For those interested, after digging further, I was able to consistently reproduce the issue with a synthetic dataset. My findings are documented here:
https://gist.github.com/igozali/d327a85646abe7ab10c2ae479bed431f -- Regards, Ivan Gozali Lecida Email: i...@lecida.com On Wed, Jan 18, 2017 at 12:14 PM, Ivan Gozali <i...@lecida.com> wrote: > Hello, > > I have a use case that seems relatively simple to solve using Spark, but > can't seem to figure out a sure way to do this. > > I have a dataset which contains time series data for various users. All > I'm looking to do is: > > - partition this dataset by user ID > - sort the time series data for each user which by then should > supposedly be contained within individual partitions, > - write each partition to a single CSV file. In the end I'd like to > end up with 1 CSV file per user ID. > > I tried using the following code snippet, but ended up getting surprising > results. I do end up with 1 csv file user ID and most of the users' time > series data were indeed sorted, but those of a small fraction of users were > unsorted. > > # repr(ds) = DataFrame[userId: string, timestamp: string, c1: float, c2: > float, c3: float, ...] > ds = load_dataset(user_dataset_path) > ds.repartition("userId").sortWithinPartitions(" > timestamp").write.partitionBy("userId").option("header", > "true").csv(output_path) > > I'm unclear as to why this could happen, and I'm not entirely sure how to > do this. I'm also not sure if this is potentially a bug in Spark. > > I'm using Spark 2.0.2 with Python 2.7.12. Any advice would be very much > appreciated! > > -- > Regards, > > > Ivan Gozali >