The short answer
At-least-once: every event is processed ≥1 time. After a crash you may reprocess, so duplicates are possible unless *you* make writes idempotent.
Exactly-once: each event affects the sink as if it ran once, even when the job retries. Teams often get there with dedupe, transactional sinks, or idempotent upserts (sometimes called effectively once).
crash mid-write
at-least-once: may write again --> duplicates unless idempotent
exactly-once: replay + commit protocol / idempotent sink --> one effectRelated: at-most-once
Process ≤1 time; may lose events on failure. Rarely acceptable for money/audit data.
Reality check
True end-to-end exactly-once across Kafka → Flink → warehouse is hard. Many teams aim for at-least-once + idempotent merges.
Interview tip: Separate *delivery* (transport) from *processing effect* (business result). Say "exactly-once semantics" carefully and mention as the practical twin.