Wow this is fantastic! I tested it out and it worked great for my runner. I am also excited for this change now and will eagerly set `cloudpickle` as the default pickler for our code.
FWIW - I noticed that the DataFlow Options documentation[1] for setting the pickling library and the Beam documentation [2] for setting the pickling library differ slightly. The DataFlow documentation mentions having to set `pickler.set_library(pickler.USE_CLOUDPICKLE)` while the Beam documentation doesn't say anything about that. It turned out unnecessary for my runner - not sure if it's just a DataFlow runner specific requirement, but just wanted to point it out. [1] https://cloud.google.com/dataflow/docs/reference/pipeline-options#pythonyaml [2] https://beam.apache.org/documentation/sdks/python-pipeline-dependencies/#pickling-and-managing-the-main-session On Wed, Apr 30, 2025 at 2:38 AM Robert Bradshaw <rober...@waymo.com> wrote: > On Tue, Apr 29, 2025 at 7:51 PM Joey Tran <joey.t...@schrodinger.com> > wrote: > > > > Does cloudpickle make --save_main_session unnecessary? As in, will more > transforms defined in __main__ "just work"? > > Yes. Or at least it "just works" much more often. (There may still be > corner cases, but I haven't run into them...) > > I, for one, am excited to see this change. Thanks, Claude, for taking > the lead on this. > > > If so, I can see why that's worthwhile. I've had a _ton_ of issues with > this, especially with new users of beam at my company. Explaining main > session and why random things throw unpickling errors or why their > transform is throwing Name errors has been a very painful experience, > especially since it usually happens with users first experiences > > > > On Tue, Apr 29, 2025, 6:14 PM Valentyn Tymofieiev via dev < > dev@beam.apache.org> wrote: > >> > >> There are several reasons: > >> - wide adoption in data processing community , see initial discussion: > [1] > >> - expectations on cloudpickle having a larger number of maintainers > and contributors. > >> - new releases of dill had breaking changes[2], which made adoption of > a new version challenging. > >> - cloudpickle is easier to vendor - it is a single file and unlike > dill, does not create side-effects in the global namespace, which might > conflict with any unvendored version. vendoring allows to eliminate a > common failure mode when the pickler library is different at submission and > runtime. > >> - previously, some bugs and feature requests Beam requested in dill > took a long time to be implemented and released. > >> > >> [1] https://lists.apache.org/thread/dvxvclhok0fx48955x6szvw4kotxh87n > >> [2] https://github.com/apache/beam/issues/22893#issuecomment-1502354194 > >> > >> On Mon, Apr 28, 2025 at 4:00 PM Joey Tran <joey.t...@schrodinger.com> > wrote: > >>> > >>> Naive question, but why is beam upgrading to cloudpickle? > >>> > >>> I saw this doc: > >>> > https://docs.google.com/document/d/1G5Q0ckX5sKQRQD1yEkLCPQL7N6B-AL9Cb1p0zlOOfQU/edit?tab=t.0 > >>> > >>> Is the main reason because cloudpickle is more actively maintained? > >>> > >>> > >>> On Mon, Apr 28, 2025 at 6:51 PM Claudius van der Merwe < > claud...@vdmza.com> wrote: > >>>> > >>>> Hi Beam Devs, > >>>> > >>>> > >>>> I am making progress on making cloudpickle the default pickling > library and removing the strict dependency on dill as outlined in > https://s.apache.org/beam-cloudpickle-next-steps. > >>>> > >>>> > >>>> The current plan is to: > >>>> > >>>> > >>>> 1. Make cloudpickle the default library in Beam 2.65.0 release (see > https://github.com/apache/beam/pull/34695). Users will be able to specify > pickle_library='dill' without any additional requirements. There will still > be a hard dependency on dill (blocked by #2) but it is a step in the right > direction. > >>>> > >>>> > >>>> 2. Remove the strict dependency on dill in Beam 2.66.0 release. Dill > is directly used for coder's encoding types in FastPrimitivesCoderImpl > [1][2]. I prefer to submit a fix for this after the branch cut so we have > more time to identify any issues. > >>>> > >>>> > >>>> Coudpickle has some fundamentally different pickling behavior to dill > that is likely to break: > >>>> > >>>> Unittests that rely on globals > >>>> > >>>> This can be fixed by using apache_beam.utils.shared [3] > >>>> > >>>> Closures and dynamic classes that reference unpicklable globals > >>>> > >>>> This can be fixed by defining functions in the top level, and using > functools.partial to bind parameters if necessary > >>>> > >>>> > >>>> [1] > https://github.com/apache/beam/blob/b9fa49a9827dd28349e382f479ebd1a8bbe27d07/sdks/python/apache_beam/coders/coder_impl.py#L529 > >>>> > >>>> [2] > https://github.com/apache/beam/blob/b9fa49a9827dd28349e382f479ebd1a8bbe27d07/sdks/python/apache_beam/coders/coder_impl.py#L595 > >>>> > >>>> [3] > https://github.com/apache/beam/blob/b9fa49a9827dd28349e382f479ebd1a8bbe27d07/sdks/python/apache_beam/internal/cloudpickle_pickler_test.py#L54 > >>>> > >>>> > >>>> I'd appreciate any feedback or concerns. > >>>> > >>>> > >>>> Best, > >>>> > >>>> Claude > >>>> > >>>> >
