Replace SubString of values in a dataframe in Pyspark

Question

I have a dataframe with some attributes and it has the next appearence:

+-------+-------+
| Atr1  | Atr2  |
+-------+-------+
|  3,06 |  4,08 |
|  3,03 |  4,08 |
|  3,06 |  4,08 |
|  3,06 |  4,08 |
|  3,06 |  4,08 |
|  ...  |  ...  |
+-------+-------+

As you can see, the values of the Atr1 and Atr2 of the dataframe are numbers that has a ',' character. This is because I have loaded those data from a CSV where the decimals of the DoubleType numbers were represented by ','.

When I load the data into a dataframe the values are cast to String, so I applied a casting from String to DoubleType for those attributes like this:

df = df.withColumn("Atr1", df["Atr1"].cast(DoubleType()))
df = df.withColumn("Atr2", df["Atr2"].cast(DoubleType()))

But when I do it, the values are converted to null

+-------+-------+
| Atr1  | Atr2  |
+-------+-------+
|  null |  null |
|  null |  null |
|  null |  null |
|  null |  null |
|  null |  null |
|  ...  |  ...  |
+-------+-------+

I guess that the reason is that DoubleType decimals must be separated by '.' instead of by ','. But I don't have the chance to edit the CSV file, so I want to replace the ',' signs from the Dataframe by '.' and then apply the casting to DoubleType.

How could I do it?

Answer 1

You can simply solve this problem by using an user defined function.

from pyspark.sql.functions import UserDefinedFunction
from pyspark.sql.functions import *

data = [Row(Atr1="3,06", Atr2="4,08"),
        Row(Atr1="3,06", Atr2="4,08"),
        Row(Atr1="3,06", Atr2="4,08")]

df = sqlContext.createDataFrame(data)

# Create an user defined function to replace ',' for '.'
udf = UserDefinedFunction(lambda x: x.replace(",","."), StringType())

out = df
   .withColumn("Atr1", udf(col("Atr1")).cast(DoubleType()))
   .withColumn("Atr2", udf(col("Atr2")).cast(DoubleType()))

##############################################################
out.show()

+----+----+
|Atr1|Atr2|
+----+----+
|3.06|4.08|
|3.06|4.08|
|3.06|4.08|
+----+----+

##############################################################

out.printSchema()

root
 |-- Atr1: double (nullable = true)
 |-- Atr2: double (nullable = true)

EDIT: More compact solution following suggestion from comments.

from pyspark.sql.functions import UserDefinedFunction
from pyspark.sql.functions import *

udf = UserDefinedFunction(lambda x: float(x.replace(",",".")), DoubleType())

out = df
    .withColumn("Atr1", udf(col("Atr1")))
    .withColumn("Atr2", udf(col("Atr2")))

Answer 2

Let's assume you have:

sdf.show()
+-------+-------+
|   Atr1|   Atr2|
+-------+-------+
|  3,06 |  4,08 |
|  3,03 |  4,08 |
|  3,06 |  4,08 |
|  3,06 |  4,08 |
|  3,06 |  4,08 |
+-------+-------+

Then the following code will produce the desired result:

strToDouble = udf(lambda x: float(x.replace(",",".")), DoubleType())

sdf = sdf.withColumn("Atr1", strToDouble(sdf['Atr1']))
sdf = sdf.withColumn("Atr2", strToDouble(sdf['Atr2']))

sdf.show()
+----+----+
|Atr1|Atr2|
+----+----+
|3.06|4.08|
|3.03|4.08|
|3.06|4.08|
|3.06|4.08|
|3.06|4.08|
+----+----+

Answer 3

You can also do it with just SQL.

val df = sc.parallelize(Array(
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08"),
      ("3,06", "4,08")
      )).toDF("a", "b")

df.registerTempTable("test")

val doubleDF = sqlContext.sql("select cast(trim(regexp_replace( a , ',' , '.')) as double) as a from test ")

doubleDF.show
+----+
|   a|
+----+
|3.06|
|3.06|
|3.06|
|3.06|
|3.06|
|3.06|
|3.06|
|3.06|
+----+

doubleDF.printSchema
root
 |-- a: double (nullable = true)

Answer 4

is it possible to pass the column name as a parameter to the col() function in your sample code? Something like this:

# Create an user defined function to replace ',' for '.'
udf = UserDefinedFunction(lambda x: x.replace(",","."), StringType())

col_name1 = "Atr1"
col_name2 = "Atr2"

out = df
   .withColumn(col_name1, udf(col(col_name1)).cast(DoubleType()))
   .withColumn(col_name2, udf(col(col_name2)).cast(DoubleType()))