Hi Abdeali,
Thanks for the support. Indeed you can go ahead and test and review my latest PR for SPARK-34265 (Instrument Python UDF execution using SQL Metrics) if you want to: https://github.com/apache/spark/pull/33559 Currently I reduced the scope of the instrumentation to just 3 simple metrics to implement: "data sent to Python workers", "data returned from Python workers", "number of output rows". In a previous attempt I had also instrumented the time for UDF execution, although there are some subtle points there, and I may need to go back to testing that at a later stage. It definitely would be good to know if people using PySpark and Python UDFs find this proposed improvement useful. I see the proposed additional instrumentation as complementary to the Python/Pandas UDF Profiler introduced in Spark 3.3. Best, Luca From: Abdeali Kothari <abdealikoth...@gmail.com> Sent: Friday, August 26, 2022 15:59 To: Luca Canali <luca.can...@cern.ch> Cc: Russell Jurney <russell.jur...@gmail.com>; Gourav Sengupta <gourav.sengu...@gmail.com>; Sean Owen <sro...@gmail.com>; Takuya UESHIN <ues...@happy-camper.st>; user <user@spark.apache.org>; Subash Prabanantham <subashpraba...@gmail.com> Subject: Re: Profiling PySpark Pandas UDF Hi Luca, I see you pushed some code to the PR 3 hrs ago. That's awesome. If I can help out in any way - do let me know I think that's an amazing feature and would be great if it can get into spark On Fri, 26 Aug 2022, 12:41 Luca Canali, <luca.can...@cern.ch <mailto:luca.can...@cern.ch> > wrote: @Abdeali as for “lightweight profiling”, there is some work in progress on instrumenting Python UDFs with Spark metrics, see https://issues.apache.org/jira/browse/SPARK-34265 However it is a bit stuck at the moment, and needs to be revived I believe. Best, Luca From: Abdeali Kothari <abdealikoth...@gmail.com <mailto:abdealikoth...@gmail.com> > Sent: Friday, August 26, 2022 06:36 To: Subash Prabanantham <subashpraba...@gmail.com <mailto:subashpraba...@gmail.com> > Cc: Russell Jurney <russell.jur...@gmail.com <mailto:russell.jur...@gmail.com> >; Gourav Sengupta <gourav.sengu...@gmail.com <mailto:gourav.sengu...@gmail.com> >; Sean Owen <sro...@gmail.com <mailto:sro...@gmail.com> >; Takuya UESHIN <ues...@happy-camper.st <mailto:ues...@happy-camper.st> >; user <user@spark.apache.org <mailto:user@spark.apache.org> > Subject: Re: Profiling PySpark Pandas UDF The python profiler is pretty cool ! Ill try it out to see what could be taking time within the UDF with it. I'm wondering if there is also some lightweight profiling (which does not slow down my processing) for me to get: - how much time the UDF took (like how much time was spent inside the UDF) - how many times the UDF was called I can see the overall time a stage took in the Spark UI - would be cool if I could find the time a UDF takes too On Fri, 26 Aug 2022, 00:25 Subash Prabanantham, <subashpraba...@gmail.com <mailto:subashpraba...@gmail.com> > wrote: Wow, lots of good suggestions. I didn’t know about the profiler either. Great suggestion @Takuya. Thanks, Subash On Thu, 25 Aug 2022 at 19:30, Russell Jurney <russell.jur...@gmail.com <mailto:russell.jur...@gmail.com> > wrote: YOU know what you're talking about and aren't hacking a solution. You are my new friend :) Thank you, this is incredibly helpful! Thanks, Russell Jurney <http://twitter.com/rjurney> @rjurney <mailto:russell.jur...@gmail.com> russell.jur...@gmail.com <http://linkedin.com/in/russelljurney> LI <http://facebook.com/jurney> FB <http://datasyndrome.com> datasyndrome.com On Thu, Aug 25, 2022 at 10:52 AM Takuya UESHIN <ues...@happy-camper.st <mailto:ues...@happy-camper.st> > wrote: Hi Subash, Have you tried the Python/Pandas UDF Profiler introduced in Spark 3.3? - https://spark.apache.org/docs/latest/api/python/development/debugging.html#python-pandas-udf Hope it can help you. Thanks. On Thu, Aug 25, 2022 at 10:18 AM Russell Jurney <russell.jur...@gmail.com <mailto:russell.jur...@gmail.com> > wrote: Subash, I’m here to help :) I started a test script to demonstrate a solution last night but got a cold and haven’t finished it. Give me another day and I’ll get it to you. My suggestion is that you run PySpark locally in pytest with a fixture to generate and yield your SparckContext and SparkSession and the. Write tests that load some test data, perform some count operation and checkpoint to ensure that data is loaded, start a timer, run your UDF on the DataFrame, checkpoint again or write some output to disk to make sure it finishes and then stop the timer and compute how long it takes. I’ll show you some code, I have to do this for Graphlet AI’s RTL utils and other tools to figure out how much overhead there is using Pandera and Spark together to validate data: https://github.com/Graphlet-AI/graphlet I’ll respond by tomorrow evening with code in a fist! We’ll see if it gets consistent, measurable and valid results! :) Russell Jurney On Thu, Aug 25, 2022 at 10:00 AM Sean Owen <sro...@gmail.com <mailto:sro...@gmail.com> > wrote: It's important to realize that while pandas UDFs and pandas on Spark are both related to pandas, they are not themselves directly related. The first lets you use pandas within Spark, the second lets you use pandas on Spark. Hard to say with this info but you want to look at whether you are doing something expensive in each UDF call and consider amortizing it with the scalar iterator UDF pattern. Maybe. A pandas UDF is not spark code itself so no there is no tool in spark to profile it. Conversely any approach to profiling pandas or python would work here . On Thu, Aug 25, 2022, 11:22 AM Gourav Sengupta <gourav.sengu...@gmail.com <mailto:gourav.sengu...@gmail.com> > wrote: Hi, May be I am jumping to conclusions and making stupid guesses, but have you tried koalas now that it is natively integrated with pyspark?? Regards Gourav On Thu, 25 Aug 2022, 11:07 Subash Prabanantham, <subashpraba...@gmail.com <mailto:subashpraba...@gmail.com> > wrote: Hi All, I was wondering if we have any best practices on using pandas UDF ? Profiling UDF is not an easy task and our case requires some drilling down on the logic of the function. Our use case: We are using func(Dataframe) => Dataframe as interface to use Pandas UDF, while running locally only the function, it runs faster but when executed in Spark environment - the processing time is more than expected. We have one column where the value is large (BinaryType -> 600KB), wondering whether this could make the Arrow computation slower ? Is there any profiling or best way to debug the cost incurred using pandas UDF ? Thanks, Subash -- Thanks, Russell Jurney <http://twitter.com/rjurney> @rjurney <mailto:russell.jur...@gmail.com> russell.jur...@gmail.com <http://linkedin.com/in/russelljurney> LI <http://facebook.com/jurney> FB <http://datasyndrome.com> datasyndrome.com -- Takuya UESHIN
