I agree with Alexey and Byron. 1. We do not have any concrete evidence of our users paying attention to any of those annotations. Experimental API that were in that state for a long while are good examples. A possible exception is a deprecated annotation. My preference would be to simplify annotations to nothing (stable enough for use and will evolve backward compatibility), and maybe deprecated annotations. 2. If you all think that Experimental annotation is needed, Byron's suggestion (more or less what we do today) but with some concrete life cycle definitions of those annotations would be useful to our users. (An example could be: experimental APIs either need to graduate or be removed in X releases.)
On Tue, Apr 4, 2023 at 9:01 AM Alexey Romanenko <aromanenko....@gmail.com> wrote: > Great and long-to-wait topic to discuss. > > My personal opinion based on what I saw on different open-source projects > is that all such annotations, like @Experimental or @Stable, are not > usefull along the time and even rather useless and misleading. What > actually play roles is artifacts publishing and public API despite how it > was annotated. Once a class/method was published and available for users to > use, it should be considered as “stable" (even if it’s not yet stable from > its developers point of view) and can’t be easily removed/changed in the > next releases. > > At Beam, we have a “good" example with @Experimental that was used to > annotate many parts of code in the beginning of its creation but then > perhaps forgotten to be removed whenever this code is already used by many > users and API can’t be just changed despite of this annotation. > > So, I’m pro to dismiss such annotations and consider all public and > user-available API as “stable”. If it’s needed to change/remove a public > API then we should follow the procedure of API deprecation and final > removing, at least, after 3 major (x.y) Beam releases. It should help to > have the clear rules for API changes and avoiding breaking changes for > users. > > — > Alexey > > > On 3 Apr 2023, at 17:04, Byron Ellis via dev <dev@beam.apache.org> wrote: > > Honestly, I think APIs could be pretty simply defined if you think of it > in terms of the user: > > @Deprecated = this was either stable or evolve but the > functionality/interface will go away at a future date > > @Stable = the user of this API opting out of changes to functionality and > interface. For example, default options don't change for a transform > annotated this way. > > Evolving (No Annotation) = the user is opting in to changes to > functionality but not to interface. We should generally try to write > backwards compatible code, but on the other hand the release model does not > force users into an upgrade > > @Experimental = this functionality / interface might be a bad idea and > could go away at any time > > > On Mon, Apr 3, 2023 at 7:22 AM Danny McCormick via dev < > dev@beam.apache.org> wrote: > >> *;tldr - I'd like "evolving" to be further defined, specifically around >> how we will make decisions about breaking behavior and API changes* >> >> I don't particularly care what tags we use as long as they're well >> documented. With that said, I think the following framing needs to be >> documented with more definition to flesh out the underlying philosophy: >> >> *> - new code is changeable/evolving by default (so we don't have to >> always remember to annotate it) but users have confidence they can use it >> in production (because we have good software engineering practices)* >> >> * > - Experimental would be reserved for more risky things* >> * > - after we are confident an API is stable, because it has been the >> same across a couple releases, we mark it* >> >> Here, we have 3 classes of APIs - "experimental", "stable", and >> "evolving" (or alternately "undefined"). >> >> "Experimental" seems clear - we can make any changes we want. "Stable" is >> reasonably straightforward as well - we will only make non-breaking changes >> except in exceptional cases (e.g. security hole, total failure of >> functionality, etc...) >> >> With "evolving" is the idea that we can still make any changes we want, >> but we think it's less likely we'll need to? Are silent behavior changes >> acceptable here (my vote would be no)? What about breaking API changes (my >> vote would be rarely)? >> >> I think being able to change our APIs is an ok goal, but outside of a >> true experimental context we should still be weighing the cost of API >> changes against the benefit; we have a problem of people not updating to >> newer SDKs, and introducing more breaking changes will just exacerbate that >> problem. Maybe my concerns are just a consequence of me not really seeing >> the same things that you're seeing, specifically: "*I'm seeing a culture >> of being afraid to change things, even when it would be good for users, >> because our API surface area is far too large and not explicitly chosen.*" >> Mostly what I've seen is a healthy concern about making it hard for users >> to upgrade versions, but my view is probably just limited here. >> >> My ideal framing for "evolving" is: an *evolving* API can make breaking >> API changes between versions, but this will be rare and weighed against the >> cost of slowing users' upgrade process. All breaking changes will be >> communicated in our change log. An *evolving* API will not make silent >> behavior changes except in exceptional cases (e.g. patching a security gap, >> fixing total failures of functionality). >> >> Thanks, >> Danny >> >> On Mon, Apr 3, 2023 at 9:02 AM Jan Lukavský <je...@seznam.cz> wrote: >> >>> Hi, >>> >>> removing @Experimental and adding explicit @Stable annotation makes >>> sense to me. FWIW, when we were designing Euphoria API, we adopted the >>> following convention: >>> >>> - the default stability of "evolving", @Experimental for really >>> experimental code [1] >>> >>> - target @Audience of API [2] (pipeline author, runner, internal, test) >>> >>> - and @StateComplexity of operators (PTransforms) [3] >>> >>> The last part is something that was planned to be used by tools that can >>> analyze the Pipeline for performance or visualize which transform(s) are >>> most state-consuming. But this ended only as plans. :) >>> >>> Jan >>> >>> [1] >>> >>> https://github.com/apache/beam/blob/master/sdks/java/extensions/euphoria/src/main/java/org/apache/beam/sdk/extensions/euphoria/core/annotation/stability/Experimental.java >>> >>> [2] >>> >>> https://github.com/apache/beam/blob/master/sdks/java/extensions/euphoria/src/main/java/org/apache/beam/sdk/extensions/euphoria/core/annotation/audience/Audience.java >>> >>> [3] >>> >>> https://github.com/apache/beam/blob/master/sdks/java/extensions/euphoria/src/main/java/org/apache/beam/sdk/extensions/euphoria/core/annotation/operator/StateComplexity.java >>> >>> >>> On 3/31/23 23:05, Kenneth Knowles wrote: >>> > Hi all, >>> > >>> > Long ago, we adopted two annotations in Beam to communicate to users: >>> > >>> > - `@Experimental` indicates that an API might change >>> > - `@Internal` indicates that an API is not meant for users. >>> > >>> > I've seen some real problems with this approach: >>> > >>> > - Users are afraid to use `@Experimental` APIs, because they are >>> > worried they are not production-ready. But it really just means they >>> > might change, and has nothing to do with that. >>> > - People write new code and do not put `@Experimental` annotations on >>> > it, even though it really should be able to change for a while, so we >>> > can do a good job. >>> > - I'm seeing a culture of being afraid to change things, even when it >>> > would be good for users, because our API surface area is far too large >>> > and not explicitly chosen. >>> > - `@Internal` is not that well-known. And now we have many target >>> > audiences: Beam devs, PTransform devs, tool devs, pipeline authors. >>> > Some of them probably want to use `@Internal` stuff! >>> > >>> > I looked at a couple sibling projects and what they have >>> > - Flink: >>> > - Spark: >>> > >>> > They have many more tags, and some of them seem to have reverse >>> > defaults to Beam. >>> > >>> > Flink: >>> > >>> https://github.com/apache/flink/tree/master/flink-annotations/src/main/java/org/apache/flink/annotation >>> > >>> > - Experimental >>> > - Internal.java >>> > - Public >>> > - PublicEvolving >>> > - VisibleForTesting >>> > >>> > Spark: >>> > >>> https://github.com/apache/spark/tree/master/common/tags/src/main/java/org/apache/spark/annotation >>> and >>> >>> > >>> https://github.com/apache/spark/tree/master/common/tags/src/main/scala/org/apache/spark/annotation >>> > >>> > - AlphaComponent >>> > - DeveloperApi >>> > - Evolving >>> > - Experimental >>> > - Private >>> > - Stable >>> > - Unstable >>> > - Since >>> > >>> > I think it would help users to understand Beam with some simple, >>> > though possibly large-scale changes. My goal would be: >>> > >>> > - new code is changeable/evolving by default (so we don't have to >>> > always remember to annotate it) but users have confidence they can use >>> > it in production (because we have good software engineering practices) >>> > - Experimental would be reserved for more risky things >>> > - after we are confident an API is stable, because it has been the >>> > same across a couple releases, we mark it >>> > >>> > A concrete proposal to achieve this would be: >>> > >>> > - Add a @Stable annotation and use it as appropriate on our primary >>> APIs >>> > - [Possibly] add an @Evolving annotation that would also be the >>> default. >>> > - Remove most `@Experimental` annotations or change them to >>> `@Evolving` >>> > - Communicate about this (somehow). If possible, surface the >>> > `@Evolving` default in documentation. >>> > >>> > The last bit is the hardest. >>> > >>> > Kenn >>> >> >
