Event-Driven Architecture
Event-driven design trades tight coupling for asynchronous workflows—and introduces ordering, duplicates, schema evolution, and distributed tracing challenges.
When to Offer This Workflow
Trigger conditions:
- Replacing long chains of synchronous HTTP calls
- Adopting Kafka, Pub/Sub, EventBridge, NATS, etc.
- Need for sagas, compensating transactions, or cross-service workflows
Initial offer:
Use six stages: (1) identify events, (2) contracts & versioning, (3) delivery semantics, (4) orchestration vs choreography, (5) observability, (6) failure & replay). Assume at-least-once delivery unless proven otherwise.
Stage 1: Identify Events
Goal: Distinguish domain events (facts that happened) from commands (requests). Assign owning bounded context per event type.
Exit condition: Event catalog: name, schema, producers, consumers, SLAs.
Stage 2: Contracts & Versioning
Goal: Schema registry or equivalent; backward-compatible evolution; consumers ignore unknown fields; deprecation policy for old versions.
Stage 3: Delivery Semantics
Goal: Partition keys for per-entity ordering; idempotent consumers; dedupe keys when exactly-once illusion is needed.
Stage 4: Orchestration vs Choreography
Goal: Central orchestrator (saga coordinator) vs decentralized choreography—trade visibility vs coupling.
Practices
- Transactional outbox when DB write and event publish must align
Stage 5: Observability
Goal: Correlation ids on events; traces spanning HTTP → broker → consumer; lag and DLQ depth metrics.
Stage 6: Failure & Replay
Goal: Dead-letter queues, replay tooling, poison message handling, and idempotent replays.
Final Review Checklist
- Event inventory with clear ownership
- Versioned contracts and compatibility rules
- Idempotent consumers; partition strategy documented
- Saga/outbox where transactional consistency required
- Tracing and replay operationalized
Tips for Effective Guidance
- Choreography can hide flows—document critical sequences as diagrams.
- Pair with message-queues and idempotency for implementation detail.
Handling Deviations
- Low volume: start with a simple queue before full Kafka topology.