On Thu, Feb 4, 2021 at 3:33 PM Kyle Weaver <kcwea...@google.com> wrote:
> This gets into the distinction of customizing what kind of environment >> one wants to have (which could be generally applicable) vs. an absolute >> designation of a particular environment (e.g. a docker image). > > > For common environment modifications, resource hints are a great idea, > since it's much easier to set an annotation than to build and set a custom > container. The limitation of this approach is we can't handle every > possible modification a user might want to make to their environment. > Custom containers give the user ultimate control over the environment, so > we forfeit a lot of flexibility if we don't provide enough options to use > them. > > Note that what we're running into in part is that "pipeline options" are >> the wrong level of granularity for specifying characteristics of an >> environment, as there is not a single environment to parameterize (or, >> possibly, even one per language). > > > Yes, this is the crux of the problem. We already expose an > environment_config as a pipeline option, so we basically have three choices: > 1. Deprecate pipeline-level environment options altogether. > 2. Find a way to generalize environment options. > 3. Keep and document the status quo (ie users can use custom containers, > but at most only one per language). > I do think it can be useful to specify a custom "top-level" environment. We should probably make it easy to use customized expansion services. > The caller should not need any visibility into the environment(s) that an >> expansion service uses, which is an implementation detail that the >> expansion service is free to change at any time. (In fact, whether it is >> (partially or fully) implemented as an external transform is an >> implementation detail that the end user should not need to care about or >> depend on.) > > > I personally think pattern matching and substitution by runners (maybe >> more sophisticated than regexp on container names) is a reasonable way to >> approach customization of environments. > > > Aren't these ideas contradictory? Pattern matching requires knowledge in > advance of which patterns to match. We'd need to know at least some > information about the environment the expansion service is expected to use > in order to replace it. > The pattern matching is not such that I want to replace the environment for this particular transform, but that /if/ I see a Java environment of a certain type /then/ I want to run it in this way. > For example, suppose I construct a pipeline that uses both Python and Java >> transforms. (I could do this from Go, Java, or Python). If I want to run >> this locally (e.g. on the Python FnAPI runner), I would prefer that the >> python bits be run in-process but would have to shell out (maybe via >> docker, maybe something cheaper) for the java bits. On the other hand, if I >> want to run this same pipeline (ideally, the same model proto, such that we >> don't have runner-dependent construction) on Flink, I might want the java >> bits to be inlined and the Python bits to be in a separate process. On >> Dataflow, both would live in containers. To do this, the Python runner >> would say "hey, I know that Python environment" and just swap it out for >> in-process, and vice versa. (For isolation/other reasons, one may want the >> option to force everything to be docker, but that's more of a "don't make >> substitutions" option than manually providing environment configs.) > > > In this example, wouldn't you normally just rebuild the pipeline? I'm not > sure what the advantage of re-using the same model proto is. > Yes, you'd re-build the pipeline. But if all you change is the --runner flag the model proto produced should not change. (And, sometimes, you may want to stash the proto itself, or pass it to one-of-N runners depending on some other condition, etc.) > It would be helpful for me to have concrete usecases of why a user wants >> to customize the container used by some transform they did not write, which >> could possibly inform the best course(s) of action here. > > > I should have led with this. Someone wanted to mount credentials into the > SDK harness [1]. So in this particular case the user just wants to mount > files into their SDK harness, which is a pretty common use case, so > resource hints are probably a more appropriate solution. > > [1] > https://lists.apache.org/thread.html/r690094f1c9ebc4e1d20f029a21ba8bc846672a65baafd57c4f52cb94%40%3Cuser.beam.apache.org%3E > Ah, that clarifies things. Would it be possible/preferable to pass the credentials as parameters to the transform itself? > > > On Thu, Feb 4, 2021 at 1:51 PM Robert Bradshaw <rober...@google.com> > wrote: > >> On Thu, Feb 4, 2021 at 12:38 PM Kyle Weaver <kcwea...@google.com> wrote: >> >>> So, an external transform is uniquely identified by its URN. An external >>>> transform identified by a URN may refer to an arbitrary composite which may >>>> have sub-transforms that refer to different environments. I think with the >>>> above proposal we'll lose this flexibility. >>>> What we need is a way to override environments (or properties of >>>> environments) that results in the final pipeline proto. Once we modify such >>>> environments in the proto it will be reflected to all transforms that >>>> utilize such environments. >>> >>> >>> As far as I can tell we currently only register a single environment for >>> the entire transform (and it's always the default). Am I missing something? >>> https://github.com/apache/beam/blob/0cfa80fd919d141a2061393ec5c12521c7d7af0b/sdks/java/expansion-service/src/main/java/org/apache/beam/sdk/expansion/service/ExpansionService.java#L447-L449 >>> >>> Anyway, I don't see how sub-transforms require overrides. We should be >>> able to propagate environment options to sub-transforms to achieve the same >>> purpose. >>> >> >> The discussion of resource hints at >> https://lists.apache.org/thread.html/ra40286b66a03a1d9f4086c9e1ecdeb9f299836d2d0361c3e3fe7c382%40%3Cdev.beam.apache.org%3E >> actually may tie into this as well. I would assume a localised request for, >> say, high memory should be propagated down to cross-language pipelines. It >> is possible that other customizations (such as making sure specific >> dependencies are available, or filesystems mounted) would fit here too. >> >> This gets into the distinction of customizing what kind of environment >> one wants to have (which could be generally applicable) vs. an absolute >> designation of a particular environment (e.g. a docker image). >> >> Note that what we're running into in part is that "pipeline options" are >> the wrong level of granularity for specifying characteristics of an >> environment, as there is not a single environment to parameterize (or, >> possibly, even one per language). If I call >> ExpansionRequset(MyFancyTransform,environment_config=docker_path) >> and MyFancyTransform is composed of two environments, to which >> does docker_path apply? What about PTransforms that use ExternalTransforms >> under the hood (e.g does some pre-processing and then calls SQL, or calls >> Kafka followed by some Python-level post-processing)? >> >> >> 'sdk_harness_container_image_overrides' is such a property (which >>>> unfortunately only works for Dataflow today). Also this only works for >>>> Docker URLs. Maybe we can extend this property to all runners or introduce >>>> a new property that works for all types of environments ? >>> >>> >>> In my original email, I wrote that sdk_harness_container_image_overrides >>> is no more flexible than having a single option per SDK, since the default >>> container images for all external transforms in each SDK are expected to be >>> the same. For example, in the case of a pipeline with two external >>> transforms that both use the same default container image, >>> sdk_harness_container_image_overrides does not let the user give those two >>> transforms different containers. >>> >>> From a design standpoint, I feel find-replace is hacky and backwards. >>> It's cleaner to specify what kind of environment we want directly in >>> the ExpansionRequest. That way all of the environment creation logic >>> belongs inside the expansion service. >>> >> >> While Environments logically belong with Transforms, it is the expansion >> service's job to attach the right environments to the transforms that it >> vends. The caller should not need any visibility into the environment(s) >> that an expansion service uses, which is an implementation detail that the >> expansion service is free to change at any time. (In fact, whether it is >> (partially or fully) implemented as an external transform is an >> implementation detail that the end user should not need to care about or >> depend on.) >> >> I personally think pattern matching and substitution by runners (maybe >> more sophisticated than regexp on container names) is a reasonable way to >> approach customization of environments. For example, suppose I construct a >> pipeline that uses both Python and Java transforms. (I could do this from >> Go, Java, or Python). If I want to run this locally (e.g. on the Python >> FnAPI runner), I would prefer that the python bits be run in-process but >> would have to shell out (maybe via docker, maybe something cheaper) for the >> java bits. On the other hand, if I want to run this same pipeline (ideally, >> the same model proto, such that we don't have >> runner-dependent construction) on Flink, I might want the java bits to be >> inlined and the Python bits to be in a separate process. On Dataflow, both >> would live in containers. To do this, the Python runner would say "hey, I >> know that Python environment" and just swap it out for in-process, and vice >> versa. (For isolation/other reasons, one may want the option to force >> everything to be docker, but that's more of a "don't make substitutions" >> option than manually providing environment configs.) >> >> On the other hand, as we go the route of custom containers, especially >> expansion services that might vend custom containers, I think we need a way >> to push down *properties* of environments (such as resource hints) through >> the expansion service that may influence the environments that get attached >> and returned. >> >> It would be helpful for me to have concrete usecases of why a user wants >> to customize the container used by some transform they did not write, which >> could possibly inform the best course(s) of action here. >> >> >> >>> >>> >>> On Wed, Feb 3, 2021 at 5:07 PM Chamikara Jayalath <chamik...@google.com> >>> wrote: >>> >>>> >>>> >>>> On Wed, Feb 3, 2021 at 12:34 PM Kyle Weaver <kcwea...@google.com> >>>> wrote: >>>> >>>>> Hi Beamers, >>>>> >>>>> Recently we’ve had some requests on user@ and Slack for instructions >>>>> on how to use custom-built containers in cross-language pipelines >>>>> (typically calling Java transforms from a predominantly Python pipeline). >>>>> Currently, it seems like there is no way to change the container used by a >>>>> cross-language transform except by modifying and rebuilding the expansion >>>>> service. The SDK does not pass pipeline options to the expansion service >>>>> (BEAM-9449 [1]). Fixing BEAM-9449 does not solve everything, however. Even >>>>> if pipeline options are passed, the existing set of pipeline options still >>>>> limits the amount of control we have over environments. Here are the >>>>> existing pipeline options that I’m aware of: >>>>> >>>>> Python [2] and Go [3] have these: >>>>> >>>>> - >>>>> >>>>> environment_type (DOCKER, PROCESS, LOOPBACK) >>>>> - >>>>> >>>>> environment_config (This one is confusingly overloaded. It’s a >>>>> string that means different things depending on environment_type. For >>>>> DOCKER, it is the Docker image URL. For PROCESS it is a JSON blob. For >>>>> EXTERNAL, it is the external service address.) >>>>> >>>>> >>>>> Whereas Java [4] has defaultEnvironmentType and >>>>> defaultEnvironmentConfig, which are named differently but otherwise act >>>>> the >>>>> same as the above. >>>>> >>>>> I was unsatisfied with environment_config for a number of reasons. >>>>> First, having a single overloaded option that can mean entirely different >>>>> things depending on context is poor design. Second, in PROCESS mode, >>>>> requiring the user to type in a JSON blob for environment_config is not >>>>> especially human-friendly (though it has also been argued that JSON makes >>>>> complex arguments like this easier to parse). Finally, we must overload >>>>> this string further to introduce new environment-specific options, such as >>>>> a mounted Docker volume (BEAM-5440 [5]). >>>>> >>>> >>>> Agree. >>>> >>>> >>>>> >>>>> To address these problems, I added a new option called >>>>> “environment_options” (BEAM-10671 [6]). (This option has been implemented >>>>> in the Python SDK, but not the other SDKs yet.) Environment_options, >>>>> similar to the “experiments” option, takes a list of strings, for example >>>>> “--environment_option=docker_container_image=my_beam_sdk:latest”. It could >>>>> be argued we should have made “docker_container_image” etc. top-level >>>>> options instead, but this “catch-all” design makes what I am about to >>>>> propose a lot easier. >>>>> >>>>> The solution proposed in PR #11638 [7] set a flag to include >>>>> unrecognized pipeline options during serialization, since otherwise >>>>> unrecognized options are dropped. In a Python pipeline, this will allow us >>>>> to set environment_config and default_environment_config to separate >>>>> values, for Python and Java containers, respectively. However, this still >>>>> limits us to one container image for all Python and Go transforms, and one >>>>> container image for all Java transforms. As more cross-language transforms >>>>> are implemented, sooner or later someone will want to have different Java >>>>> SDK containers for different external transforms. >>>>> >>>>> (I should also mention the sdk_harness_container_image_overrides >>>>> pipeline option [8], which is currently only supported by the Dataflow >>>>> runner. It lets us basically perform a find/replace on container image >>>>> strings. This is not significantly more flexible than having a single >>>>> option per SDK, since the default container images for all external >>>>> transforms in each SDK are expected to be the same.) >>>>> >>>>> Environments logically belong with transforms, and that’s how it works >>>>> in the Runner API [9]. The problem now is that from the user’s >>>>> perspective, >>>>> the environment is bound to the expansion service. After addressing >>>>> BEAM-9449, the problem will be that one or two environments at most are >>>>> bound to the pipeline. Ideally, though, users should have fully granular >>>>> control over environments at the transform level. >>>>> >>>>> All this context for a very simple proposal: we should have all >>>>> ExternalTransform subclasses take optional environment_type and >>>>> environment_options fields in their constructors. As with their >>>>> corresponding pipeline options, these options would default to DOCKER and >>>>> none, respectively. Then we could overwrite the environment_type and >>>>> environment_options in the pipeline options passed to the expansion >>>>> service >>>>> with these values. (Alternatively, we could pass environment_type and >>>>> environment_options to the expansion service individually to avoid having >>>>> to overwrite their original values, but their original values should be >>>>> irrelevant to the expansion service anyway.) >>>>> >>>>> What do you think? >>>>> >>>> >>>> So, an external transform is uniquely identified by its URN. An >>>> external transform identified by a URN may refer to an arbitrary composite >>>> which may have sub-transforms that refer to different environments. I think >>>> with the above proposal we'll lose this flexibility. >>>> What we need is a way to override environments (or properties of >>>> environments) that results in the final pipeline proto. Once we modify such >>>> environments in the proto it will be reflected to all transforms that >>>> utilize such environments. >>>> >>>> 'sdk_harness_container_image_overrides' is such a property (which >>>> unfortunately only works for Dataflow today). Also this only works for >>>> Docker URLs. Maybe we can extend this property to all runners or introduce >>>> a new property that works for all types of environments ? >>>> >>>> Thanks, >>>> Cham >>>> >>>> >>>>> >>>>> [1] https://issues.apache.org/jira/browse/BEAM-9449 >>>>> >>>>> [2] >>>>> https://github.com/apache/beam/blob/f2c9b6e1aa5d38385f4c168107c85d4fe7f0f259/sdks/python/apache_beam/options/pipeline_options.py#L1097-L1115 >>>>> >>>>> [3] >>>>> https://github.com/apache/beam/blob/b56b61a9a6401271f14746000ecc38b17aab753d/sdks/go/pkg/beam/options/jobopts/options.go#L41-L53 >>>>> >>>>> [4] >>>>> https://github.com/apache/beam/blob/b56b61a9a6401271f14746000ecc38b17aab753d/sdks/java/core/src/main/java/org/apache/beam/sdk/options/PortablePipelineOptions.java#L53-L71 >>>>> >>>>> [5] https://issues.apache.org/jira/browse/BEAM-5440 >>>>> >>>>> [6] https://issues.apache.org/jira/browse/BEAM-10671 >>>>> >>>>> [7] https://github.com/apache/beam/pull/11638 >>>>> >>>>> [8] >>>>> https://github.com/apache/beam/blob/f2c9b6e1aa5d38385f4c168107c85d4fe7f0f259/sdks/python/apache_beam/options/pipeline_options.py#L840-L850 >>>>> >>>>> [9] >>>>> https://github.com/apache/beam/blob/b56b61a9a6401271f14746000ecc38b17aab753d/model/pipeline/src/main/proto/beam_runner_api.proto#L194 >>>>> >>>>>