Spark: Replace Null value in a Nested column

问题

I would like to replace all the n/a values in the below dataframe to unknown. It can be either scalar or complex nested column. If it's a StructField column I can loop through the columns and replace n\a using WithColumn. But I would like this to be done in a generic way inspite of the type of the column as I dont want to specify the column names explicitly as there are 100's in my case?

case class Bar(x: Int, y: String, z: String)
case class Foo(id: Int, name: String, status: String, bar: Seq[Bar])

val df = spark.sparkContext.parallelize(
Seq(
  Foo(123, "Amy", "Active", Seq(Bar(1, "first", "n/a"))),
  Foo(234, "Rick", "n/a", Seq(Bar(2, "second", "fifth"),Bar(22, "second", "n/a"))),
  Foo(567, "Tom", "null", Seq(Bar(3, "second", "sixth")))
)).toDF

df.printSchema
df.show(20, false)

Result:

+---+----+------+---------------------------------------+
|id |name|status|bar                                    |
+---+----+------+---------------------------------------+
|123|Amy |Active|[[1, first, n/a]]                      |
|234|Rick|n/a   |[[2, second, fifth], [22, second, n/a]]|
|567|Tom |null  |[[3, second, sixth]]                   |
+---+----+------+---------------------------------------+   

Expected Output:

+---+----+----------+---------------------------------------------------+
|id |name|status    |bar                                                |
+---+----+----------+---------------------------------------------------+
|123|Amy |Active    |[[1, first, unknown]]                              |
|234|Rick|unknown   |[[2, second, fifth], [22, second, unknown]]        |
|567|Tom |null      |[[3, second, sixth]]                               |
+---+----+----------+---------------------------------------------------+

Any suggestion on this?

回答1:

If you like playing with RDDs, here's a simple, generic and evolutive solution :

  val naToUnknown = {r: Row =>
    def rec(r: Any): Any = {
      r match {
        case row: Row => Row.fromSeq(row.toSeq.map(rec))
        case seq: Seq[Any] => seq.map(rec)
        case s: String if s == "n/a" => "unknown"
        case _ => r
      }
    }
    Row.fromSeq(r.toSeq.map(rec))
  }

  val newDF = spark.createDataFrame(df.rdd.map{naToUnknown}, df.schema)
  newDF.show(false)

Output :

+---+----+-------+-------------------------------------------+
|id |name|status |bar                                        |
+---+----+-------+-------------------------------------------+
|123|Amy |Active |[[1, first, unknown]]                      |
|234|Rick|unknown|[[2, second, fifth], [22, second, unknown]]|
|567|Tom |null   |[[3, second, sixth]]                       |
+---+----+-------+-------------------------------------------+

回答2:

It is somehow easy to replace nested values when you have only simple columns and structs. For array fields, you'll have to explode the structure before replacing or use UDF / higher-order functions, see my other answer here.

You can define a generic function that loops through DataFrame schema and apply a lambda function func to replace what you want:

def replaceNestedValues(schema: StructType, func: Column => Column, path: Option[String] = None): Seq[Column] = {
  schema.fields.map(f => {
    val p = path.fold(s"`${f.name}`")(c => s"$c.`${f.name}`")
    f.dataType match {
      case s: StructType => struct(replaceNestedValues(s, func, Some(p)): _*).alias(f.name)
      case _ => func(col(p)).alias(f.name)
    }
  })
}

Before using this function, explode the array structure bar like this:

val df2 = df.select($"id", $"name", $"status", explode($"bar").alias("bar"))

Then, define a lambda function that takes a column and replace it with unknown when it is equal to n/a using when/otherwise functions, and apply transformation to columns using the above function:

val replaceNaFunc: Column => Column = c => when(c === lit("n/a"), lit("unknown")).otherwise(c)
val replacedCols = replaceNestedValues(df2.schema, replaceNaFunc)

Select new columns and groupBy to get back the bar array:

df2.select(replacedCols: _*).groupBy($"id", $"name", $"status").agg(collect_list($"bar").alias("bar")).show(false)

Gives:

+---+----+-------+-------------------------------------------+                  
|id |name|status |bar                                        |
+---+----+-------+-------------------------------------------+
|234|Rick|unknown|[[2, second, fifth], [22, second, unknown]]|
|123|Amy |Active |[[1, first, unknown]]                      |
|567|Tom |null   |[[3, second, sixth]]                       |
+---+----+-------+-------------------------------------------+

回答3:

You can define an UDF to deal with your Array and replace the items you want:

UDF

 val replaceNA =  udf((x:Row) => {
      val z = x.getString(2)
      if ( z == "n/a")
        Bar(x.getInt(0), x.getString(1), "unknow")
      else
        Bar(x.getInt(0), x.getString(1), x.getString(2))
      })

Once you have that UDF, you can explode your dataframe to hace each item in bar as a single row:

 val explodedDF = df.withColumn("exploded", explode($"bar"))
+---+----+------+--------------------+------------------+
| id|name|status|                 bar|          exploded|
+---+----+------+--------------------+------------------+
|123| Amy|Active|   [[1, first, n/a]]|   [1, first, n/a]|
|234|Rick|   n/a|[[2, second, fift...|[2, second, fifth]|
|234|Rick|   n/a|[[2, second, fift...| [22, second, n/a]|
|567| Tom|  null|[[3, second, sixth]]|[3, second, sixth]|
+---+----+------+--------------------+------------------+ 

Then apply the previously defined UDF to replace the items:

val replacedDF = explodedDF.withColumn("exploded", replaceNA($"exploded"))
+---+----+------+--------------------+--------------------+
| id|name|status|                 bar|            exploded|
+---+----+------+--------------------+--------------------+
|123| Amy|Active|   [[1, first, n/a]]|  [1, first, unknow]|
|234|Rick|   n/a|[[2, second, fift...|  [2, second, fifth]|
|234|Rick|   n/a|[[2, second, fift...|[22, second, unknow]|
|567| Tom|  null|[[3, second, sixth]]|  [3, second, sixth]|
+---+----+------+--------------------+--------------------+

And finally grouping all together with collect_list to return it to it's original state

 val resultDF = replacedDF.groupBy("id", "name", "status")
      .agg(collect_list("exploded").as("bar")).show(false)
+---+----+------+------------------------------------------+
|id |name|status|bar                                       |
+---+----+------+------------------------------------------+
|234|Rick|n/a   |[[2, second, fifth], [22, second, unknow]]|
|567|Tom |null  |[[3, second, sixth]]                      |
|123|Amy |Active|[[1, first, unknow]]                      |
+---+----+------+------------------------------------------+

Putting al together in a single step:

import org.apache.spark.sql._

 val replaceNA =  udf((x:Row) => {
          val z = x.getString(2)
          if ( z == "n/a")
            Bar(x.getInt(0), x.getString(1), "unknow")
          else
            Bar(x.getInt(0), x.getString(1), x.getString(2))
          }) 

df.withColumn("exploded", explode($"bar"))
 .withColumn("exploded", replaceNA($"exploded"))
 .groupBy("id", "name", "status")
 .agg(collect_list("exploded").as("bar"))

来源：https://stackoverflow.com/questions/59536407/spark-replace-null-value-in-a-nested-column