IT lexicon Infra & Ops Exactly-once semantics

Exactly-once semantics

Infra & Ops På svenska → Updated: 2026-05-23

Guarantee in distributed messaging systems: each message is processed exactly once — never duplicated, never dropped — despite faults, retries and partial failures.

Strictly speaking impossible between independent systems (FLP impossibility). Practically possible inside a system via a transactional producer + idempotent consumer + atomic offset commit. Kafka supports "exactly-once" since 0.11 (transactional producer + read_committed isolation).

"Effectively-once" is often what people mean: the producer may duplicate but the consumer is idempotent. That's a common pattern — use idempotency keys (request-uuid → dedupe table) instead of relying on broker magic.

← Back to the lexicon