SparkSQL:如何处理用户定义函数中的空值?

SparkSQL: How to deal with null values in user defined function?

给定 Table 1,其中一列 "x" 类型为 String。 我想创建 Table 2,其列 "y" 是 "x".

中给出的日期字符串的整数表示

基本 是在 "y".

列中保留 null 个值

Table 1(数据框 df1):

+----------+
|         x|
+----------+
|2015-09-12|
|2015-09-13|
|      null|
|      null|
+----------+
root
 |-- x: string (nullable = true)

Table 2(数据框 df2):

+----------+--------+                                                                  
|         x|       y|
+----------+--------+
|      null|    null|
|      null|    null|
|2015-09-12|20150912|
|2015-09-13|20150913|
+----------+--------+
root
 |-- x: string (nullable = true)
 |-- y: integer (nullable = true)

而将列 "x" 的值转换为列 "y" 的值的用户定义函数 (udf) 是:

val extractDateAsInt = udf[Int, String] (
  (d:String) => d.substring(0, 10)
      .filterNot( "-".toSet)
      .toInt )

并且有效,无法处理空值。

尽管如此,我可以做类似

的事情
val extractDateAsIntWithNull = udf[Int, String] (
  (d:String) => 
    if (d != null) d.substring(0, 10).filterNot( "-".toSet).toInt 
    else 1 )

我没找到办法,通过udfs给"produce" null值(当然,因为Ints不能null)。

我目前创建df2 (Table 2)的方案如下:

// holds data of table 1  
val df1 = ... 

// filter entries from df1, that are not null
val dfNotNulls = df1.filter(df1("x")
  .isNotNull)
  .withColumn("y", extractDateAsInt(df1("x")))
  .withColumnRenamed("x", "right_x")

// create df2 via a left join on df1 and dfNotNull having 
val df2 = df1.join( dfNotNulls, df1("x") === dfNotNulls("right_x"), "leftouter" ).drop("right_x")

问题:

代码摘录

val extractDateAsNullableInt = udf[NullableInt, String] (
  (d:String) => 
    if (d != null) d.substring(0, 10).filterNot( "-".toSet).toInt 
    else null )

这是 Option 派上用场的地方:

val extractDateAsOptionInt = udf((d: String) => d match {
  case null => None
  case s => Some(s.substring(0, 10).filterNot("-".toSet).toInt)
})

或者在一般情况下使其稍微更安全:

import scala.util.Try

val extractDateAsOptionInt = udf((d: String) => Try(
  d.substring(0, 10).filterNot("-".toSet).toInt
).toOption)

所有功劳归于 Dmitriy Selivanov who've pointed out this solution as a (missing?) edit here

替代方法是在 UDF 外部处理 null

import org.apache.spark.sql.functions.{lit, when}
import org.apache.spark.sql.types.IntegerType

val extractDateAsInt = udf(
   (d: String) => d.substring(0, 10).filterNot("-".toSet).toInt
)

df.withColumn("y",
  when($"x".isNull, lit(null))
    .otherwise(extractDateAsInt($"x"))
    .cast(IntegerType)
)

补充代码

使用@zero323 的 nice 答案,我创建了以下代码,以使用户定义的函数可用于处理所描述的空值。希望对其他人有帮助!

/**
 * Set of methods to construct [[org.apache.spark.sql.UserDefinedFunction]]s that
 * handle `null` values.
 */
object NullableFunctions {

  import org.apache.spark.sql.functions._
  import scala.reflect.runtime.universe.{TypeTag}
  import org.apache.spark.sql.UserDefinedFunction

  /**
   * Given a function A1 => RT, create a [[org.apache.spark.sql.UserDefinedFunction]] such that
   *   * if fnc input is null, None is returned. This will create a null value in the output Spark column.
   *   * if A1 is non null, Some( f(input) will be returned, thus creating f(input) as value in the output column.
   * @param f function from A1 => RT
   * @tparam RT return type
   * @tparam A1 input parameter type
   * @return a [[org.apache.spark.sql.UserDefinedFunction]] with the behaviour describe above
   */
  def nullableUdf[RT: TypeTag, A1: TypeTag](f: Function1[A1, RT]): UserDefinedFunction = {
    udf[Option[RT],A1]( (i: A1) => i match {
      case null => None
      case s => Some(f(i))
    })
  }

  /**
   * Given a function A1, A2 => RT, create a [[org.apache.spark.sql.UserDefinedFunction]] such that
   *   * if on of the function input parameters is null, None is returned.
   *     This will create a null value in the output Spark column.
   *   * if both input parameters are non null, Some( f(input) will be returned, thus creating f(input1, input2)
   *     as value in the output column.
   * @param f function from A1 => RT
   * @tparam RT return type
   * @tparam A1 input parameter type
   * @tparam A2 input parameter type
   * @return a [[org.apache.spark.sql.UserDefinedFunction]] with the behaviour describe above
   */
  def nullableUdf[RT: TypeTag, A1: TypeTag, A2: TypeTag](f: Function2[A1, A2, RT]): UserDefinedFunction = {
    udf[Option[RT], A1, A2]( (i1: A1, i2: A2) =>  (i1, i2) match {
      case (null, _) => None
      case (_, null) => None
      case (s1, s2) => Some((f(s1,s2)))
    } )
  }
}

Scala 实际上有一个很好的工厂函数,Option(),它可以使这个更加简洁:

val extractDateAsOptionInt = udf((d: String) => 
  Option(d).map(_.substring(0, 10).filterNot("-".toSet).toInt))

在内部,Option 对象的 apply 方法只是为您做 null 检查:

def apply[A](x: A): Option[A] = if (x == null) None else Some(x)