Exactlyonce

Exactly-once is a concept in distributed computing and data processing that refers to guarantees about how many times a given operation or event is applied to a system. The goal is that each intended effect occurs exactly one time, even in the presence of failures, retries, or duplicate messages. This stands in contrast to at-least-once delivery (where duplicates may occur) and at-most-once delivery (where some events can be lost).

In practice, exactly-once is difficult to achieve across multiple services and data stores. Common strategies include designing idempotent operations (where repeating the same input has no additional effect), using deduplication keys to detect and ignore repeated processing, and enforcing transactional boundaries that couple data changes with message publication. Patterns such as the transactional outbox—writing domain data and outgoing messages within one transaction—help align state changes with message delivery. Distributed transactions, like two-phase commits, are another approach but can introduce complexity and performance considerations.

Architectures often rely on a combination of approaches. Idempotent producers and consumers in message streams, coupled with a durable store of processed identifiers, are frequently used. For long-running workflows, sagas with compensating actions can provide an exactly-once-like outcome without locking across services.

Limitations include increased system complexity, potential latency, and partial guarantees across heterogeneous systems or partitions. In practice, teams frequently pursue practical exactly-once behavior for critical operations (notably financial transactions) by prioritizing idempotence and deduplication and applying more stringent guarantees within a bounded scope. The term appears across databases, messaging platforms, and event-driven architectures, with varying levels of strictness.