Tuần 10 - Ngày 5: Serverless Patterns

Mục tiêu học tập

• Hiểu serverless ecosystem trên AWS
• Apply common serverless patterns
• Compare serverless vs container vs EC2

1. Serverless Definition

Đặc điểm

• No server management (you don't see EC2)
• Auto-scaling (0 → millions)
• Pay per use (per request/invocation)
• High availability built-in

AWS Serverless Services

• Compute: Lambda, Fargate
• Storage: S3
• Database: DynamoDB, Aurora Serverless v2
• Messaging: SQS, SNS, EventBridge
• API: API Gateway, AppSync
• Orchestration: Step Functions
• Streaming: Kinesis (somewhat)
• Other: SES (email), Cognito, AWS Glue

2. Serverless vs Other

Serverless (Lambda + DynamoDB)

• Pros: No infra, pay per use, auto-scale, fast deploy
• Cons: Cold start, 15-min limit, less control

Containers (Fargate + ECS)

• Pros: Containerized, more control, longer-running
• Cons: More config, cost may be higher

EC2

• Pros: Full control, all OS access, longest run
• Cons: Manage infra, scaling, patches

Decision

• Event-driven, short tasks: Lambda
• Long-running web apps: Fargate
• Specific OS, GPU, high control: EC2

3. Pattern 1: Simple REST API

Architecture

Client → API Gateway → Lambda → DynamoDB

Components

• API Gateway: HTTP routing, auth
• Lambda: business logic
• DynamoDB: data persistence

Cost (low traffic)

• 1M requests/month:
• Lambda: ~$0 (free tier)
• API Gateway: $1 (HTTP API)
• DynamoDB On-Demand: ~$1
• Total: ~$2/month for 1M requests

Scale to millions

• Auto-scale Lambda + DynamoDB
• Linear cost (no overhead)

4. Pattern 2: File Processing Pipeline

Architecture

User uploads file → S3
                    ↓ (S3 event)
                  Lambda
                    ↓
                Process (transcode, OCR, etc.)
                    ↓
                  S3 (output)
                    ↓ (event)
                  SNS → Notify user

Use case

• Image thumbnail generation
• PDF text extraction
• Video transcoding (with AWS Elemental MediaConvert)
• File virus scanning

Lambda limitations

• 15-min execution → use Step Functions or ECS for longer
• 10 GB memory max → use ECS/Batch for heavy compute

5. Pattern 3: Real-time Analytics

Architecture

Use case

• Click stream analytics
• IoT sensor data
• Real-time fraud detection

6. Pattern 4: GraphQL Backend

Architecture

Use case

• Mobile app with complex data needs
• Real-time features (chat, dashboards)
• Offline support (Amplify)

7. Pattern 5: Workflow Orchestration

Architecture

Use case

• ETL pipelines
• Saga (distributed transactions)
• ML training pipelines
• Multi-step business processes

8. Pattern 6: Event-Driven Microservices

Architecture

Communication

• Sync: API calls between services (avoid as much as possible)
• Async: events via EventBridge

9. Pattern 7: Static Website + CDN

Architecture

Developer → CI/CD → S3 (static files)
                     ↓
                   CloudFront (CDN + cache)
                     ↓
                   Route 53 (DNS)
                     ↓
                   User

Use case

• Marketing sites
• Single-page apps (React, Vue, Angular)
• Documentation sites

Cost

• S3: cents
• CloudFront: $0.085/GB after free tier
• Route 53: $0.50/zone
• Total: < $5/month for moderate traffic

10. Pattern 8: Cron / Scheduled Jobs

Architecture

Use case

• Daily reports
• Database cleanup
• Backup automation
• Periodic health checks

vs Cron on EC2

• No EC2 instance idle 99% of time
• Pay per execution
• Auto-scale (multiple schedules → multiple Lambdas)

11. Patterns 9: Async HTTP API

Vấn đề

• HTTP request timeout (29 sec on API Gateway, 30 sec ALB)
• Long-running tasks need async pattern

Architecture

1. Client → POST /jobs → API Gateway → Lambda
2. Lambda creates job_id in DynamoDB, sends to SQS
3. API returns 202 Accepted + job_id
4. Worker (Lambda/ECS) processes from SQS
5. Worker updates job status in DynamoDB
6. Client polls GET /jobs/{id} OR
   Client uses WebSocket for push notification

12. Cost Optimization

Lambda

• Right-size memory (1769 MB = 1 vCPU)
• Provisioned Concurrency only when needed
• ARM (Graviton2) Lambda: 20% cheaper
• Lambda Power Tuning tool: find optimal memory

DynamoDB

• On-Demand for unpredictable (no minimum)
• Provisioned + Auto Scaling for known load (cheaper at scale)
• Reserved capacity for steady workload

API Gateway

• HTTP API ($1/M) instead of REST API ($3.50/M) when features OK

S3 + CloudFront

• CloudFront caching reduces S3 requests
• Use S3 Transfer Acceleration only when needed

13. Common Mistakes

Anti-patterns

1. Long-running Lambda (close to 15 min) — use Step Functions/ECS
2. Lambda + RDS without RDS Proxy — connection storms
3. Synchronous chain of Lambdas — use Step Functions
4. No DLQ for async invocations — losing failed events
5. Lambda in VPC unnecessarily — adds cold start, NAT cost
6. Cold start critical paths — use Provisioned Concurrency

Câu hỏi ôn tập

1. Khi nào dùng Step Functions thay vì Lambda chain?

   Xem đáp án

   Step Functions khi: (1) Workflow có retry/error handling phức tạp per step, (2) Cần audit trail và visibility vào execution state, (3) Cần parallel hoặc conditional branching dễ manage, (4) Workflow dài > 15 phút (vượt Lambda timeout), (5) Cần human approval step. Lambda chain (gọi Lambda từ Lambda) phù hợp cho simple linear 2-3 step flows — rẻ hơn và đơn giản hơn nhưng khó debug và không có built-in retry per step.

2. API Gateway HTTP vs REST: cái nào rẻ hơn?

   Xem đáp án

   HTTP API rẻ hơn ~70%: $1/million requests vs ~$3.5/million cho REST API. HTTP API cũng có lower latency. REST API cần khi muốn: API Keys/Usage Plans, request transformation (mapping templates), WAF integration, access logging chi tiết, edge-optimized endpoint, caching tích hợp. HTTP API đủ cho hầu hết modern serverless APIs với JWT authorization.

3. Lambda + RDS có vấn đề gì? Cách giải quyết?

   Xem đáp án

   Lambda scale horizontally (nhiều concurrent executions) — mỗi execution mở DB connection → connection exhaustion khi hàng nghìn Lambda concurrent (RDS max connections: 1,000-16,000 tùy instance). Cũng slow cold start khi mở new connection. Giải pháp: RDS Proxy — pool và reuse DB connections, giảm từ N Lambda connections xuống còn ~connections cần thiết thực sự. Hoặc dùng Aurora Serverless v2 với connection pooling tích hợp.

4. Async HTTP API pattern dùng services nào?

   Xem đáp án

   Pattern: API Gateway → SQS → Lambda (hoặc Step Functions). Flow: (1) API GW nhận request → ghi vào SQS, trả 202 Accepted + requestId, (2) Lambda consume từ SQS xử lý async, (3) Client poll /status/{requestId} hoặc nhận webhook. SQS buffer giúp: retry khi Lambda fail (DLQ), scale Lambda theo queue depth, không mất requests khi Lambda overloaded. EventBridge cũng có thể thay SQS cho event-driven scenarios.

5. Static website serverless dùng services gì?

   Xem đáp án

   Standard pattern: S3 (lưu HTML/CSS/JS, enable static website hosting) + CloudFront (CDN + HTTPS với ACM cert + custom domain + cache) + Route 53 (DNS ALIAS record trỏ vào CloudFront). Optional: Lambda@Edge hoặc CloudFront Functions cho URL rewrite, auth headers. Cost gần như zero cho low-traffic sites (S3 storage rẻ, CloudFront free tier 1 TB/month 12 tháng đầu). Không cần server, không cần EC2.

Bài tập thực hành

• Build REST API serverless: API Gateway + Lambda + DynamoDB
• Build file processing: S3 → Lambda → S3
• Build scheduled cron job: EventBridge → Lambda
• Build async API: submit job → poll status
• Build static site: S3 + CloudFront + Route 53

Tài liệu tham khảo chính thức

• AWS Serverless
• Serverless Application Lens
• Lambda Power Tuning

---

Tiếp theo: Quiz Tuần 10