Description
The new message headers functionality introduced in Kafka 0.11.0.0 (https://cwiki.apache.org/confluence/display/KAFKA/KIP-82+-+Add+Record+Headers) does not appear to be respected by Kafka Streams, specifically message headers set on input topics to a Kafka Streams topology do not get propagated to the corresponding output topics of the topology.
It appears that it's at least partially due to the SourceNodeRecordDeserializer not properly respecting message headers here:
where it isn't using the new ConsumerRecord constructor which supports headers:
For additional background here is the line before which we noticed that we still have the message headers, and after which we no longer have them:
In terms of a potential solution there are a few different scenarios to consider:
1. A stream processor with one input and one output, i.e. 1-to-1, (A map/transformation for example). This is the simplest case, and one proposal would be to directly propagate any message headers from input to output.
2. A stream processor with one input and many outputs, i.e. 1-to-many, (A flatmap step for example).
3. A stream processor with multiple inputs per output, i.e. many-to-1, (A join step for example).
One proposal for supporting all possible scenarios would be to expose overloads in the Kafka Streams DSL methods to allow the user the ability to specify logic for handling of message headers.
For additional background the use case is similar to a distributed tracing use case, where the following previous work may be useful for aiding in design discussions:
Dapper (https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/36356.pdf)
or
Zipkin (https://github.com/openzipkin/zipkin)
Attachments
Issue Links
- is blocked by
-
KAFKA-6850 KIP-244: Add Record Header support to Kafka Streams Processor API
-
- Resolved
-
- is duplicated by
-
-
- Closed
-
- is related to
-
KAFKA-4125 Provide low-level Processor API meta data in DSL layer
-
- Open
-
- is superceded by
-
-
- Resolved
-
- links to
