pyspark - apply a UDF to a list of columns and return several dataframes

Question

I have created a udf, that aims to ffill and bfill a column, and return back a new imputed dataframe. The error is not in the function as it works well.

See below my function:

def ffill_bfill(df,partition_by_col,order_by_col,col_to_imp):
  
    '''Forward fill and Backward fill a column by a column/set of columns (order_col).  
    Parameters:
    ------------
    df: Dataframe that the columns are in (Company wide? Company Narrow?) 
    order_col: String or List of string. This is the Year column until we get more granular time data!!
    fill_col: String (Only work for a column). The name of the column to be imputed!!

    Return:
    ---------
    df: Dataframe 
        Return df with the filled_cols. 
    '''
      
    # create the series containing the forward filled values             
    window_ff = Window.partitionBy(partition_by_col).orderBy(order_by_col).rowsBetween(-sys.maxsize, 0)
      
    # create the series containing the backward filled values  
    window_bf = Window.partitionBy(partition_by_col).orderBy(order_by_col).rowsBetween(0, sys.maxsize)
      
    # create the series containing the BACKWARD filled values for the two columns 
    s_bf = func.first(df[col_to_imp], ignorenulls=True).over(window_bf)
    
    # create the series containing the FORWARD filled values for the two columns
    s_ff = func.last(df[col_to_imp], ignorenulls=True).over(window_ff)

    # add the IMPUTED column to a dataframe 
    imputed_df = df_company_wide.withColumn(f'{col_to_imp}_bf', s_bf)\
                                .withColumn(f'{col_to_imp}_ff', s_ff)
  
    # Fill in the nulls with the imputed values
    imputed_df = imputed_df.withColumn(f'{col_to_imp}_imp',coalesce(col_to_imp,f'{col_to_imp}_ff',f'{col_to_imp}_bf'))
  
    # Create the imputed dataframes
    cols_to_use = ['isin','company','year',col_to_imp]+[s for s in imputed_df.columns if col_to_imp in s and 'imp' in s]
    imputed_df_final = imputed_df.select(cols_to_use)

    return imputed_df_final

The issue is with the way I apply the function:

My intention is to apply the function, in 4 columns, and return back 4 imputed dataframes. I try to do that with the below code:

# Get the columns to be imputed in a list
features_to_impute = ['mobile_maximum_plan_for_one',
                     'mobile_minimum_plan_for_one',
                     'slowest_internet_speed',
                     'fastest_internet_speed']

# Return a dataframe and make available for SQL
for feature in features_to_impute:
  f"{feature}_imp"= ffill_bfill(df_company_wide,partition_by_col='isin',order_by_col='year',col_to_imp=f"'{feature}'")  
  f"{feature}_imputed".createOrReplaceTempView(f"{feature}_imputed")

When I run the above command, I get the error:

SyntaxError: can't assign to literal
  File "<command-575233896480136>", line 21
    f"{feature}_imp"= ffill_bfill(df_company_wide,partition_by_col='isin',order_by_col='year',col_to_imp=f"'{feature}'")
    ^
SyntaxError: can't assign to literal

but when I try to apply the function, on 1 column at a time (like below), it works:

mobile_maximum_plan_for_one_imputed = ffill_bfill(df_company_wide,partition_by_col='isin',order_by_col='year',col_to_imp='mobile_maximum_plan_for_one')
mobile_minimum_plan_for_one_imputed.show()

+------------+----------------+------+---------------------------+-------------------------------+
|        isin|         company|  year|mobile_minimum_plan_for_one|mobile_minimum_plan_for_one_imp|
+------------+----------------+------+---------------------------+-------------------------------+
|BE0003810273|        Proximus|2015.0|                       null|              11.19820828667413|
|BE0003810273|        Proximus|2016.0|                       null|              11.19820828667413|
|BE0003810273|        Proximus|2017.0|                       null|              11.19820828667413|
|BE0003810273|        Proximus|2018.0|                       null|              11.19820828667413|
|BE0003810273|        Proximus|2019.0|          11.19820828667413|              11.19820828667413|
|CH0008742519|        Swisscom|2015.0|                       null|                          29.82|
|CH0008742519|        Swisscom|2016.0|                       null|                          29.82|
|CH0008742519|        Swisscom|2017.0|                       null|                          29.82|
|CH0008742519|        Swisscom|2018.0|                      29.82|                          29.82|
|CH0008742519|        Swisscom|2019.0|                      29.82|                          29.82|

Can someone shed some light on how to fix the for loop, to successful bring back the 4 different dataframes with the imputed values? A good explanation will add add a lot of value!

Many thanks in advance.

score 1 · Answer 1 · answered Mar 23 '21 at 12:21

Your code failed to run because you're assigning the dataframe to a string instead of a variable. In any case, using variable variable names is not considered a good practice. You can consider using a dictionary for this purpose.

features = dict()
for feature in features_to_impute:
  features[f"{feature}_imp"] = ffill_bfill(df_company_wide,partition_by_col='isin',order_by_col='year',col_to_imp=f"'{feature}'")  
  features[f"{feature}_imputed"].createOrReplaceTempView(f"{feature}_imputed")

pyspark - apply a UDF to a list of columns and return several dataframes

1 Answers1