Current state: Vote in progress
Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]
Pull Request: https://github.com/apache/kafka/pull/6171
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
MirrorMaker 2.0 (KIP-382) needs to know the downstream offsets of replicated records in order to provide cross-cluster offset translation. Currently, WorkerSourceTask receives this information from KafkaProducer but throws it away. It's possible that other Connectors may benefit from this change, e.g. see KIP-381, which also proposes to notify SourceTasks of ACK'd records. In particular, this proposal makes it possible to distinguish when records have been durably stored vs when they have been skipped altogether by a SourceConnector.
The callback commitRecord() will be overloaded with an extra parameter:
Currently, SourceTask includes a commitRecord() callback, which is invoked under these conditions:
- record is ACK'd from producer
- SourceRecord is filtered out from transformation chain, and thus is never sent to Kafka
- SourceTask gives up after retries and skips the SourceRecord.
The new overloaded version will be invoked only when a record is ACK'd, which implies the record was not filtered and was not skipped. This is somewhat intuitive, as the RecordMetadata can only be filled in by a producer ACK.
After an ACK, WorkerSourceTask currently logs MetadataRecord.offset() and partition() before invoking the commitRecord() callback. I propose to also invoke the new overloaded version as well:
Compatibility, Deprecation, and Migration Plan
This is a new callback and won't affect existing code beyond an additional no-op function call.
- We could introduce a new method name, but overloading an existing method is a lighter touch.
- We could invoke commitRecord() with recordMetadata == null when records are skipped, but this would be surprising and error prone to those using the API.
- SinkRecord has offset and partition fields, so we could potentially use commitRecord(SinkRecord), but this would be a confusing abuse of SinkRecord's semantics. In particular, it would be weird to have SinkRecord in the SourceTask interface.
- We could add kafkaOffset and kafkaPartition fields to SourceRecord, which would be null prior to ACK and then filled in by WorkerSourceTask after ACK. However, this breaks the "value class" semantics of SourceRecord.
- We could extend SourceRecord with AckedSourceRecord or LoggedSourceRecord, but this seems overkill.
- We could avoid using RecordMetadata in the API, and instead include kafkaOffset, kafkaPartition, kafkaTopic, and kafkaTimestamp as parameters to commitRecord(). But this is a lot of parameters, and RecordMatadata is already in the clients API.
- We could pass just the kafkaOffset to commitRecord(), but an offset is mostly meaningless without an associated partition. We could include both kafkaOffset and kafkaPartition, but this doesn't account for transformations that may change the downstream topic name. We could include kafkaTopic as well, but then we might as well include the entire RecordMetadata.