Skip to main content

Reliability Guarantees

Use this matrix to understand what behavior is guaranteed in each mode.

Dispatch-Level Matrix

CapabilityLocal DispatchTransport Dispatch
DeliverySingle in-process invocationAt-least-once baseline
OrderingDeterministic in-processProvider/partition scoped
Duplicate handlingNot expected in single invocation pathRequires idempotency/inbox strategy
Cancellation propagationRequiredRequired
Retry behaviorOptional policyRequired for production transports
Dead-letter handlingOptionalRequired for production transports

Provider Family Baselines

Provider FamilyDeliveryOrderingDead-Letter
KafkaAt-least-oncePartition-orderedSupported via policy
RabbitMQAt-least-onceQueue-order best effort under competing consumersSupported via reject/dead-letter policy
Azure Service BusAt-least-onceEntity/session scoped orderingSupported via dead-letter subqueue
AWS SQS/SNS/EventBridge adaptersAt-least-onceProfile-dependentSupported via adapter policy
Google Pub/SubAt-least-onceOrdering-key scoped when enabledSupported via provider policy

Failure-Path Guarantees

Release tests verify:

  1. retry requeue increments delivery attempt,
  2. poison flow routes to dead-letter when requeue is disabled,
  3. timeout windows return no phantom message,
  4. cancellation tokens are honored.

Non-Guarantees

The framework does not promise:

  • global ordering across all partitions/queues,
  • exactly-once distributed delivery without external idempotency controls,
  • zero-loss under infrastructure outage without durable persistence patterns.

See Also