Thanks all for the input. It helped a lot for the doc. From the feedback,

- The main concern is that RawType/CoderLogicalType break the strong
mapping of schema<->coder. This is a valid concern.

- On the other hand, it is a way to make schemas the fundamental concept
(which is a goal of Beam 3) under the situation that Beam and its ecosystem
has already evolved for years with many Beam pipelines using (non-portable)
coders+custom types.

>From these feedbacks, I suggest we proceed with CoderLogicalType approach,
given the requirements noted in "Requirement" section of the doc, and in
addition,

- We should clearly document that this approach, if implemented, should not
be used to bypass the schema framework. We always encourage schema-fy
structured types.

I'll start drafting changes for each supported SDK.

Thanks again!




On Wed, Sep 3, 2025 at 1:51 PM Yi Hu <[email protected]> wrote:

> Hi all,
>
> Please find the following design doc for a portable RAW field type
> enabling arbitrary (serializable) data type to be included and take
> advantage of the Beam portable schema framework
>
> https://s.apache.org/beam-portable-raw-type
>
> It aims to solve https://github.com/apache/beam/issues/23374 (as well as
> https://github.com/apache/beam/issues/19817) as part of schema
> improvement for Beam 3 (https://github.com/apache/beam/issues/34672).
>
> It also includes an appendix of term disambiguation between
> Beam/Flink/Avro schema systems that might find useful in general.
>
> I proposed two alternative designs. Any feedback is welcome!
>
> Regards,
>
> Yi
>
> --
>
> Yi Hu, (he/him/his)
>
> Software Engineer
>
>
>

Reply via email to