Current state: Accepted [VOTE] KIP-251: Allow timestamp manipulation in Processor API
Discussion thread: [DISCUSS] KIP-251: Allow timestamp manipulation in Processor API
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Kafka Streams has a defined "contract" about timestamp propagation at the Processor API level: all processors within a sub-topology, see the timestamp from the input topic record that is currently processed and this timestamp will be used for all result records when writing them to a topic, too. For punctuation, the used output record timestamp is the current event-time or wall-clock time (depending on the punctuation type) that triggers the punctuation (i.e., for event-time the timestamp of last processed record that fires the punctuation).
For the DSL and also for custom operators, it would be desirable to allow timestamp manipulation at Processor API level for individual records that are forwarded. This allows to support a larger scope of possible semantics for stream processing.
Using Processor API, users use
ProcessorContext to forward record to downstream operators by calling
forward(). Thus, we need to add overloads for
forward() to allow user to pass in a timestamp for the output record.
We also add a new auxiliary class to specify optional argument for
We add one new overload of
ProcessorContext#forward() that take one additional parameter of type
To. to specify optional argument like childName or output record timestamp. If users call the new overloads and set a timestamp on the
To object, the output record gets the specified timestamp assigned. For the existing methods or if no timestamp is specified on
To, the default contract using the input record timestamp for the output record will be used. Note, that the new
To class does not include forwarding by index anymore—we want to remove forwarding by index and only support forwarding to all or by name.
Compatibility, Deprecation, and Migration Plan
This change is backward compatible, as we don't alter the contract of existing methods, and only add new method with new functionality.
We can test this feature with unit tests, by implementing test Processors that manipulate the output record timestamp using the new API, and by verifying that the output record have the assigned timestamps. Existing test ensure, that the old behavior is preserved.