Source(None,Some("1/8/2020"),"pqr","10/4/2020"),Source(Some("1/9/2020"),"xyz","10/2/2020"),Source(Some("1/8/2020"),Some("1/7/2020"),"10/6/2020")
).toDF().as[Source].collect()

var env1NextRowInvalid = false
var env2NextRowInvalid = false
var env3NextRowInvalid = false
val source1 = source.map(i => {
  val env1Flag = if (env1NextRowInvalid == false && (i.Env1_date.getOrElse("") == """1/9/2020""" || i.Env1_date.getOrElse("") == "")) "valid" else "invalid"
  env1NextRowInvalid = if(env1NextRowInvalid == false) (i.Env1_date == "1/9/2020") else true
  val env2Flag = if (env2NextRowInvalid == false && (i.Env2_date.getOrElse("") == """1/9/2020""" || i.Env2_date.getOrElse("") == "")) "valid" else "invalid"
  env2NextRowInvalid = if(env2NextRowInvalid == false) (i.Env2_date.getOrElse("") == "1/9/2020") else true
  val env3Flag = if (env3NextRowInvalid == false && (i.Env3_date.getOrElse("") == """1/9/2020""" || i.Env3_date.getOrElse("") == "")) "valid" else "invalid"
  env3NextRowInvalid = if(env3NextRowInvalid == false) (i.Env3_date.getOrElse("") == "1/9/2020") else true
  Source1(i.Env1_date,i.Env2_date,i.Env3_date,i.Pid,i.orderDate,env1Flag,env2Flag,env3Flag)
})

val resDF = source1.toSeq.toDF()
resDF.show(false)
//  +---------+---------+---------+---+---------+---------+---------+---------+
//  |Env1_date|Env2_date|Env3_date|Pid|orderDate|Env1_Flag|Env2_Flag|Env3_Flag|
//  +---------+---------+---------+---+---------+---------+---------+---------+
//  |null     |null     |1/9/2020 |abc|10/6/2020|valid    |valid    |valid    |
//  |null     |1/9/2020 |1/8/2020 |pqr|10/4/2020|valid    |valid    |invalid  |
//  |1/9/2020 |null     |null     |xyz|10/2/2020|valid    |invalid  |invalid  |
//  |1/8/2020 |1/7/2020 |null     |abc|10/6/2020|invalid  |invalid  |invalid  |
//  +---------+---------+---------+---+---------+---------+---------+---------+

执行此操作的一种方法是，将所有数据收集到驱动程序并将其作为常规数组进行处理，然后再次将其转换为DF。但是请注意，数据应适合驱动程序。

我编写了可以处理您提供的数据的代码。如果您稍微调整一下（尤其是数据比较部分），您应该会得到预期的结果。

  // This is how your data is going to look like when you collect it with df.collect
  val arrayData = Array(
    Array("null","null","1/9/2020",Array("null","1/8/2020",Array("1/9/2020",Array("1/8/2020","1/7/2020","uvw","10/1/2020"),)

  // just printing
  arrayData.foreach(arr => println(arr.mkString(" \t| ")))
  println("-".repeat(30))

  // rotates the array,so column become rows and vice verse
  def shiftArray(arr: Array[Array[String]])
    = for(i <- arr(0).indices.toArray) yield arr.map(arr => arr(i))

  // the function that does the validation part
  val someDate = "1/9/2020"
  def processColumn(arr: Array[String]) = {
    val (startingNulls,rest) = arr.span(_ == "null")
    val startingNullsValidated: Array[String] = startingNulls.map(_ => "Valid")
    val restValidated: Array[String] = rest.map(date => if (date == someDate) "Valid" else "Invalid") // implement custom date comparison
    startingNullsValidated ++ restValidated
  }

  val shiftedArray: Array[Array[String]] = shiftArray(arrayData)

  // you need to validate only first 3 columns,so i used take/slice
  val validatedArray = {
    val columnsToProcess = shiftedArray.take(3)
    val otherColumns = shiftedArray.slice(3,shiftedArray.length)
    val processedColumns = for (arr <- columnsToProcess) yield processColumn(arr)
    processedColumns ++ otherColumns
  }

  // rotate array back
  val shiftBackValidatedArray = shiftArray(validatedArray)

  // just printing the final result
  shiftBackValidatedArray.foreach(arr => println(arr.mkString(" \t| ")))

这是上面打印线的输出

null    | null  | 1/9/2020  | abc   | 10/6/2020
null    | 1/9/2020  | 1/8/2020  | pqr   | 10/4/2020
1/9/2020    | null  | null  | xyz   | 10/2/2020
1/8/2020    | 1/7/2020  | null  | uvw   | 10/1/2020
------------------------------
Valid   | Valid     | Valid     | abc   | 10/6/2020
Valid   | Valid     | Invalid   | pqr   | 10/4/2020
Valid   | Invalid   | Invalid   | xyz   | 10/2/2020
Invalid     | Invalid   | Invalid   | uvw   | 10/1/2020

apache-spark apache-spark-1.5 apache-spark-sql scala scala