[SPARK-12774] DataFrame.mapPartitions apply function operates on Pandas DataFrame instead of a generator or rows

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Won't Fix
• Component/s: PySpark
• Labels: dataframe mapPartitions pandas
• Assignee: Unassigned
• Reporter: Josh
• Created: 12/Jan/16 10:11
• Updated: 08/Oct/16 04:51
• Resolved: 08/Oct/16 04:51

Description

Currently DataFrame.mapPatitions is analogous to DataFrame.rdd.mapPatitions in both Spark and pySpark. The function that is applied to each partition f must operate on a list generator. This is however very inefficient in Python. It would be more logical and efficient if the apply function f operated on Pandas DataFrames instead and also returned a DataFrame. This avoids unnecessary iteration in Python which is slow.

Currently:

def apply_function(rows):
    df = pd.DataFrame(list(rows))
    df = df % 100   # Do something on df
    return df.values.tolist()

table = sqlContext.read.parquet("")
table = table.mapPatitions(apply_function)

New apply function would accept a Pandas DataFrame and return a DataFrame:

def apply_function(df):
    df = df % 100   # Do something on df
    return df