Hi Kevin, Thanks, these are some great points. Just to clarify, I do agree that the subject should be an option (like in the case of RegistryAvroFormatFactory). We could fallback to subject and auto-register schemas, if schema-Id not provided explicitly. In general, I think it would be good to be more explicit about the schemas used ( https://docs.confluent.io/platform/curren/schema-registry/schema_registry_onprem_tutorial.html#auto-schema-registration <https://docs.confluent.io/platform/current/schema-registry/schema_registry_onprem_tutorial.html#auto-schema-registration> ). This would also help prevent us from overriding the ids in incompatible ways.
Under the current implementation of FlinkToProtoSchemaConverter we might end up overwriting the field-Ids. If we are able to locate a prior schema, the approach you outlined makes a lot of sense. Let me explore this a bit further and get back(in terms of feasibility). Thanks again! - Anupam On Wed, Mar 13, 2024 at 2:28 AM Kevin Lam <kevin....@shopify.com.invalid> wrote: > Hi Anupam, > > Thanks again for your work on contributing this feature back. > > Sounds good re: the refactoring/re-organizing. > > Regarding the schema-id, in my opinion this should NOT be a configuration > option on the format. We should be able to deterministically map the Flink > type to the ProtoSchema and register that with the Schema Registry. > > I think it can make sense to provide the `subject` as a parameter, so that > the serialization format can look up existing schemas. > > This would be used in something I mentioned something related above > > > Another topic I had is Protobuf's field ids. Ideally in Flink it would be > > nice if we are idiomatic about not renumbering them in incompatible ways, > > similar to what's discussed on the Schema Registry issue here: > > https://github.com/confluentinc/schema-registry/issues/2551 > > > When we construct the ProtobufSchema from the Flink LogicalType, we > shouldn't renumber the field ids in an incompatible way, so one approach > would be to use the subject to look up the most recent version, and use > that to evolve the field ids correctly. > > > On Tue, Mar 12, 2024 at 2:33 AM Anupam Aggarwal <anupam.aggar...@gmail.com > > > wrote: > > > Hi Kevin, > > > > Thanks for starting the discussion on this. > > I will be working on contributing this feature back (This was developed > by > > Dawid Wysakowicz and others at Confluent). > > > > I have opened a (very initial) draft PR here > > https://github.com/apache/flink/pull/24482 with our current > > implementation. > > Thanks for the feedback on the PR, I haven’t gotten around to > > re-organizing/refactoring the classes yet, but it would be inline with > some > > of your comments. > > > > On the overall approach there are some slight variations from the initial > > proposal. > > Our implementation relies on an explicit schema-id being passed through > the > > config. This could help in cases where one Flink type could potentially > map > > to multiple proto types. > > We could make the schema-Id optional and fall back to deriving it from > the > > rowType (during serialization) if not present? > > > > The message index handling is still TBD. I am thinking of replicating > logic > > in AbstractKafkaProtobufSerializer.java > > < > > > https://github.com/confluentinc/schema-registry/blob/342c8a9d3854d4253d785214f5dcfb1b6cc59a06/protobuf-serializer/src/main/java/io/confluent/kafka/serializers/protobuf/AbstractKafkaProtobufSerializer.java#L157 > > > > > (|Deserializer). > > Please let me know if this makes sense / or in case you have any other > > feedback. > > > > Thanks > > Anupam > > > > On Thu, Feb 29, 2024 at 8:54 PM Kevin Lam <kevin....@shopify.com.invalid > > > > wrote: > > > > > Hey Robert, > > > > > > Awesome thanks, that timeline works for me. Sounds good re: deciding on > > > FLIP once we have the PR, and thanks for looking into the field ids. > > > > > > Looking forward to it! > > > > > > On Thu, Feb 29, 2024 at 5:09 AM Robert Metzger <metrob...@gmail.com> > > > wrote: > > > > > > > Hey Kevin, > > > > > > > > Thanks a lot. Then let's contribute the Confluent implementation to > > > > apache/flink. We can't start working on this immediately because of a > > > team > > > > event next week, but within the next two weeks, we will start working > > on > > > > this. > > > > It probably makes sense for us to open a pull request of what we have > > > > already, so that you can start reviewing and maybe also contributing > to > > > the > > > > PR. > > > > I hope this timeline works for you! > > > > > > > > Let's also decide if we need a FLIP once the code is public. > > > > We will look into the field ids. > > > > > > > > > > > > On Tue, Feb 27, 2024 at 8:56 PM Kevin Lam > > <kevin....@shopify.com.invalid > > > > > > > > wrote: > > > > > > > > > Hey Robert, > > > > > > > > > > Thanks for your response. I have a partial implementation, just for > > the > > > > > decoding portion. > > > > > > > > > > The code I have is pretty rough and doesn't do any of the > refactors I > > > > > mentioned, but the decoder logic does pull the schema from the > schema > > > > > registry and use that to deserialize the DynamicMessage before > > > converting > > > > > it to RowData using a DynamicMessageToRowDataConverter class. For > the > > > > other > > > > > aspects, I would need to start from scratch for the encoder. > > > > > > > > > > Would be very happy to see you drive the contribution back to open > > > source > > > > > from Confluent, or collaborate on this. > > > > > > > > > > Another topic I had is Protobuf's field ids. Ideally in Flink it > > would > > > be > > > > > nice if we are idiomatic about not renumbering them in incompatible > > > ways, > > > > > similar to what's discussed on the Schema Registry issue here: > > > > > https://github.com/confluentinc/schema-registry/issues/2551 > > > > > > > > > > > > > > > On Tue, Feb 27, 2024 at 5:51 AM Robert Metzger < > rmetz...@apache.org> > > > > > wrote: > > > > > > > > > > > Hi all, > > > > > > > > > > > > +1 to support the format in Flink. > > > > > > > > > > > > @Kevin: Do you already have an implementation for this inhouse > that > > > you > > > > > are > > > > > > looking to upstream, or would you start from scratch? > > > > > > I'm asking because my current employer, Confluent, has a Protobuf > > > > Schema > > > > > > registry implementation for Flink, and I could help drive > > > contributing > > > > > this > > > > > > back to open source. > > > > > > If you already have an implementation, let's decide which one to > > use > > > :) > > > > > > > > > > > > Best, > > > > > > Robert > > > > > > > > > > > > On Thu, Feb 22, 2024 at 2:05 PM David Radley < > > > david_rad...@uk.ibm.com> > > > > > > wrote: > > > > > > > > > > > > > Hi Kevin, > > > > > > > Some thoughts on this. > > > > > > > I suggested an Apicurio registry format in the dev list, and > was > > > > > advised > > > > > > > to raise a FLIP for this, I suggest the same would apply here > (or > > > the > > > > > > > alternative to FLIPs if you cannot raise one). I am prototyping > > an > > > > Avro > > > > > > > Apicurio format, prior to raising the Flip, and notice that > the > > > > > > readSchema > > > > > > > in the SchemaCoder only takes a byte array ,but I need to pass > > down > > > > the > > > > > > > Kafka headers (where the Apicurio globalId identifying the > schema > > > > > lives). > > > > > > > > > > > > > > I assume: > > > > > > > > > > > > > > * for the confluent Protobuf format you would extend the > > > Protobuf > > > > > > > format to drive some Schema Registry logic for Protobuf > (similar > > to > > > > the > > > > > > way > > > > > > > Avro does it) where the magic byte _ schema id can be obtained > > and > > > > the > > > > > > > schema looked up using the Confluent Schema registry. > > > > > > > * It would be good if any protobuf format enhancements for > > > Schema > > > > > > > registries pass down the Kafka headers (I am thinking as a > > > > Map<String, > > > > > > > Object> for Avro) as well as the message payload so Apicurio > > > registry > > > > > > could > > > > > > > work with this. > > > > > > > * It would make sense to have the Confluent schema lookup > in > > > > common > > > > > > > code, which is part of the SchemaCoder readSchema logic. > > > > > > > * I assume the ProtobufSchemaCoder readSchema would return > a > > > > > Protobuf > > > > > > > Schema object. > > > > > > > > > > > > > > > > > > > > > > > > > > > > I also wondered whether these Kafka only formats should be > moved > > to > > > > the > > > > > > > Kafka connector repo, or whether they might in the future be > used > > > > > outside > > > > > > > Kafka – e.g. Avro/Protobuf files in a database. > > > > > > > Kind regards, David. > > > > > > > > > > > > > > > > > > > > > From: Kevin Lam <kevin....@shopify.com.INVALID> > > > > > > > Date: Wednesday, 21 February 2024 at 18:51 > > > > > > > To: dev@flink.apache.org <dev@flink.apache.org> > > > > > > > Subject: [EXTERNAL] [DISCUSS] FLINK-34440 Support Debezium > > Protobuf > > > > > > > Confluent Format > > > > > > > I would love to get some feedback from the community on this > JIRA > > > > > issue: > > > > > > > > https://issues.apache.org/jira/projects/FLINK/issues/FLINK-34440 > > > > > > > > > > > > > > I am looking into creating a PR and would appreciate some > review > > on > > > > the > > > > > > > approach. > > > > > > > > > > > > > > In terms of design I think we can mirror the > > > > `debezium-avro-confluent` > > > > > > and > > > > > > > `avro-confluent` formats already available in Flink: > > > > > > > > > > > > > > 1. `protobuf-confluent` format which uses DynamicMessage > > > > > > > < > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://protobuf.dev/reference/java/api-docs/com/google/protobuf/DynamicMessage > > > > > > > > > > > > > > > for encoding and decoding. > > > > > > > - For encoding the Flink RowType will be used to > > dynamically > > > > > > create a > > > > > > > Protobuf Schema and register it with the Confluent Schema > > > > > > > Registry. It will > > > > > > > use the same schema to construct a DynamicMessage and > > > serialize > > > > > it. > > > > > > > - For decoding, the schema will be fetched from the > > registry > > > > and > > > > > > use > > > > > > > DynamicMessage to deserialize and convert the Protobuf > > object > > > > to > > > > > a > > > > > > > Flink > > > > > > > RowData. > > > > > > > - Note: here there is no external .proto file > > > > > > > 2. `debezium-avro-confluent` format which unpacks the > Debezium > > > > > > Envelope > > > > > > > and collects the appropriate UPDATE_BEFORE, UPDATE_AFTER, > > > INSERT, > > > > > > DELETE > > > > > > > events. > > > > > > > - We may be able to refactor and reuse code from the > > existing > > > > > > > DebeziumAvroDeserializationSchema + > > > > > DebeziumAvroSerializationSchema > > > > > > > since > > > > > > > the deser logic is largely delegated to and these Schemas > > are > > > > > > > concerned > > > > > > > with the handling the Debezium envelope. > > > > > > > 3. Move the Confluent Schema Registry Client code to a > > separate > > > > > maven > > > > > > > module, flink-formats/flink-confluent-common, and extend it > to > > > > > support > > > > > > > ProtobufSchemaProvider > > > > > > > < > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://github.com/confluentinc/schema-registry/blob/ca226f2e1e2091c67b372338221b57fdd435d9f2/protobuf-provider/src/main/java/io/confluent/kafka/schemaregistry/protobuf/ProtobufSchemaProvider.java#L26 > > > > > > > > > > > > > > > . > > > > > > > > > > > > > > > > > > > > > Does anyone have any feedback or objections to this approach? > > > > > > > > > > > > > > Unless otherwise stated above: > > > > > > > > > > > > > > IBM United Kingdom Limited > > > > > > > Registered in England and Wales with number 741598 > > > > > > > Registered office: PO Box 41, North Harbour, Portsmouth, Hants. > > PO6 > > > > 3AU > > > > > > > > > > > > > > > > > > > > > > > > > > > >
