kennknowles commented on pull request #13006: URL: https://github.com/apache/beam/pull/13006#issuecomment-703916817
> #12519 (comment) > > > > > @kennknowles > > > > > > I did a first pass. I think big picture there are two things: > > > > > > > > > > > > 1. Without unfolding all the code not in the code review, I am not already familiar with the expected level of SDF support of all the runners. Some of them seem to previously reject non-SDF reads but now could end up having non-SDF reads. > > > > > > > > > > > > > > > [BEAM-10670](https://issues.apache.org/jira/browse/BEAM-10670) is about migrating all runners to use SDF first and only SDF. All runners that supported BoundedSource/UnboundedSource support SDF to the level necessary to execute the BoundedSource/UnboundedSource as SDF wrappers. The only ones that don't are Spark (underway in #12603) and Dataflow runner v1 which would have a degraded experience in batch and would break pipeline update in streaming. This was a focus of mine over the past ~6 weeks. > > > > > > > > > > > > Do they have equivalent performance? My architectural idea is that we can have expansion that looks like: > > > > > > > > > Nexmark did show improvements ([#12519 (comment)](https://github.com/apache/beam/pull/12519#issuecomment-680278642)) > > > > ``` > > > > Read(BoundedSource) { > > > > Impulse -> SDF > > > > } > > > > ``` > > > > > > > > > > > > preserving all the bounded/unbounded source data but having a portable expansion. "Portable" meaning that all non-primitives (like Read) are expanded to primitives (SDF). But there is no reason a runner has to support SDF at all, if they can recognize the composite and implement it directly. > > > > So my basic idea would be to have all such runners continue to use their existing execution paths, unless the runner authors/maintainers believe whatever SDF support they have is good enough. FWIW I don't think this is my invention. I thought this was the plan to seamlessly support Read and SDF in the same portable pipeline, with runners picking and choosing what they can execute CC @xinyuiscool @iemejia @mxm > > > > > > > > > I brought up the migration to using SDF to power BoundedSource/UnboundedSource on [dev@](https://lists.apache.org/thread.html/r1ba6fe6ac2bd2b28aa7ef31f0d87ad716fc878f2515085fdbc275333%40%3Cdev.beam.apache.org%3E). If you would like to reconsider the migration path it would make sense to continue this discussion on the original thread. > > > > > > 1. I do not think runners core construction should be manipulated by global flags, but each runner should decide what to do with the flag, if anything. If a runner has a well-used non-portable variant that is not going away any time soon, then it can still use the portable pipeline and just reject any with non-Java UDFs. If a runner only supports SDF in some modes but only supports Read in others, or wants to translate Read specially for efficiency (or any other reason it wants to - it can do whatever it wants) then it can do that. > > > > > > > > > > > > > > > I can do that but it seems like it will lead to possible inconsistency across runners around how this choice is made. Currently CHANGES.md says to get the old behavior use this flag and it would be unfortunate to say "consult the runner documentation as to how to get the old behavior". > > > > > > > > > > > > There should not be behavior differences. The expansion above should have identical PCollection contents whether the runner implements Read or SDF. And the runner should make sure that pipeline maintain their performance by using the best translation they can. Users should not need to be aware of this, except perhaps via debugging interfaces that expose the expansion. > > > > > > > > > The purpose of having the fallback is to give users an opt-out mechanism during the migration should they hit issues. I don't believe our runners have comprehensive testing of pipelines at scale enough to state that the migration will be perfect. > > > > > After a few releases, the intent is that this logic is removed from runners-core and everyone except for Dataflow v1 runner always use the SDF wrapper and we move the necessary logic within runners-core to the Dataflow module deleting the remainder (the PrimitiveBoundedRead override). > > > > > > > > > > > > My position on runners-core and runners-core-construction is that these are just libraries with collections of useful functionality for implementing a runner. Neither one should ever be user-visible. Deleting dead code from either library is nice-to-have cleanup. If any runner wants to keep the code it can be copied into the runner. > > > > > > > > > I disagree. As someone who went through all our runners changing a core piece, dead code would only make Apache Beam harder to maintain. > > > > > > "core" is just a name because I/we didn't have a clever name for it. Both libraries were created for this purpose and really shouldn't be seen as more than a utility library. They are not the definition of any part of the Beam model, just helpers. > > I didn't mean "core" as in the modules with the name "core" in them but the "core" concept of how data is ingested into a pipeline by a runner. OK. I am only referring to the module name. The modules are just piles of functionality that a runner can use as they please. The thing I don't like is having a global user-controlled flag that tweaks runner behavior when the runner may not even be aware. Basically moving a flag from the SDK to the runner is better but still not great. ---------------------------------------------------------------- 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]
