NATS JetStream

JetStream vs core NATS: core NATS is extremely fast but ephemeral — if no subscriber is listening, messages are lost. JetStream persists messages to disk/memory streams, tracks consumer positions, and supports acknowledgements. This makes JetStream suitable for work queues, event sourcing, and audit logs while retaining NATS's low-latency characteristics.

JetStream stream: a stream named ORDERS with subjects ["orders.>"] captures orders.created, orders.updated, and orders.shipped in one store. Retention policies control eviction: MaxAge: 7d deletes old messages, MaxBytes: 10GB caps storage. Streams can have replicas across NATS cluster nodes for fault tolerance.

Push vs pull consumers: push consumers are simple but can overwhelm slow subscribers. Pull consumers are preferred for work queues — multiple instances call Fetch(batch) and only receive what they can process, with the server balancing work across pullers. Pull consumers also work well behind load balancers and in Lambda-style functions.

NATS KV: js.CreateKeyValue(nats.KeyValueConfig{Bucket: "config"}) creates a KV bucket. kv.Put("feature_flag", []byte("true")) stores a value. kv.Watch("feature_flag") returns a watcher that delivers updates in real time — useful for distributed configuration management. History depth preserves past values for audit or rollback.

JetStream delivery guarantees: at-least-once: consumers acknowledge with msg.Ack(); unacknowledged messages are redelivered after AckWait timeout. Exactly-once: publishers send with a Nats-Msg-Id header; JetStream deduplicates within a window. Consumers implement idempotent processing. The combination prevents both message loss (at-least-once) and duplicate processing (idempotent consumer).