We used Flink and S3 compatible backends (for the Flink state) to ensure exactly once processing.

I didn't say Kafka magically solves these problems for you, but it was required for the scalability we needed.