[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: PTransforms and Fusion

Empty pipelines have neither subtransforms or a spec, which is what I don't think is useful -- especially given the only usecase (which is really "nop") would create non-timer loops in the representations. I'd rather have a well-known nop primitive instead. Even now, for the A example, I don't think it's unreasonable to add a (well-known) identity transform inside a normal composite to retain the extrema at either end. It could be ignored at runtime at no cost.

To clarify my support for A1, native transforms would have a spec and would be passed through in the shared code even through they're not primitives.

On Tue, Sep 11, 2018 at 12:56 AM Robert Bradshaw <robertwb@xxxxxxxxxx> wrote:
For (A), it really boils down to the question of what is a legal pipeline. A1 takes the position that all empty transforms must be on a whitelist (which implies B1, unless we make the whitelist extensible, which starts to sound a lot like B3). Presumably if we want to support B2, we cannot remove all empty unknown transforms, just those whose outputs are a subset of the inputs. 

The reason I strongly support A3 is that empty PTransforms are not just noise, they are expressions of user intent, and the pipeline graph should reflect that as faithfully as possible. This is the whole point of composite transforms--one should not be required to expose what is inside (even whether it's empty). Consider, for example, an A, B -> C transform that mixes A and B in proportions to produce C. In the degenerate case where we want 100% for A or 100% from B, it's reasonable to implement this by just returning A or B directly. But when, say, visualizing the pipeline graph, I don't think it's desirable to have the discontinuity of the composite transform suddenly disappearing when the mixing parameter is at either extreme. 

If a runner cannot handle these empty pipelines (as is the case for those relying on the current Java libraries) it is an easy matter for it to drop them, but that doesn't mean we should withhold this information (by making it illegal and dropping it in every SDK) from a runner (or any other tool) that would want to see this information. 

- Robert

On Tue, Sep 11, 2018 at 4:20 AM Henning Rohde <herohde@xxxxxxxxxx> wrote:
For A, I am in favor of A1 and A2 as well. It is then up to each SDK to not generate "empty" transforms in the proto representation as we avoid noise as mentioned. The shared Java libraries are also optional and we should not assume every runner will use them. I'm not convinced empty transforms would have value for pipeline structure over what can be accomplished with normal composites. I suspect empty transforms, such as A, B -> B, B, will just be confusion generators.

For B, I favor B2 for the reasons Thomas mentions. I also agree with the -1 for B1.

On Mon, Sep 10, 2018 at 2:51 PM Thomas Weise <thw@xxxxxxxxxx> wrote:
For B, note the prior discussion [1].

B1 and B2 cannot be supported at the same time.

Native transforms will almost always be customizations. Users do not create customizations without reason. They would start with what is there and dig deeper only when needed. Right now there are no streaming connectors in the Python SDK - should the user not use the SDK? Or is it better (now and in general) to have the option of a custom connector, even when it is not portable?

Regarding portability, IMO it should be up to the user to decide how much of it is necessary/important. The IO requirements are normally dictated by the infrastructure. If it has Kafka, Kinesis or any other source (including those that Beam might never have a connector for), the user needs to be able to integrate it.

Overall extensibility is important and will help users adopt Beam. This has come up in a few other areas (think Docker environments). I think we need to provide the flexibility and enable, not prevent alternatives and therefore -1 for B1 (unsurprisingly :). 

On Mon, Sep 10, 2018 at 10:14 AM Robert Bradshaw <robertwb@xxxxxxxxxx> wrote:
A) I think it's a bug to not handle empty PTransforms (which are useful at pipeline construction, and may still have meaning in terms of pipeline structure, e.g. for visualization). Note that such transforms, if truly composite, can't output any PCollections that do not appear in their inputs (which is how we distinguish them from primitives in Python). Thus I'm in favor of A3, and as a stopgap we can drop these transforms as part of/just before decoding in the Java libraries (rather than in the SDKs during encoding as in A2). 

B) I'm also for B1 or B2. 

On Mon, Sep 10, 2018 at 3:58 PM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
> A) What should we do with these "empty" PTransforms?

We can't translate them, so dropping them seems the most reasonable
choice. Should we throw an error/warning to make the user aware of this?
Otherwise might be unexpected for the user.

>> A3) Handle the "empty" PTransform case within all of the shared libraries.

What can we do at this point other than dropping them?

> B) What should we do with "native" PTransforms?

I support B1 and B2 as well. Non-portable PTransforms should be
discouraged in the long run. However, the available PTransforms are not
even consistent across the different SDKs yet (e.g. no streaming
connectors in Python), so we should continue to provide a way to utilize
the "native" transforms of a Runner.


On 07.09.18 19:15, Lukasz Cwik wrote:
> A primitive transform is a PTransform that has been chosen to have no
> default implementation in terms of other PTransforms. A primitive
> transform therefore must be implemented directly by a pipeline runner in
> terms of pipeline-runner-specific concepts. An initial list of primitive
> PTransforms were defined in [2] and has since been updated in [3].
> As part of the portability effort, libraries that are intended to be
> shared across multiple runners are being developed to support their
> migration to a portable execution model. One of these is responsible for
> fusing multiple primitive PTransforms together into a pipeline runner
> specific concept. This library made the choice that a primitive
> PTransform is a PTransform that doesn't contain any other PTransforms.
> Unfortunately, while Ryan was attempting to enable testing of validates
> runner tests for Flink using the new portability libraries, he ran into
> an issue where the Apache Beam Java SDK allows for a person to construct
> a PTransform that has zero sub PTransforms and also isn't one of the
> defined Apache Beam primitives. In this case the PTransform was trivial
> as it was not applying any additional transforms to input PCollection
> and just returning it. This caused an issue within the portability
> libraries since they couldn't handle this structure.
> To solve this issue, I had proposed that we modify the portability
> library that does fusion to use a whitelist of primitives preventing the
> issue from happening. This solved the problem but caused an issue for
> Thomas as he was relying on this behaviour of PTransforms with zero sub
> transforms being primitives. Thomas has a use-case where he wants to
> expose the internal Flink Kafka and Kinesis connectors and to build
> Apache Beam pipelines that use the Flink native sources/sinks. I'll call
> these "native" PTransforms, since they aren't part of the Apache Beam
> model and are runner specific.
> This brings up two topics:
> A) What should we do with these "empty" PTransforms?
> B) What should we do with "native" PTransforms?
> The typical flow of a pipeline representation for a portable pipeline is:
> language specific representation -> proto representation -> job service
> -> shared libraries that simplify/replace the proto representation with
> a simplified version (e.g. fusion) -> runner specific conversion to
> native runner concepts (e.g. GBK -> runner implementation of GBK)
> ------------------
> A) What should we do with these "empty" PTransforms?
> To give a little more detail, these transforms typically can happen if
> people have conditional logic such as loops that would perform an
> expansion but do nothing if the condition is immediately unsatisfied. So
> allowing for PTransforms that are empty is useful when building a pipeline.
> What should we do:
> A1) Stick with the whitelist of primitive PTransforms.
> A2) When converting the pipeline from language specific representation
> into the proto representation, drop any "empty" PTransforms. This means
> that the pipeline representation that is sent to the runner doesn't
> contain the offending type of PTransform and the shared libraries
> wouldn't have to change.
> A3) Handle the "empty" PTransform case within all of the shared libraries.
> I like doing both A1 and A2. A1 since it helps simplify the shared
> libraries since we know the whole list of primitives we need to
> understand and A2 because it removes noise within the pipeline shape
> from its representation.
> ------------------
> B) What should we do with "native" PTransforms?
> Some approaches that we could take as a community:
> B1) Prevent the usage of "native" PTransforms within Apache Beam since
> they hurt portability of pipelines across runners. This can be done by
> specifically using whitelists of allowed primitive PTransforms in the
> shared libraries and explicitly not allowing for shared libraries to
> have extension points customizing this.
> B2) We embrace that closed systems internal to companies will want to
> use their own extensions and enable support for "native" PTransforms but
> actively discourage "native" PTransforms in the open ecosystem.
> B3) We embrace and allow for "native" PTransforms in the open ecosystem.
> "native" PTransforms are useful in closed systems since they allow
> companies to solve certain scenarios which would not be practical to
> expose the Apache Beam community. It does take more work for the
> community to support these types of extensions. To my knowledge, Google
> is likely to want to do something similar to handle internal use cases
> similar to what Thomas is trying to do.
> I'm for either B1 or B2 since the risk of embracing and allowing for
> "native" PTransforms in the open ecosystem is likely to fragment the
> project and also is counter to what portability is really about.
> 1: https://github.com/apache/beam/pull/6328
> 2:
> https://docs.google.com/document/d/1bao-5B6uBuf-kwH1meenAuXXS0c9cBQ1B2J59I3FiyI/edit#heading=h.tt55lhd3k6by
> 3:
> https://github.com/apache/beam/blob/6df2ef3ec9c835097e79b4441ce47ff09a458894/model/pipeline/src/main/proto/beam_runner_api.proto#L180