Ensuring exactly-once delivery in a message queue system like Kafka or NATS is a challenging problem because it requires addressing multiple aspects: message delivery, acknowledgment, and duplication due to retries or failures. Here’s how it can be achieved or approximated:

1. Idempotent Producers

• Mechanism: The producer assigns a unique identifier (e.g., a sequence number or UUID) to each message it sends. The broker keeps track of these identifiers to ensure duplicate messages aren’t stored multiple times.
• Example in Kafka: Kafka provides idempotent producers. When enabled, the producer appends messages to a topic partition with a unique sequence number, ensuring duplicates caused by retries are discarded.

2. Transactional Messaging

• Mechanism: Transactions are used to bundle message sends and acknowledgments. This ensures that messages are either fully committed or not at all.
• Example in Kafka: Kafka’s exactly-once semantics (EOS) allow producers to produce messages and consumers to commit offsets as a single atomic operation.
  The producer uses the transactional.id to track its state across retries and restarts.

3. Deduplication by Consumers

• Mechanism: Consumers can implement deduplication logic based on a unique message identifier (such as a UUID or sequence number) included in the message.
• Requirement: Consumers must have a way to maintain state about already processed message IDs (e.g., in a database or cache).
• Example in Practice: Many message systems assume at-least-once delivery and delegate deduplication responsibility to the consumer.

4. Acknowledgment Mechanisms

• Mechanism: Messages are delivered and acknowledged by consumers. If a consumer fails to acknowledge, the message may be retried, but careful design ensures duplicate deliveries are avoided.
• Example in Kafka: Kafka consumers track offsets, and the system ensures that offsets are committed only when a message has been successfully processed.
• Example in NATS: NATS JetStream uses acknowledgment modes to track the processing state of messages. It also supports durable subscriptions to avoid delivering the same message multiple times.

5. Partitioning and Ordering

• Mechanism: By assigning messages to partitions and ensuring consumers process a single partition sequentially, systems can reduce the complexity of managing message ordering and deduplication.
• Example in Kafka: Kafka partitions guarantee order within a partition, enabling deterministic message processing.

6. Storage Guarantees

• Mechanism: Persistent storage (like Kafka’s commit log) ensures that messages are reliably stored until acknowledged by consumers. This prevents loss or duplication due to transient failures.
• Example: Kafka ensures durability with replication, and NATS JetStream provides persistence with disk-based storage.

Limitations

• Performance Overhead: Achieving exactly-once semantics requires additional coordination (e.g., tracking offsets, deduplication), which can impact throughput and latency.
• Infrastructure Complexity: Managing transactions and deduplication requires more infrastructure, such as stateful brokers or additional database interactions.

Summary

Both Kafka and NATS provide tools for exactly-once delivery or close approximations:

• Kafka relies on idempotent producers, transactional messaging, and consumer offset management.
  Kafka实现原理详细介绍见文章：Kafka Transactions: Part 1: Exactly-Once Messaging

• NATS JetStream focuses on durable storage, acknowledgment mechanisms, and consumer-driven deduplication.

The implementation depends on the system design, trade-offs between performance and reliability, and the application’s tolerance for complexity.