Performance for pyspark dataframe is very slow after using a @pandas_udf

[RRO]

Hello,

I am currently working on a time series forecasting with FBProphet. Since I have data with many time series groups (~3000) I use a @pandas_udf to parallelize the training.

@pandas_udf(schema, PandasUDFType.GROUPED_MAP)
def forecast_netprofit(prophtrain):
 
     ... 
 
     return results_pd
 
 
time_series_id_column_names = ['Grp1', 'Grp2', 'Grp3']
 
results = (prophtrain
           .groupby(time_series_id_column_names)
           .apply(forecast_netprofit)
          )

Now every time I want to display or do some operations on the results dataframe the performance is really low. For example: Just to display the first 1000 rows takes around 6min.

Is there a reason why the performance of the results is so slow and can I fix that somehow?

[-werners-]

Spark will run on the whole dataset in background and return 1000 rows of that. So it might be that, not necessarily the function itself.

You can test that by f.e. starting with a dataset of 1000 records and apply the function on that.

[RRO]

Alright, the dataset has around 80.000 rows and 12 columns - so it should not be to much to handle. I have different datasets that are bigger than that can be displayed within seconds. That is why I think it might be somehow related to the function...

[-werners-]

could be, although it should use Arrow these days.

What version of spark do you use?

[RRO]

Databricks Runtime Version: 10.3 ML (includes Apache Spark 3.2.1, Scala 2.12)

[Kaniz]

Hi @Rafael Rockenbach​ , This blog will demonstrate a performance benchmark in Apache Spark™ between Scala UDF, PySpark UDF, and PySpark Pandas UDF.

UDF-s are really slow in general because Spark cannot optimize them as it does with SQL/premade functions. 

[Hubert-Dudek]

Please specify the type hint in the function so you will save some time. Something similar to (can be different hints needed, it is an example):

@pandas_udf(schema)

def forecast_netprofit(prophtrain: pd.Series) -> pd.Series

you could consider also using .agg(forecast_netprofit(prophtrain)) instead of .apply()

[RRO · Accepted solution]

Thank you for the answers.

Unfortunately this did not solve the performance issue.

What I did now is I saved the results into a table:

results.write.mode("overwrite").saveAsTable("db.results")

This is probably not the best solution but after I do that I can work with the results data from the table.