SPRKSCL1174

org.apache.spark.sql.functions.udf

Message: The single-parameter udf function is supported in Snowpark but it might require manual intervention. Please check the documentation to learn how to manually modify the code to make it work in Snowpark.

Category: Warning.

Description

This issue appears when the SMA detects an use of the single-parameter org.apache.spark.sql.functions.udf function in the code. Then, it might require a manual intervention.

The Snowpark API provides an equivalent com.snowflake.snowpark.functions.udf function that allows you to create a user-defined function from a lambda or function in Scala, however, there are some caveats about creating udf in Snowpark that might require you to make some manual changes to your code in order to make it work properly.

Scenarios

The Snowpark udf function should work as intended for a wide range of cases without requiring manual intervention. However, there are some scenarios that would requiere you to manually modify your code in order to get it work in Snowpark. Some of those scenarios are listed below:

Scenario 1

Input

Below is an example of creating UDFs in an object with the App Trait.

The Scala's App trait simplifies creating executable programs by providing a main method that automatically runs the code within the object definition. Extending App delays the initialization of the fields until the main method is executed, which can affect the UDFs definitions if they rely on initialized fields. This means that if an object extends App and the udf references an object field, the udf definition uploaded to Snowflake will not include the initialized value of the field. This can result in null values being returned by the udf.

For example, in the following code the variable myValue will resolve to null in the udf definition:

object Main extends App {
  ...
  val myValue = 10
  val myUdf = udf((x: Int) => x + myValue) // myValue in the `udf` definition will resolve to null
  ...
}

Output

The SMA adds the EWI SPRKSCL1174 to the output code to let you know that the single-parameter udf function is supported in Snowpark but it requires manual intervention.

object Main extends App {
  ...
  val myValue = 10
  /*EWI: SPRKSCL1174 => The single-parameter udf function is supported in Snowpark but it might require manual intervention. Please check the documentation to learn how to manually modify the code to make it work in Snowpark.*/
  val myUdf = udf((x: Int) => x + myValue) // myValue in the `udf` definition will resolve to null
  ...
}

Recommended fix

To avoid this issue, it is recommended to not extend App and implement a separate main method for your code. This ensure that object fields are initialized before udf definitions are created and uploaded to Snowflake.

object Main {
  ...
  def main(args: Array[String]): Unit = {
    val myValue = 10
    val myUdf = udf((x: Int) => x + myValue)
  }
  ...
}

For more details about this topic, see Caveat About Creating UDFs in an Object With the App Trait.

Scenario 2

Input

Below is an example of creating UDFs in Jupyter Notebooks.

def myFunc(s: String): String = {
  ...
}

val myFuncUdf = udf((x: String) => myFunc(x))
df1.select(myFuncUdf(col("name"))).show()

Output

The SMA adds the EWI SPRKSCL1174 to the output code to let you know that the single-parameter udf function is supported in Snowpark but it requires manual intervention.

def myFunc(s: String): String = {
  ...
}

/*EWI: SPRKSCL1174 => The single-parameter udf function is supported in Snowpark but it might require manual intervention. Please check the documentation to learn how to manually modify the code to make it work in Snowpark.*/
val myFuncUdf = udf((x: String) => myFunc(x))
df1.select(myFuncUdf(col("name"))).show()

Recommended fix

To create a udf in a Jupyter Notebook, you should define the implementation of your function in a class that extends Serializable. For example, you should manually convert it into this:

object ConvertedUdfFuncs extends Serializable {
  def myFunc(s: String): String = {
    ...
  }

  val myFuncAsLambda = ((x: String) => ConvertedUdfFuncs.myFunc(x))
}

val myFuncUdf = udf(ConvertedUdfFuncs.myFuncAsLambda)
df1.select(myFuncUdf(col("name"))).show()

For more details about how to create UDFs in Jupyter Notebooks, see Creating UDFs in Jupyter Notebooks.

Additional recommendations

• To learn more about how to create user-defined functions in Snowpark, please refer to the following documentation: Creating User-Defined Functions (UDFs) for DataFrames in Scala
• For more support, you can email us at [e-mail] or post an issue in the SMA.