Kafka 0.11 (it will be released very soon) add supports for transactions. Thanks to that, Flink might be able to implement Kafka sink supporting "exactly-once" semantic. API changes and whole transactions support is described in KIP-98.
The goal is to mimic implementation of existing BucketingSink. New FlinkKafkaProducer011 would
- upon creation begin transaction, store transaction identifiers into the state and would write all incoming data to an output Kafka topic using that transaction
- on `snapshotState` call, it would flush the data and write in state information that current transaction is pending to be committed
- on `notifyCheckpointComplete` we would commit this pending transaction
- in case of crash between `snapshotState` and `notifyCheckpointComplete` we either abort this pending transaction (if not every participant successfully saved the snapshot) or restore and commit it.