diff --git a/rfc/rfc-57/rfc-57.md b/rfc/rfc-57/rfc-57.md new file mode 100644 index 0000000000000..52e8f1745f48d --- /dev/null +++ b/rfc/rfc-57/rfc-57.md @@ -0,0 +1,111 @@ + +# RFC-57: DeltaStreamer Protobuf Support + + + +## Proposers + +- @the-other-tim-brown + +## Approvers +- @bhasudha +- @vinothchandar + +## Status + +JIRA: https://issues.apache.org/jira/browse/HUDI-4399 + +> Please keep the status updated in `rfc/README.md`. + +## Abstract + +Support consuming Protobuf messages from Kafka with the DeltaStreamer. + +## Background +Hudi's DeltaStreamer currently supports consuming Avro and JSON data from Kafka but it does not support Protobuf. Adding support will require: +1. Parsing the data from Kafka into Protobuf Messages +2. Generating a schema from a Protobuf Message class +3. Converting from Protobuf to Avro and Row + +## Implementation + +### Parsing Data from Kafka +Users will provide a classname for the Protobuf Message that is contained within a jar that is on the path. We will then implement a deserializer that parses the bytes from the kafka message into a protobuf Message. + +Configuration options: +hoodie.deltastreamer.schemaprovider.proto.className - The class to use + +### ProtobufClassBasedSchemaProvider +This new SchemaProvider will allow the user to provide a Protobuf Message class and get an Avro Schema. In the proto world, there is no concept of a nullable field so people use wrapper types such as Int32Value and StringValue to represent a nullable field. The schema provider will also allow the user to treat these wrapper fields as nullable versions of the fields they are wrapping instead of treating them as a nested message. In practice, this means that the user can choose between representing a field `Int32Value my_int = 1;` as `my_int.value` or simply `my_int` when writing the data out to the file system. + +#### Field Mappings +Protobuf -> Avro +bool -> boolean +float -> float +double -> double +enum -> enum +string -> string +bytes -> bytes +int32 -> int +sint32 -> int +fixed32 -> int +sfixed32 -> int +uint32 -> long [1] +int64 -> long +uint64 -> long +sint64 -> long +fixed64 -> long +sfixed64 -> long +message -> record [2] +repeated -> array +map -> array [3] + + +[1] Handling of Unsigned Integers and Longs: Protobuf provides support for unsigned integers and longs while Avro does not. The schema provider will convert unsigned integers and longs to Avro long type in the schema definition. +[2] All messages will be translated to a union[null, record] with a default of null. +[3] Protobuf maps allow non-string keys while avro does not so we convert maps to an array of records containing a key and a value. + +#### Schema Evolution +**Adding a Field:** +Protobuf has a default value for all fields and the translation from proto to avro schema will carry over this default value so there are no errors when adding a new field to the proto definition. +**Removing a Field:** +If a user removes a field in the Protobuf schema, the schema provider will not be able to add this field to the avro schema it generates. To avoid issues when writing data, users must use `hoodie.datasource.write.reconcile.schema=true` to properly reconcile the schemas if a field is removed from the proto definition. Users can avoid this situation by using `deprecated` field option in proto instead of removing the field from the schema. + +Configuration Options: +hoodie.deltastreamer.schemaprovider.proto.className - The class to use +hoodie.deltastreamer.schemaprovider.proto.flattenWrappers (Default: false) - By default the wrapper classes will be treated like any other message and have a nested `value` field. When this is set to true, we do not have a nested `value` field and treat the field as nullable in the generated Schema + +### ProtoToAvroConverter and ProtoToRowConverter + +A utility will be provided that can take in a Protobuf Message and convert it to an Avro GenericRecord. This will be used inside the SourceFormatAdapter to properly convert to an avro RDD. This change will be adding a new `Source.SourceType` as well so other sources in the future can implement this source type, for example Protobuf messages on PubSub. + +To convert to `Dataset` we will initially convert to Avro and then to Row to reduce the amount of changes required to initially ship the feature. Then we will do a fast-follow with a direct proto to row conversion that follows a similar approach as the proto to avro converter. + +Special handling for maps: +Protobuf allows you to use any integral or string type as a key while Avro requires a string. To account for this, we convert all maps to lists of entries in that map. + +## Rollout/Adoption Plan + +This change simply adds new functionality and will not impact an existing users. The changes to the SourceFormatAdapter will only impact Proto source types and this is the first of its kind. Users will need to start running a new DeltaStreamer to get this functionality. + +## Test Plan + +- The new source will have testing that mirrors what we are currently doing for the TestJsonKafkaSource. This will exercise both reading the records from kafka into an `JavaRDD` and a `DataSet` +- The converter code will also be executed on the above path, but it will be more thoroughly tested in its own unit tests. The unit tests will use a sample protobuf message that covers nested messages, primitive, and wrapped field types. +- The schema provider will similarly be tested within its own unit tests to validate the behavior matches expectations. +