lukecwik edited a comment on issue #11406: [BEAM-9748] Move Reparallelize transform to Reshuffle URL: https://github.com/apache/beam/pull/11406#issuecomment-614121831 > The comments inside Reparallelize explain how this transform differs from Reshuffle.viaRandomKey(): it performs dramatically better on Dataflow in case the input PCollection is generated highly sequentially, as in the case of reading several GB of JDBC results. It almost certainly performs somewhat worse if the input PCollection is generated in a well-parallelized way, but I haven't measured that; I haven't measured the former case for non-Dataflow runners either. > > I think it's reasonable to move this to Reshuffle, but rename it to something more clear: maybe Reshuffle.forSequentiallyGeneratedInput()? This was true at some point in time for Dataflow but I don't think we have rerun benchmarks to see how it performs now. Shuffle performance has improved a lot while side input performance has remained relatively flat for the iterable use case.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
