Cost optimization is not “pick the cheapest service.” It is meeting the required outcome at the lowest total cost, including operations, data transfer, idle capacity, licenses, and recovery risk.
Apply cost levers in a stable order
- Remove idle resources and unnecessary data movement.
- Right-size and enable elasticity.
- Match storage class and retention to access.
- Commit only the steady baseline.
- Use Spot for interruption-tolerant excess work.
- Prefer managed/serverless when reduced operations outweigh unit price.
- Measure with Cost Explorer, Budgets, anomaly detection, tags, and the Cost and Usage Report.
Watch NAT Gateway processing, cross-AZ transfer, cross-Region transfer, public IPv4, idle load balancers, over-retained logs, and frequent retrieval from archival storage. Sometimes a VPC endpoint or CloudFront reduces both cost and latency.
High-frequency pairs
| Do not confuse | Decision |
|---|---|
| Multi-AZ / read replica | Failover / read scaling |
| SG / NACL | Stateful resource / stateless subnet |
| SQS / SNS | Work queue / fan-out |
| Kinesis Streams / Firehose | Retain and process / buffer and deliver |
| CloudWatch / CloudTrail / Config | Operations / API audit / configuration |
| CloudFront / Global Accelerator | HTTP cache / network acceleration |
| Gateway / interface endpoint | S3+DynamoDB route / ENI-based PrivateLink |
| Pilot light / warm standby | Core only / scaled-down full stack |
A practice-question log that improves decisions
Record four fields: decisive phrase, chosen architecture, rejected alternative, and rule to reuse. Revisit errors by rule—for example “availability versus read scale”—rather than by service chapter. When two answers seem correct, locate the requirement one option does not satisfy, then compare operational complexity and cost.
Final lens: the exam rewards explicit trade-offs. “Always use serverless,” “always use Multi-Region,” and other absolutes are usually wrong.
Final checklist
- Can I trace identity, network, data, and failure paths?
- Can I explain why the runner-up answer fails one requirement?
- Do I know which services scale reads, writes, compute, or queues?
- Did I choose recovery scope from RTO/RPO?
- Did I include operations and transfer in the cost?
Cost optimization 不是「選最便宜的 service」，而是在滿足 outcome 前提下，降低含 operations、data transfer、idle capacity、license 與 recovery risk 的 total cost。
按穩定順序套 cost levers
- 刪 idle resource 與不必要 data movement。
- Right-size 並啟用 elasticity。
- 依 access 選 storage class/retention。
- 只對 steady baseline 做 commitment。
- Interruption-tolerant 額外工作用 Spot。
- 降低 operations 的價值大於 unit price 差異時，選 managed/serverless。
- 用 Cost Explorer、Budgets、anomaly detection、tags、Cost and Usage Report 量測。
特別注意 NAT Gateway processing、cross-AZ/Region transfer、public IPv4、idle load balancer、過度保留 logs、頻繁取回 archival storage。VPC endpoint 或 CloudFront 有時能同時降成本與 latency。
高頻混淆對照
| 別混淆 | 決策 |
|---|---|
| Multi-AZ / read replica | Failover / read scaling |
| SG / NACL | Stateful resource / stateless subnet |
| SQS / SNS | Work queue / fan-out |
| Kinesis Streams / Firehose | Retain/process / buffer/deliver |
| CloudWatch / CloudTrail / Config | Operations / API audit / configuration |
| CloudFront / Global Accelerator | HTTP cache / network acceleration |
| Gateway / interface endpoint | S3+DynamoDB route / ENI-based PrivateLink |
| Pilot light / warm standby | Core only / scaled-down full stack |
真的會讓你變強的刷題紀錄
每題寫四格：決定性關鍵詞、chosen architecture、rejected alternative、可重用 rule。依 rule（例如「availability vs read scaling」）重看錯題，不要依 service chapter 背。兩個答案都像對時，找哪個沒滿足某條 requirement，再比 operations 與 cost。
最終視角： 這場考試獎勵明確 trade-off。「一律 serverless」、「一律 Multi-Region」這種絕對句通常是錯的。
最後檢查
- 我能追 identity、network、data、failure path 嗎？
- 我能說明第二好答案違反哪條需求嗎？
- 我知道哪個 service 擴 read、write、compute 或 queue 嗎？
- Recovery scope 是根據 RTO/RPO 嗎？
- Cost 有算 operations 與 transfer 嗎？