Pandas API on spark runs too slow according to pandas

Question

I am making transform on my dataframe. while the process takes just 3 seconds with pandas, when I use pyspark and Pandas API on spark it takes approximately 30 minutes, yes 30 minutes! my data is 10k rows. The following is my pandas approach;

def find_difference_between_two_datetime(time1, time2):
      return int((time2-time1).total_seconds())

processed_data = pd.DataFrame()
for unique_ip in data.ip.unique():
      session_ids = []
      id = 1
      id_prefix = str(unique_ip) + "_"
      session_ids.append(id_prefix + str(id))
      ip_data = data[data.ip == unique_ip]
      timestamps= [time for time in ip_data.time]
      for item in zip(timestamps, timestamps[1:]):
             if find_difference_between_two_datetime(item[0], item[1]) > 30:
                    id +=1
             session_ids.append(id_prefix + str(id))

      ip_data["session_id"] = session_ids
      processed_data = pd.concat([processed_data, ip_data])

processed_data = processed_data.reset_index(drop=True)
processed_data

And the following is my pyspark - Pandas API on spark approach;

import pyspark.pandas as ps
def find_difference_between_two_datetime_spark(time1, time2):
        return int((time2-time1)/ 1000000000)

spark_processed_data = ps.DataFrame()
for unique_ip in data.ip.unique().to_numpy():
     session_ids = []
     id = 1
     id_prefix = str(unique_ip) + "_"
     session_ids.append(id_prefix + str(id))
     ip_data = data[data.ip == unique_ip]
     timestamps= ip_data.time.to_numpy()
     for item in zip(timestamps, timestamps[1:]):
          if find_difference_between_two_datetime_spark(item[0], item[1]) > 30:
                 id +=1
          session_ids.append(id_prefix + str(id))
     ip_data["session_id"] = session_ids
     spark_processed_data = ps.concat([spark_processed_data, ip_data])

spark_processed_data = spark_processed_data.reset_index(drop=True)
spark_processed_data

What I am missing about spark environment, I think it is not normal to run this code too slowly?

Answer 1

Spark offers distributed processing which can be very good for large datasets, but for smaller data sets, it can actually make things slower. You can take a look at this thread for more information.

10k rows is not a lot of data and you won’t really benefit much from Spark.