On Fri, Apr 27, 2018 at 12:34 PM Kenneth Knowles <k...@google.com> wrote:
> On Fri, Apr 27, 2018 at 12:18 PM Thomas Weise <t...@apache.org> wrote: >> The ability to specify with URN and implement custom transforms is also important. Such transforms may not qualify for inclusion in Beam for a variety of reasons (only relevant for a specific environment or use case, dependencies/licensing, ...). > They don't need to be included in Beam - by design, a third party library transform can specify its own URN and Payload to be put in the proto representation. I'm not sure of the state of the code here, but I think the current path is a shared dep on runners-core-construction and some ServiceLoader shenanigans. Shading may be in place that breaks this. Agreed, however I was saying that this particular record-like coder probably would make sense as a Beam standard coder rather than having every third-party define their own (or find some other shared location). >> For my specific experiment, I prefer the custom URN over trying to bend the implementation to mimic an SDF based KafkaIO that it wouldn't (and doesn't need to) be semantically equivalent to. At this point Beam doesn't have the spec and implementation for said KafkaIO, but it would be great to see an example how it would look like. Following a Beam spec would absolutely make sense if the custom implementation is purely for optimization or similar purpose. >> I wanted to circle back to the coder related question. I see that we now have a proto definition for the standard transforms and coders, which is really nice: https://github.com/apache/beam/blob/42fac771814b119c162d40e9300f5a0d3afe0f48/model/pipeline/src/main/proto/beam_runner_api.proto#L521 >> This enables interoperability between languages with some standard types (KV, ITERABLE etc.), but for a structure like KafkaRecord a custom coder would be required, implemented in both Java and Python. Any thoughts on providing a generic tuple/record coder as part of the spec? >> Thanks, >> Thomas >> On Fri, Apr 27, 2018 at 8:53 AM, Lukasz Cwik <lc...@google.com> wrote: >>> On Thu, Apr 26, 2018 at 8:38 PM Chamikara Jayalath >>> <chamik...@google.com> wrote: >>>> On Thu, Apr 26, 2018 at 5:59 PM Eugene Kirpichov <kirpic...@google.com> wrote: >>>>> I agree with Thomas' sentiment that cross-language IO is very important because of how much work it takes to produce a mature connector implementation in a language. Looking at implementations of BigQueryIO, PubSubIO, KafkaIO, FileIO in Java, only a very daring soul would be tempted to reimplement them entirely in Python and Go. >>>>> I'm imagining pretty much what Kenn is describing: a pipeline would specify some transforms by URN + payload, and rely on the runner to do whatever it takes to run this - either by expanding it into a Beam implementation of this transform that the runner chooses to use (could be in the same language or in a different language; either way, the runner would indeed need to invoke the respective SDK to expand it given the parameters), or by doing something entirely runner-specific (e.g. using the built-in Flink Kafka connector). >>>>> I don't see a reason to require that there *must* exist a Beam implementation of this transform. There only, ideally, must be a runner- and language-agnostic spec for the URN and payload; of course, then the transform is only as portable as the set of runners that implement this URN. >>>> For a transform in general it's true that we don't need a Beam implementation, but more specifically for IOs I think there are many benefits to having the implementation in Beam. For example, >>>> IO connector will offer same behavior and feature set across various runners/SDKs. >>>> Beam community will be able to view/modify/improve the IO connector. >>>> existing IO connectors will serve as examples for users who wish to develop new IO connectors >>> More runners will be able to execute the users pipeline. >>>>> I actually really like the idea that the transform can be implemented in a completely runner-specific way without a Beam expansion to back it up - it would let us unblock a lot of the work earlier than full-blown cross-language IO is delivered or even than SDFs work in all languages/runners. >>>> If there are existing established connectors (for example, Kafka for Flink in this case) I agree. But for anybody developing a new IO connector, I think we should encourage developing that in Beam (in some SDK) given that the connector will be available to all runners (and to all SDKs once we have cross-language transforms). >>>> Thanks, >>>> Cham >>>>> On Wed, Apr 25, 2018 at 10:02 PM Kenneth Knowles <k...@google.com> wrote: >>>>>> It doesn't have to be 1:1 swapping KafkaIO for a Flink Kafka connector, right? I was imagining: Python SDK submits pipeline with a KafkaIO (with URN + payload) maybe bogus contents. It is replaced with a small Flink subgraph, including the native Flink Kafka connector and some compensating transfoms to match the required semantics. To me, this is preferable to making single-runner transform URNs, since that breaks runner portability by definition. >>>>>> Kenn >>>>>> On Wed, Apr 25, 2018 at 7:40 PM Chamikara Jayalath < chamik...@google.com> wrote: >>>>>>> On Wed, Apr 25, 2018 at 6:57 PM Reuven Lax <re...@google.com> wrote: >>>>>>>> On Wed, Apr 25, 2018 at 6:51 PM Kenneth Knowles <k...@google.com> wrote: >>>>>>>>> The premise of URN + payload is that you can establish a spec. A native override still needs to meet the spec - it may still require some compensating code. Worrying about weird differences between runners seems more about worrying that an adequate spec cannot be determined. >>>>>>>> My point exactly. a SDF-based KafkaIO can run in the middle of a pipeline. E.g. we could have TextIO producing a list of topics, and the SDF KafkaIO run after that on this dynamic (not known until runtime) list of topics. If the native Flink source doesn't work this way, then it doesn't share the same spec and should have a different URN. >>>>>>> Agree that if they cannot share the same spec, SDF and native transforms warrant different URNs. Native Kafka might be able to support a PCollection of topics/partitions as an input though by utilizing underlying native Flink Kafka connector as a library. On the other hand, we might decide to expand SDF based ParDos into to other transforms before a runner gets a chance to override in which case this kind of replacements will not be possible. >>>>>>> Thanks, >>>>>>> Cham >>>>>>>>> Runners will already invoke the SDF differently, so users treating every detail of some implementation as the spec are doomed. >>>>>>>>> Kenn >>>>>>>>> On Wed, Apr 25, 2018, 17:04 Reuven Lax <re...@google.com> wrote: >>>>>>>>>> On Tue, Apr 24, 2018 at 5:52 PM Chamikara Jayalath < chamik...@google.com> wrote: >>>>>>>>>>> On Tue, Apr 24, 2018 at 3:44 PM Henning Rohde < hero...@google.com> wrote: >>>>>>>>>>>> > Note that a KafkaDoFn still needs to be provided, but could be a DoFn that >>>>>>>>>>>> > fails loudly if it's actually called in the short term rather than a full >>>>>>>>>>>> > Python implementation. >>>>>>>>>>>> For configurable runner-native IO, for now, I think it is reasonable to use a URN + special data payload directly without a KafkaDoFn -- assuming it's a portable pipeline. That's what we do in Go for PubSub-on-Dataflow and something similar would work for Kafka-on-Flink as well. I agree that non-native alternative implementation is desirable, but if one is not present we should IMO rather fail at job submission instead of at runtime. I could imagine connectors intrinsic to an execution engine where non-native implementations are not possible. >>>>>>>>>>> I think, in this case, KafkaDoFn can be a SDF that would expand similar to any other SDF by default (initial splitting, GBK, and a map-task equivalent, for example) but a runner (Flink in this case) will be free to override it with an runner-native implementation if desired. I assume runner will have a chance to perform this override before the SDF expansion (which is not fully designed yet). Providing a separate source/sink transforms for Flink native Kafka will be an option as well, but that will be less desirable from a Python user API perspective. >>>>>>>>>> Are we sure that the internal SDF will provide the same functionality as the native one? What if the Kafka SDF is in the middle of a pipeline - can Flink support that? Having a separate transform for the Flink native source might be a better user experience than having one that changes its behavior in strange ways depending on the runner. >>>>>>>>>>>> On Tue, Apr 24, 2018 at 3:09 PM Robert Bradshaw < rober...@google.com> wrote: >>>>>>>>>>>>> On Tue, Apr 24, 2018 at 1:14 PM Thomas Weise <t...@apache.org> wrote: >>>>>>>>>>>>> > Hi Cham, >>>>>>>>>>>>> > Thanks for the feedback! >>>>>>>>>>>>> > I should have probably clarified that my POC and questions aren't >>>>>>>>>>>>> specific to Kafka as source, but pretty much any other source/sink that we >>>>>>>>>>>>> internally use as well. We have existing Flink pipelines that are written >>>>>>>>>>>>> in Java and we want to use the same connectors with the Python SDK on top >>>>>>>>>>>>> of the already operationalized Flink stack. Therefore, portability isn't a >>>>>>>>>>>>> concern as much as the ability to integrate is. >>>>>>>>>>> Thanks for the clarification. Agree that providing runner-native implementations of established source/sinks can be can be desirable in some cases. >>>>>>>>>>>>> > --> >>>>>>>>>>>>> > On Tue, Apr 24, 2018 at 12:00 PM, Chamikara Jayalath >>>>>>>>>>>>> > <chamik...@google.com> >>>>>>>>>>>>> wrote: >>>>>>>>>>>>> >> Hi Thomas, >>>>>>>>>>>>> >> Seems like we are working on similar (partially) things :). >>>>>>>>>>>>> >> On Tue, Apr 24, 2018 at 9:03 AM Thomas Weise < t...@apache.org> wrote: >>>>>>>>>>>>> >>> I'm working on a mini POC to enable Kafka as custom streaming source >>>>>>>>>>>>> for a Python pipeline executing on the (in-progress) portable Flink runner. >>>>>>>>>>>>> >>> We eventually want to use the same native Flink connectors for sources >>>>>>>>>>>>> and sinks that we also use in other Flink jobs. >>>>>>>>>>>>> >> Could you clarify what you mean by same Flink connector ? Do you mean >>>>>>>>>>>>> that Beam-based and non-Beam-based versions of Flink will use the same >>>>>>>>>>>>> Kafka connector implementation ? >>>>>>>>>>>>> > The native Flink sources as shown in the example below, not the Beam >>>>>>>>>>>>> KafkaIO or other Beam sources. >>>>>>>>>>>>> >>> I got a simple example to work with the FlinkKafkaConsumer010 reading >>>>>>>>>>>>> from Kafka and a Python lambda logging the value. The code is here: https://github.com/tweise/beam/commit/79b682eb4b83f5b9e80f295464ebf3499edb1df9 >>>>>>>>>>>>> >>> I'm looking for feedback/opinions on the following items in particular: >>>>>>>>>>>>> >>> * Enabling custom translation on the Flink portable runner (custom >>>>>>>>>>>>> translator could be loaded with ServiceLoader, additional translations >>>>>>>>>>>>> could also be specified as job server configuration, pipeline option, ...) >>>>>>>>>>>>> >>> * For the Python side, is what's shown in the commit the recommended >>>>>>>>>>>>> way to define a custom transform (it would eventually live in a reusable >>>>>>>>>>>>> custom module that pipeline authors can import)? Also, the example does not >>>>>>>>>>>>> have the configuration part covered yet.. >>>>>>>>>>>>> >> The only standard unbounded source API offered by Python SDK is the >>>>>>>>>>>>> Splittable DoFn API. This is the part I'm working on. I'm trying to add a >>>>>>>>>>>>> Kafka connector for Beam Python SDK using SDF API. JIRA is >>>>>>>>>>>>> https://issues.apache.org/jira/browse/BEAM-3788. I'm currently comparing >>>>>>>>>>>>> different Kafka Python client libraries. Will share more information on >>>>>>>>>>>>> this soon. >>>>>>>>>>>>> >> I understand this might not be possible in all cases and we might want >>>>>>>>>>>>> to consider adding a native source/sink implementations. But this will >>>>>>>>>>>>> result in the implementation being runner-specific (each runner will have >>>>>>>>>>>>> to have it's own source/sink implementation). So I think we should try to >>>>>>>>>>>>> add connector implementations to Beam using the standard API whenever >>>>>>>>>>>>> possible. We also have plans to implement support for cross SDK transforms >>>>>>>>>>>>> in the future (so that we can utilize Java implementation from Python for >>>>>>>>>>>>> example) but we are not there yet and we might still want to implement a >>>>>>>>>>>>> connector for a given SDK if there's good client library support. >>>>>>>>>>>>> > It is great that the Python SDK will have connectors that are written in >>>>>>>>>>>>> Python in the future, but I think it is equally if not more important to be >>>>>>>>>>>>> able to use at least the Java Beam connectors with Python SDK (and any >>>>>>>>>>>>> other non-Java SDK). Especially in a fully managed environment it should be >>>>>>>>>>>>> possible to offer this to users in a way that is largely transparent. It >>>>>>>>>>>>> takes significant time and effort to mature connectors and I'm not sure it >>>>>>>>>>>>> is realistic to repeat that for all external systems in multiple languages. >>>>>>>>>>>>> Or, to put it in another way, it is likely that instead of one over time >>>>>>>>>>>>> rock solid connector per external system there will be multiple less mature >>>>>>>>>>>>> implementations. That's also the reason we internally want to use the Flink >>>>>>>>>>>>> native connectors - we know what they can and cannot do and want to >>>>>>>>>>>>> leverage the existing investment. >>>>>>>>>>>>> There are two related issues here: how to specify transforms (such as >>>>>>>>>>>>> sources) in a language-independent manner, and how specific runners can >>>>>>>>>>>>> recognize and run them, but URNs solve both. For this we use URNs: the >>>>>>>>>>>>> composite ReadFromKafka PTransform (that consists of a Impulse + >>>>>>>>>>>>> SDF(KafkaDoFn)) to encodes to a URN with an attached payload that fully >>>>>>>>>>>>> specifies this read. (The KafkaDoFn could similarly have a URN and >>>>>>>>>>>>> payload.) A runner that understands these URNs is free to make any >>>>>>>>>>>>> (semantically-equivalent) substitutions it wants for this transform. >>>>>>>>>>>>> Note that a KafkaDoFn still needs to be provided, but could be a DoFn that >>>>>>>>>>>>> fails loudly if it's actually called in the short term rather than a full >>>>>>>>>>>>> Python implementation. Eventually, we would like to be able to call out to >>>>>>>>>>>>> another SDK to expand full transforms (e.g. more complicated ones like >>>>>>>>>>>>> BigQueryIO). >>>>>>>>>>>>> >>> * Cross-language coders: In this example the Kafka source only >>>>>>>>>>>>> considers the message value and uses the byte coder that both sides >>>>>>>>>>>>> understand. If I wanted to pass on the key and possibly other metadata to >>>>>>>>>>>>> the Python transform (similar to KafkaRecord from Java KafkaIO), then a >>>>>>>>>>>>> specific coder is needed. Such coder could be written using protobuf, Avro >>>>>>>>>>>>> etc, but it would also need to be registered. >>>>>>>>>>>>> >> I think this requirement goes away if we implement Kafka in Python SDK. >>>>>>>>>>>>> > Wouldn't this be needed for any cross language pipeline? Or rather any >>>>>>>>>>>>> that isn't only using PCollection<byte[]>? Is there a language agnostic >>>>>>>>>>>>> encoding for KV<?,?>, for example? >>>>>>>>>>>>> Yes, Coders are also specified by URN (+components and/or payload), and >>>>>>>>>>>>> there are a couple of standard ones, including KV. See https://github.com/apache/beam/blob/master/model/pipeline/src/main/resources/org/apache/beam/model/common_urns.md >>>>>>>>>>>>> This is not a closed set. >>>>>>>>>>>>> - Robert
