apache-sparkapache-spark-sqludfnumberformatexception

spark dataframe parse csv with non US format strange error

I have a dataframe in spark which contains a column of

df.select("y_wgs84").show
+----------------+
|         y_wgs84|
+----------------+
|47,9882373902965|
|47,9848921211406|
|47,9781530280939|
|47,9731284286555|
|47,9889813907224|
|47,9881440349524|
|47,9744969812356|
|47,9779388492231|
|48,0107946653620|
|48,0161245749621|
|48,0176065577678|
|48,0029496680229|
|48,0061848607139|
|47,9947482295108|
|48,0055828684523|
|48,0148743653486|
|48,0163361315735|
|48,0071490870937|
|48,0178054077099|
|47,8670099558802|
+----------------+

As these were read by spark.read.csv() its schema is of type String. Now I want to convert it to a double as follows:

val format = NumberFormat.getInstance(Locale.GERMANY)
def toDouble: UserDefinedFunction = udf[Double, String](format.parse(_).doubleValue)
df2.withColumn("y_wgs84", toDouble('y_wgs84)).collect

but it fails with java.lang.NumberFormatException: For input string: ".E0" Strangely though, when grepping the file, there is no single record containing an E.

Additionally, df.select("y_wgs84").as[String].collect.map(format.parse(_).doubleValue) this will work just fine. What is wrong here when calling the function as an UDF in spark?

Solution

  • Actually, thread safety is the problem. So changing the parsing function to 

    def toDouble: UserDefinedFunction = udf[Double, String](_.replace(',', '.').toDouble)

    works just fine.