exactly-once
A message processing semantic guarantee stating that an operation will not be executed repeatedly and cause duplicate side effects even under crash-and-retry scenarios. It is verified here through crash tests and idempotency keys.
Timeline (last 90 days)
Exactly-once semantics means that in a data pipeline each record is processed only once even through failure retries, typically implemented via idempotent writes, transaction logs or two-phase commit.
There are three tiers of message delivery semantics: at-most-once (may be lost), at-least-once (at least once), and exactly-once (theoretically optimal, extremely difficult in engineering).
"Exactly-once" is a distributed systems semantic guarantee: even if a message is sent multiple times, the receiver will process it only once.
Deduplication is a capability that needs to be actually implemented and accepted; merely assigning a unique identifier to an event does not constitute an official "exactly-once" guarantee.
Merely assigning a unique identifier to an event is not enough to prevent duplicate side effects; the receiver still needs to implement deduplication logic (such as an idempotency table, a database unique constraint, or a distributed lock).
Exactly-once is a property of the receiving system and cannot be imposed by the caller.
Exactly-once semantics when involving uncontrollable external systems can generally only be achieved if the external system participates in two-phase commit (2PC) or provides an idempotent interface.
No distributed systems technique can provide exactly-once semantics for an arbitrary external API that does not participate in the transaction.
All Facts (8)
Exactly-once semantics means that in a data pipeline each record is processed only once even through failure retries, typically implemented via idempotent writes, transaction logs or two-phase commit.
50%UnverifiedThere are three tiers of message delivery semantics: at-most-once (may be lost), at-least-once (at least once), and exactly-once (theoretically optimal, extremely difficult in engineering).
50%Unverified"Exactly-once" is a distributed systems semantic guarantee: even if a message is sent multiple times, the receiver will process it only once.
50%UnverifiedDeduplication is a capability that needs to be actually implemented and accepted; merely assigning a unique identifier to an event does not constitute an official "exactly-once" guarantee.
50%UnverifiedMerely assigning a unique identifier to an event is not enough to prevent duplicate side effects; the receiver still needs to implement deduplication logic (such as an idempotency table, a database unique constraint, or a distributed lock).
50%UnverifiedExactly-once is a property of the receiving system and cannot be imposed by the caller.
50%UnverifiedExactly-once semantics when involving uncontrollable external systems can generally only be achieved if the external system participates in two-phase commit (2PC) or provides an idempotent interface.
50%UnverifiedNo distributed systems technique can provide exactly-once semantics for an arbitrary external API that does not participate in the transaction.
50%