Design Resilient Architectures
Event-Driven and Loosely Coupled Systems
CoreChoose APIs, queues, topics, event buses, brokers, and workflows from delivery and coupling requirements.
Aligned to the current SAA-C03 exam guide, verified August 16, 2026.
Why this matters
The key architecture choice is where to tolerate time, failure, and consumer independence. Begin with synchronous versus asynchronous behavior, then decide whether work must be buffered, copied to many consumers, routed by event content, kept broker-compatible, or coordinated as explicit state.
Must Know
- Synchronous calls give an immediate response but couple caller availability and latency to the downstream dependency. Asynchronous boundaries absorb delay and let components scale or recover independently.
- SQS durably buffers work for pull-based consumers and supplies backpressure. Standard queues favor scale with at-least-once delivery; FIFO queues add ordering and deduplication semantics where required.
- SNS pushes a published message to multiple subscribers. Pair SNS with one SQS queue per consumer when each consumer needs durable, independent backlog and retry.
- EventBridge routes events from AWS services, custom applications, and partners by event patterns; it is an event bus, not a worker backlog.
- Step Functions makes workflow state, branches, retries, timeouts, and service coordination explicit. Messaging transports work but does not by itself model a business process.
- Amazon MQ preserves common broker protocols for applications that need traditional broker compatibility; cloud-native systems usually minimize operations with SQS, SNS, or EventBridge.
- API Gateway is a managed API front door with authorization, throttling, transformations, and service integrations; a load balancer is a traffic distributor to targets.
Compare and Distinguish
- SQS vs SNS vs EventBridge: choose buffering/backpressure, push fan-out, or content-based event routing. Combine SNS/EventBridge with SQS when subscribers need durable queues.
- Step Functions vs SQS: Step Functions wins when progress and branch state must be coordinated; SQS wins when independent workers pull tasks and workflow state is unnecessary.
- Amazon MQ vs AWS-native messaging: MQ wins for protocol compatibility and minimal legacy changes; managed native services win for cloud-native scaling and lower broker administration.
- API Gateway vs ALB: API Gateway wins for managed API features and serverless/service integration; ALB wins for Layer 7 distribution to application targets such as containers or instances.
- Synchronous vs asynchronous: synchronous wins when the caller needs an immediate outcome; asynchronous wins when buffering, retry isolation, or independent scaling is more important.
Scenario examples
- Scenario: Checkout must accept orders while fulfillment is offline. Think: decouple with SQS so requests persist and workers drain backlog later.
- Scenario: One event must reach billing, analytics, and email independently. Think: fan out to separate queues so one slow consumer cannot block the others.
- Scenario: Events from many producers need routing by source and detail. Think: EventBridge patterns route without producers knowing consumers.
- Scenario: A multi-step approval needs waits, branches, retries, and visible state. Think: Step Functions orchestrates the process.
Exam traps
- SNS alone is not a durable per-consumer work backlog.
- SQS does not broadcast one copy independently to every consumer unless the design fans out into separate queues.
- EventBridge routing does not make downstream handlers idempotent.
- A FIFO queue does not guarantee exactly-once side effects in external systems.
- Step Functions is not a replacement for every queue, and API Gateway is not a worker buffer.
Key takeaways
- Choose integration from response timing, delivery semantics, ordering, fan-out, routing, and state.
- Use queues to absorb bursts and isolate failures.
- Give each independent fan-out consumer its own durable backlog when required.
- Prefer explicit orchestration only when the business process needs visible state.
How it works
- Producers place work or events on a managed integration boundary instead of calling every consumer directly.
- Consumers process at their own rate; visibility and retry behavior allow unfinished queue work to reappear.
- Topics and event buses select subscribers/routes, while queues retain consumable backlog.
- Workflow orchestration persists execution state and invokes steps according to defined transitions.
When to use it
- Use SQS for durable buffering and worker decoupling.
- Use SNS for push notifications and fan-out; add SQS for durable independent subscribers.
- Use EventBridge for loosely coupled event routing across sources and targets.
- Use Step Functions for stateful coordination and Amazon MQ when broker protocol compatibility is a hard requirement.
Security and governance implications
- Encrypt messages where required and restrict producers, consumers, and subscriptions by role and resource policy.
- Avoid sensitive payloads in events when references or tokenization meet the requirement.
- Validate API callers and apply throttling at the managed front door.
- Monitor queue age, dead-letter volume, failures, and workflow timeouts.
Operational and diagnostic signals
- If backlog grows, compare arrival rate, processing capacity, failure rate, and downstream throttling before scaling blindly.
- If duplicates cause side effects, fix consumer idempotency rather than assuming delivery will never retry.
- If fan-out consumers interfere, confirm each has an independent queue and failure policy.
More detail
- Design consumers to be idempotent because retries and duplicate deliveries can occur.
- Dead-letter queues isolate repeatedly failing messages for diagnosis without blocking healthy work.
- Ordering has throughput and design implications; require it only when the business sequence truly matters.
- A successful asynchronous acceptance response means work was accepted, not necessarily completed.
Ready for the quiz?
- Which requirement makes SQS better than a direct call?
- When should SNS fan out into SQS queues?
- What clue favors EventBridge over SNS?
- When is Step Functions a better fit than messaging?
- Why might a legacy application require Amazon MQ?
Related objectives
- D2.1.K1
- D2.1.K2
- D2.1.K5
- D2.1.K11
- D2.1.K16
- D2.1.S1
- D2.1.S3
- D2.1.S7