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

Re: Evolving a Coder for an added field

Reuven was one of the people I reached out to on this matter and he replied on this thread.

On Mon, Nov 26, 2018 at 7:07 AM Robert Bradshaw <robertwb@xxxxxxxxxx> wrote:
Modifying an existing coder is a non-starter until we have a versioning story. Creating an entirely new coder should definitely be possible, and using it either opt-in or, if a good enough case can be made, possibly even opt-out could get this unblocked.  

On Mon, Nov 26, 2018 at 3:05 PM Jeff Klukas <jklukas@xxxxxxxxxxx> wrote:
Lukasz - Were you able to get any more context on the possibility of versioning coders from other folks at Google?

It sounds like adding versioning for coders and/or schemas is potentially a large change. At this point, should I just write up some highlights from this thread in a JIRA issue for future tracking?

On Mon, Nov 12, 2018 at 8:23 PM Reuven Lax <relax@xxxxxxxxxx> wrote:
A few thoughts:

1. I agree with you about coder versioning. The lack of a good story around versioning has been a huge pain here, and it's unfortunate that nobody ever worked on this.

2. I think versioning schemas will be easier than versioning coders (especially for adding new fields). In many cases I suggest we start looking at migrating as much as possible to schemas, and in Beam 3.0 maybe we can migrate all of our internal payload to schemas. Schemas support nested fields, repeated fields, and map fields - which can model most thing.

3. There was a Beam proposal for a way to generically handle incompatible schema updates via snapshots. The idea was that such updates can be accompanied by a transform that maps a pipeline snapshot into a new snapshot with the encodings modified. 


On Tue, Nov 13, 2018 at 3:16 AM Jeff Klukas <jklukas@xxxxxxxxxxx> wrote:
Conversation here has fizzled, but sounds like there's basically a consensus here on a need for a new concept of Coder versioning that's accessible at the Java level in order to allow an evolution path. Further, it sounds like my open PR [0] for adding a new field to Metadata is essentially blocked until we have coder versioning in place.

Is there any existing documentation of these concepts, or should I go ahead and file a new Jira issue summarizing the problem? I don't think I have a comprehensive enough understanding of the Coder machinery to be able to design a solution, so I'd need to hand this off or simply leave it in the Jira backlog.

On Tue, Nov 6, 2018 at 4:38 AM Robert Bradshaw <robertwb@xxxxxxxxxx> wrote:
Yes, a Coder author should be able to register a URN with a mapping
from (components + payload) -> Coder (and vice versa), and this should
be more lightweight than manually editing the proto files.
On Mon, Nov 5, 2018 at 7:12 PM Thomas Weise <thw@xxxxxxxxxx> wrote:
> +1
> I think that coders should be immutable/versioned. The SDK should know about all the available versions and be able to associate the data (stream or at rest) with the corresponding coder version via URN. We can also look how that is solved elsewhere, for example the Kafka schema registry.
> Today we only have a few URNs for standard coders: https://github.com/apache/beam/blob/master/model/pipeline/src/main/proto/beam_runner_api.proto#L617
> I imagine we will need a coder registry where IOs and users can add their versioned coders also?
> Thanks,
> Thomas
> On Mon, Nov 5, 2018 at 7:54 AM Jean-Baptiste Onofré <jb@xxxxxxxxxxxx> wrote:
>> It makes sense to have a more concrete URN including the version.
>> Good idea Robert.
>> Regards
>> JB
>> On 05/11/2018 16:52, Robert Bradshaw wrote:
>> > I think we'll want to allow upgrades across SDK versions. A runner
>> > should be able to recognize when a coder (or any other aspect of the
>> > pipeline) has changed and adapt/reject accordingly. (Until we remove
>> > coders from sources/sinks, there's also possibly the expectation that
>> > one should be able to read data from a source written with that same
>> > coder across versions as well.)
>> >
>> > I think it really comes down to how coders are named. If we decide to
>> > let coders change arbitrarily between versions, probably the URN for
>> > SerializedJavaCoder should have the SDK version number in it. Coders
>> > that are stable across SDKs can have better, more stable URNs defined
>> > and registered.
>> >
>> > I am more OK with changing the registry to infer different coders as
>> > the SDK evolves (which would be detected and manually overwritten with
>> > the old ones, on a case-by-case basis, if they still exist). This
>> > should still be done with caution as it will make upgrading harder.
>> > Highly composite, experimental coders should possibly be designed in
>> > an intrinsically extensible way.
>> >
>> > On Mon, Nov 5, 2018 at 4:24 PM Jean-Baptiste Onofré <jb@xxxxxxxxxxxx> wrote:
>> >>
>> >> That's really a pita. It's an important and impacting change.
>> >>
>> >> I would go to 1.
>> >>
>> >> For LTS, as already said, I would create a LTS branch and only cherry
>> >> pick some changes. Using master as LTS release branch won't work IMHO.
>> >>
>> >> Regards
>> >> JB
>> >>
>> >> On 05/11/2018 15:47, Ismaël Mejía wrote:
>> >>> For some extra context this change touches more than FileIO, in
>> >>> reality this will affect updates in any file-based pipelines because
>> >>> the metadata on each file will have now an extra field for the
>> >>> lastModifiedDate.
>> >>>
>> >>> The PR looks perfect, only issue is the backwards compatibility Coder
>> >>> question. Knowing that probably Dataflow is the only one affected, I
>> >>> would like to know what can we do?
>> >>>
>> >>> [1] Should we merge and the Coder updatability be tied to SDK versions
>> >>> (which makes sense and is probably more aligned with the LTS
>> >>> discussion)?
>> >>> [2] Should we have a MetadataCoderV2? (does this imply a repeated
>> >>> Matadata object) ? In this case where is the right place to identify
>> >>> and decide what coder to use?
>> >>>
>> >>> Other ideas... ?
>> >>>
>> >>> Last thing, the link that Luke shared does not seem to work (looks
>> >>> like a googley-friendly URL, here it is the full URL for those
>> >>> interested in the drain/update proposal:
>> >>>
>> >>> [2] https://docs.google.com/document/d/1UWhnYPgui0gUYOsuGcCjLuoOUlGA4QaY91n8p3wz9MY/edit#
>> >>> On Fri, Nov 2, 2018 at 10:11 PM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
>> >>>>
>> >>>> I think the idea is that you would use one coder for paths where you don't need this information and would have FileIO provide a separate path that uses your updated coder.
>> >>>> Existing users would not be impacted and users of the new FileIO that depend on this information would not be able to have updated their pipeline in the first place.
>> >>>>
>> >>>> If the feature in FileIO is experimental, we could choose to break it for existing users though since I don't know how feasible my suggestion above is.
>> >>>>
>> >>>>
>> >>>>
>> >>>> On Fri, Nov 2, 2018 at 12:56 PM Jeff Klukas <jklukas@xxxxxxxxxxx> wrote:
>> >>>>>
>> >>>>> Lukasz - Thanks for those links. That's very helpful context.
>> >>>>>
>> >>>>> It sounds like there's no explicit user contract about evolving Coder classes in the Java SDK and users might reasonably assume Coders to be stable between SDK versions. Thus, users of the Dataflow or Flink runners might reasonably expect that they can update the Java SDK version used in their pipeline when performing an update.
>> >>>>>
>> >>>>> Based in that understanding, evolving a class like Metadata might not be possible except in a major version bump where it's obvious to users to expect breaking changes and not to expect an "update" operation to work.
>> >>>>>
>> >>>>> It's not clear to me what changing the "name" of a coder would look like or whether that's a tenable solution here. Would that change be able to happen within the SDK itself, or is it something users would need to specify?
>> >>
>> >> --
>> >> Jean-Baptiste Onofré
>> >> jbonofre@xxxxxxxxxx
>> >> http://blog.nanthrax.net
>> >> Talend - http://www.talend.com
>> --
>> Jean-Baptiste Onofré
>> jbonofre@xxxxxxxxxx
>> http://blog.nanthrax.net
>> Talend - http://www.talend.com