Examples of custom visual scripts

The following examples perform equivalent transformations. However, the second example (SparkSQL) is the cleanest and most efficient, followed by the Pandas UDF and finally the low level mapping in the first example. The following example is a complete example of a simple transformation to add up two columns:

from awsglue import DynamicFrame
 
# You can have other auxiliary variables, functions or classes on this file, it won't affect the runtime
def record_sum(rec, col1, col2, resultCol):
    rec[resultCol] = rec[col1] + rec[col2]
    return rec
 
 
# The number and name of arguments must match the definition on json config file
# (expect self which is the current DynamicFrame to transform
# If an argument is optional, you need to define a default value here
#  (resultCol in this example is an optional argument)
def custom_add_columns(self, col1, col2, resultCol="result"):
    # The mapping will alter the columns order, which could be important
    fields = [field.name for field in self.schema()]
    if resultCol not in fields:
        # If it's a new column put it at the end
        fields.append(resultCol)
    return self.map(lambda record: record_sum(record, col1, col2, resultCol)).select_fields(paths=fields)
 
 
# The name we assign on DynamicFrame must match the configured "functionName"
DynamicFrame.custom_add_columns = custom_add_columns
        

The following example is an equivalent transform leveraging the SparkSQL API.

from awsglue import DynamicFrame
 
# The number and name of arguments must match the definition on json config file
# (expect self which is the current DynamicFrame to transform
# If an argument is optional, you need to define a default value here
#  (resultCol in this example is an optional argument)
def custom_add_columns(self, col1, col2, resultCol="result"):
    df = self.toDF()
    return DynamicFrame.fromDF(
        df.withColumn(resultCol, df[col1] + df[col2]) # This is the conversion logic
        , self.glue_ctx, self.name) 
 
 
# The name we assign on DynamicFrame must match the configured "functionName"
DynamicFrame.custom_add_columns = custom_add_columns           
        

The following example uses the same transformations but using a pandas UDF, which is more efficient that using a plain UDF. For more information about writing pandas UDFs see: Apache Spark SQL documentation.

from awsglue import DynamicFrame
import pandas as pd
from pyspark.sql.functions import pandas_udf
 
# The number and name of arguments must match the definition on json config file
# (expect self which is the current DynamicFrame to transform
# If an argument is optional, you need to define a default value here
#  (resultCol in this example is an optional argument)
def custom_add_columns(self, col1, col2, resultCol="result"):
    @pandas_udf("integer")  # We need to declare the type of the result column
    def add_columns(value1: pd.Series, value2: pd.Series) → pd.Series:
        return value1 + value2
 
    df = self.toDF()
    return DynamicFrame.fromDF(
        df.withColumn(resultCol, add_columns(col1, col2)) # This is the conversion logic
        , self.glue_ctx, self.name) 
 
# The name we assign on DynamicFrame must match the configured "functionName"
DynamicFrame.custom_add_columns = custom_add_columns