Event-Driven Architecture

Start by deciding whether a message is a command sent to one capability owner or an event recording a completed fact. Then choose how it travels. Queues with competing consumers absorb bursts and handle poison messages through retry schedules and dead-letter handling. Publish-subscribe fans out to independent subscriber groups. Durable streams add retention, replay, and partition keys that keep required ordering intact while you consume in parallel.

Brokers redeliver, and handlers fail after acting. You learn how acknowledgement timing around durable side effects decides between at-most-once and at-least-once outcomes. You also learn how idempotency keys give you durable duplicate suppression, and why broker guarantees alone never make external effects exactly once. The transactional outbox removes the dual-write failure between saving state and publishing, and consumer inbox records deduplicate effects atomically.

A multi-service process is a sequence of local transactions, so something must drive it forward when one step stalls or fails. You compare choreography through reacting services with orchestration through an explicit coordinator that persists process-manager state. The card covers timeouts treated as workflow events, compensating actions, irreversible effects that need manual resolution, and clear completion and terminal failure states.

Messages outlive the code that wrote them, and retained events stay in the log after the schema changes. You learn to predict whether a change breaks old producers, old consumers, or retained events using backward, forward, and full compatibility and tolerant readers. You then migrate contracts in production with upcasting, parallel contract versions, and deprecation windows. Event sourcing extends the same discipline: rebuild projections from aggregate streams and correct history through new facts instead of edits.

When a flow stalls, the evidence is spread across queues, streams, and relays. You learn to read queue depth, oldest-message age, stream consumer lag, retry and dead-letter health, and outbox relay lag, and to follow correlation across the flow. Recovery means controlled replay with a bounded scope and rate, side-effect isolation during replay, and backlog draining that does not overload downstream services.