At-most-once
Fire and forget: at most once delivery sends each message once, never retries, and loses whatever was in flight when something fails.
Also called: fire-and-forget
At-most-once - fire and forget: the sender pushes each message out once, never retries, and so gives you at most once delivery, where a message arrives one time or not at all. The mechanism is the absence of one. Either the producer does not wait for an acknowledgement, or the consumer commits its position before processing, so a crash in the gap loses whatever was in flight. Kafka's documentation lists it first among the delivery semantics because it is the cheapest option: no retry buffer, no duplicate risk, no coordination.
To be fair, it has honest uses. Sampled metrics, load-shedding telemetry, anything where a gap is noise and a queue backing up hurts more than a hole in the graph.
Do you need it? Only when you have decided losing data is fine, and written that decision down somewhere. Anything someone bills, audits or alerts against wants at-least-once or better. In ubik the question does not come up: every source offset lives inside the checkpoint, so a crash replays instead of dropping.