CODEWITHSUNDEEP
apache-sparkpysparkapache-spark-sql
Rob
Rob

Reputation: 369

How to add multiple columns using UDF?

Question

I want to add the return values of a UDF to an existing dataframe in seperate columns. How do I achieve this in a resourceful way?

Here's an example of what I have so far.

from pyspark.sql.functions import udf
from pyspark.sql.types import ArrayType, StructType, StructField, IntegerType  

df = spark.createDataFrame([("Alive",4)],["Name","Number"])
df.show(1)

+-----+------+
| Name|Number|
+-----+------+
|Alive|     4|
+-----+------+

def example(n):
        return [[n+2], [n-2]]

#  schema = StructType([
#          StructField("Out1", ArrayType(IntegerType()), False),
#          StructField("Out2", ArrayType(IntegerType()), False)])

example_udf = udf(example)

Now I can add a column to the dataframe as follows

newDF = df.withColumn("Output", example_udf(df["Number"]))
newDF.show(1)
+-----+------+----------+
| Name|Number|Output    |
+-----+------+----------+
|Alive|     4|[[6], [2]]|
+-----+------+----------+

However I don't want the two values to be in the same column but rather in separate ones.

Ideally I'd like to split the output column now to avoid calling the example function two times (once for each return value) as explained here and here, however in my situation I'm getting an array of arrays and I can't see how a split would work there (please note that each array will contain multiple values, separated with a ",".

How the result should look like

What I ultimately want is this

+-----+------+----+----+
| Name|Number|Out1|Out2|
+-----+------+----+----+
|Alive|     4|   6|   2|
+-----+------+----+----+

Note that the use of the StructType return type is optional and doesn't necessarily have to be part of the solution.

EDIT: I commented out the use of StructType (and edited the udf assignment) since it's not necessary for the return type of the example function. However it has to be used if the return value would be something like 

return [6,3,2],[4,3,1]

Upvotes: 20

Views: 25893

Answers (3)

Zhang Tong
Zhang Tong

Reputation: 4719

To return a StructType, just using Row

from pyspark.sql.types import StructType,StructField,IntegerType,Row
from pyspark.sql import functions as F

df = spark.createDataFrame([("Alive", 4)], ["Name", "Number"])


def example(n):
    return Row('Out1', 'Out2')(n + 2, n - 2)


schema = StructType([
    StructField("Out1", IntegerType(), False),
    StructField("Out2", IntegerType(), False)])

example_udf = F.UserDefinedFunction(example, schema)

newDF = df.withColumn("Output", example_udf(df["Number"]))
newDF = newDF.select("Name", "Number", "Output.*")

newDF.show(truncate=False)

Upvotes: 29

Tony Fraser
Tony Fraser

Reputation: 757

In scala

import spark.implicits
val df = Seq(("Alive", 4)).toDF("Name", "Number")

Without a UDF

df.
  withColumn("OutPlus",  $"Number" + 2).
  withColumn("OutMinus", $"Number" - 2).
  show
+-----+------+-------+--------+
| Name|Number|OutPlus|OutMinus|
+-----+------+-------+--------+
|Alive|     4|      6|       2|
+-----+------+-------+--------+

With a UDF using explode

import org.apache.spark.sql.functions.udf
def twoItems(_i: Int) = Seq((_i + 2, _i - 2))
val twoItemsUdf = udf(twoItems(_: Int))

val exploded = df.
  withColumn("Out", explode(twoItemsUdf($"Number"))).
  withColumn("OutPlus", $"Out._1").
  withColumn("OutMinus", $"Out._2")

exploded.printSchema

root
 |-- Name: string (nullable = true)
 |-- Number: integer (nullable = false)
 |-- Out: struct (nullable = true)
 |    |-- _1: integer (nullable = false)
 |    |-- _2: integer (nullable = false)
 |-- OutPlus: integer (nullable = true)
 |-- OutMinus: integer (nullable = true)

  exploded.drop("Out").show

+-----+------+-------+--------+
| Name|Number|OutPlus|OutMinus|
+-----+------+-------+--------+
|Alive|     4|      6|       2|
+-----+------+-------+--------+

Upvotes: 3

Manish Mehra
Manish Mehra

Reputation: 1501

Better way to solve above problem is by casting the output in an array and then exploding it

import pyspark.sql.functions as f
import pyspark.sql.types as t

df = spark.createDataFrame([("Alive", 4)], ["Name", "Number"])


def example(n):
    return t.Row('Out1', 'Out2')(n + 2, n - 2)


schema = StructType([
    StructField("Out1", t.IntegerType(), False),
    StructField("Out2", t.IntegerType(), False)])

example_udf = f.udf(example, schema)

newDF = df.withColumn("Output", f.explode(f.array(example_udf(df["Number"]))))
newDF = newDF.select("Name", "Number", "Output.*")

newDF.show(truncate=False)
newDF.explain()

Notice the output of explain, you will observe that example method is actually getting called only once!!

Upvotes: 22

Related Questions