Hi Assaf, Thanks for your suggestion.
I also found one other improvement which is to iteratively convert Pandas DFs to RDDs and take a union of those(similar to dataframes). Basically calling createDataFrame is heavy + checkpointing of DataFrames is a brand new feature. Instead create a huge union of RDDs and finally apply createDataFrame in the end. Thanks and Regards, Saatvik On Wed, Jun 21, 2017 at 2:03 AM, Mendelson, Assaf <[email protected]> wrote: > If you do an action, most intermediate calculations would be gone for the > next iteration. > > What I would do is persist every iteration, then after some (say 5) I > would write to disk and reload. At that point you should call unpersist to > free the memory as it is no longer relevant. > > > > Thanks, > > Assaf. > > > > *From:* Saatvik Shah [mailto:[email protected]] > *Sent:* Tuesday, June 20, 2017 8:50 PM > *To:* Mendelson, Assaf > *Cc:* [email protected] > *Subject:* Re: Merging multiple Pandas dataframes > > > > Hi Assaf, > > Thanks for the suggestion on checkpointing - I'll need to read up more on > that. > > My current implementation seems to be crashing with a GC memory limit > exceeded error if Im keeping multiple persist calls for a large number of > files. > > > > Thus, I was also thinking about the constant calls to persist. Since all > my actions are Spark transformations(union of large number of Spark > Dataframes from Pandas dataframes), this entire process of building a large > Spark dataframe is essentially a huge transformation. Is it necessary to > call persist between unions? Shouldnt I instead wait for all the unions to > complete and call persist finally? > > > > > > On Tue, Jun 20, 2017 at 2:52 AM, Mendelson, Assaf <[email protected]> > wrote: > > Note that depending on the number of iterations, the query plan for the > dataframe can become long and this can cause slowdowns (or even crashes). > A possible solution would be to checkpoint (or simply save and reload the > dataframe) every once in a while. When reloading from disk, the newly > loaded dataframe's lineage is just the disk... > > Thanks, > Assaf. > > > -----Original Message----- > From: saatvikshah1994 [mailto:[email protected]] > Sent: Tuesday, June 20, 2017 2:22 AM > To: [email protected] > Subject: Merging multiple Pandas dataframes > > Hi, > > I am iteratively receiving a file which can only be opened as a Pandas > dataframe. For the first such file I receive, I am converting this to a > Spark dataframe using the 'createDataframe' utility function. The next file > onward, I am converting it and union'ing it into the first Spark > dataframe(the schema always stays the same). After each union, I am > persisting it in memory(MEMORY_AND_DISK_ONLY level). After I have converted > all such files to a single spark dataframe I am coalescing it. Following > some tips from this Stack Overflow > post(https://stackoverflow.com/questions/39381183/ > managing-spark-partitions-after-dataframe-unions). > > Any suggestions for optimizing this process further? > > > > -- > View this message in context: http://apache-spark-user-list. > 1001560.n3.nabble.com/Merging-multiple-Pandas-dataframes-tp28770.html > Sent from the Apache Spark User List mailing list archive at Nabble.com. > > --------------------------------------------------------------------- > To unsubscribe e-mail: [email protected] > > > > > -- > > *Saatvik Shah,* > > *1st Year,* > > *Masters in the School of Computer Science,* > > *Carnegie Mellon University* > > *https://saatvikshah1994.github.io/ <https://saatvikshah1994.github.io/>* > -- *Saatvik Shah,* *1st Year,* *Masters in the School of Computer Science,* *Carnegie Mellon University* *https://saatvikshah1994.github.io/ <https://saatvikshah1994.github.io/>*
