使用Scala，Spark UDF中的类型多态性将一系列地图平展到地图

Question

I have the following function that flattens a sequence of maps of string to double. How can I make type string to double generic?

val flattenSeqOfMaps = udf { values: Seq[Map[String, Double]] => values.flatten.toMap }
flattenSeqOfMaps: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,MapType(StringType,DoubleType,false),Some(List(ArrayType(MapType(StringType,DoubleType,false),true))))

I need something like,

val flattenSeqOfMaps[S,D] = udf { values: Seq[Map[S, D]] => values.flatten.toMap }

Thanks.

Edit 1: I'm using spark 2.3. I am aware of higher order functions in spark 2.4

Edit 2: I got a bit closer. What do I need in place of f _ in val flattenSeqOfMaps = udf { f _} . Please compare joinMap type signature and flattenSeqOfMaps type signature below

scala> val joinMap = udf { values: Seq[Map[String, Double]] => values.flatten.toMap }
joinMap: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,MapType(StringType,DoubleType,false),Some(List(ArrayType(MapType(StringType,DoubleType,false),true))))

scala> def f[S,D](values: Seq[Map[S, D]]): Map[S,D] = { values.flatten.toMap}
f: [S, D](values: Seq[Map[S,D]])Map[S,D]

scala> val flattenSeqOfMaps = udf { f _}
flattenSeqOfMaps: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,MapType(NullType,NullType,true),Some(List(ArrayType(MapType(NullType,NullType,true),true))))

Edit 3: the following code worked for me.

scala> val flattenSeqOfMaps = udf { f[String,Double] _}
flattenSeqOfMaps: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,MapType(StringType,DoubleType,false),Some(List(ArrayType(MapType(StringType,DoubleType,false),true))))

Answer 1

While you could define your function as

import scala.reflect.runtime.universe.TypeTag

def flattenSeqOfMaps[S : TypeTag, D: TypeTag] = udf { 
  values: Seq[Map[S, D]] => values.flatten.toMap
}

and then use specific instances:

val df = Seq(Seq(Map("a" -> 1), Map("b" -> 1))).toDF("val")

val flattenSeqOfMapsStringInt = flattenSeqOfMaps[String, Int]

df.select($"val", flattenSeqOfMapsStringInt($"val") as "val").show

+--------------------+----------------+
|                 val|             val|
+--------------------+----------------+
|[[a -> 1], [b -> 1]]|[a -> 1, b -> 1]|
+--------------------+----------------|

it is also possible to use built-in functions, without any need for explicit generics:

import org.apache.spark.sql.functions.{expr, flatten, map_from_arrays}

def flattenSeqOfMaps_(col: String) = {
  val keys = flatten(expr(s"transform(`$col`, x -> map_keys(x))"))
  val values = flatten(expr(s"transform(`$col`, x -> map_values(x))"))
  map_from_arrays(keys, values)
}

df.select($"val", flattenSeqOfMaps_("val") as "val").show

+--------------------+----------------+
|                 val|             val|
+--------------------+----------------+
|[[a -> 1], [b -> 1]]|[a -> 1, b -> 1]|
+--------------------+----------------+

Answer 2

The following code worked for me.

scala> def f[S,D](values: Seq[Map[S, D]]): Map[S,D] = { values.flatten.toMap}
f: [S, D](values: Seq[Map[S,D]])Map[S,D]

scala> val flattenSeqOfMaps = udf { f[String,Double] _}
flattenSeqOfMaps: org.apache.spark.sql.expressions.UserDefinedFunction = UserDefinedFunction(<function1>,MapType(StringType,DoubleType,false),Some(List(ArrayType(MapType(StringType,DoubleType,false),true))))