What is exactly-once delivery?

Exactly-once delivery is a guarantee that every event or row in a pipeline is processed exactly one time: never duplicated by a retry, never dropped by a failure. It's the strictest of the three delivery guarantees (at-most-once, at-least-once, exactly-once) and the hardest to build correctly end to end.

How it works generically

True exactly-once delivery usually needs two things working together: at-least-once delivery, so nothing gets silently dropped, paired with idempotent processing or transactional writes, so a redelivered event doesn't create a duplicate. Most systems that claim exactly-once are really at-least-once delivery plus an idempotent consumer, since guaranteeing exactly-once at the network layer alone is close to impossible in a distributed system.

Why it matters

Financial totals, inventory counts, and usage-based billing all break if a single event gets counted twice or missed once. A dashboard that's off by even a small duplicate rate erodes trust fast, and by the time someone notices the numbers are wrong, the bad data may already be baked into a report or an invoice.

Where OptimaFlo stands today

OptimaFlo does not offer a blanket exactly-once guarantee across every layer and engine. Bronze ingestion appends rows tracked by a watermark cursor, with a best-effort client-side dedup filter for REST sources that OptimaFlo's own code documents as incomplete for at least one ingestion mode. Athena has a real server-side MERGE capability built for Gold, but it is not wired into the canonical Gold-incremental path; the path that is wired uses a local, keyed merge (DELETE+INSERT by key), not a server-side MERGE INTO. Every warehouse write method is explicit that it does not deduplicate: the caller is responsible for a clean source. Retry safety in OptimaFlo depends on which layer and engine you're looking at, not a single platform-wide promise.

Related terms

• Idempotent pipeline: The property that gets you close to exactly-once behavior even when true exactly-once delivery isn't guaranteed.
• Change data capture: CDC systems often promise at-least-once or exactly-once delivery of change events; see how that compares to OptimaFlo's read model.