AWS cho AI/Agent Developers — Day 6: CI/CD cho AI Agents với CodePipeline

5 ngày qua chúng ta đã build:

✅ Day 1 — MCP server trên ECS Fargate
✅ Day 2 — DynamoDB Global Tables
✅ Day 3 — ElastiCache + Bedrock caching
✅ Day 4 — Lambda + Bedrock serverless
✅ Day 5 — Route53 multi-region

Hôm nay kết nối tất cả bằng CI/CD. Ship agent updates bằng tay không phải production.

Pipeline: GitHub → CodeBuild → Deploy staging → Smoke tests → Manual approval → Deploy us-east-1 → Deploy eu-west-1 → Invalidate CloudFront

1
Git push main → Build Docker → Deploy Staging → Smoke Tests
2
                                                   │
3
                                          Manual Approval
4
                                                   │
5
                                    ┌──────────────┴──────────────┐
6
                                    ▼                             ▼
7
                          us-east-1 Blue/Green          eu-west-1 Blue/Green
8
                                    │                             │
9
                                    └───────── CloudFront ────────┘

Step 1: buildspec.yml#

1
phases:
2
  install:
3
    commands:
4
      - npm ci && npm run build
5
  pre_build:
6
    commands:
7
      - aws ecr get-login-password | docker login --username AWS --password-stdin $ECR_REPOSITORY_URI
8
  build:
9
    commands:
10
      - docker build -t $ECR_REPOSITORY_URI:latest .
11
      - docker tag $ECR_REPOSITORY_URI:latest $ECR_REPOSITORY_URI:$IMAGE_TAG
12
  post_build:
13
    commands:
14
      - docker push $ECR_REPOSITORY_URI:latest
15
      - docker push $ECR_REPOSITORY_URI:$IMAGE_TAG
16
      - printf '[{"name":"mcp-server","imageUri":"%s:%s"}]' $ECR_REPOSITORY_URI $IMAGE_TAG > imagedefinitions.json

Step 2: Agent Smoke Tests#

Không chỉ check HTTP 200 — còn phải check agent có trả lời đúng không.

1
# Test 1: Health check
2
curl -s -o /dev/null -w "%{http_code}" https://staging.agent.yourdomain.com/health
3

4
# Test 2: Agent trả lời prompt
5
curl -s -X POST https://staging.agent.yourdomain.com/agent \
6
  -d '{"prompt":"Say hello in one word","sessionId":"smoke-test"}'
7

8
# Test 3: Tool execution
9
curl -s -X POST https://staging.agent.yourdomain.com/agent \
10
  -d '{"prompt":"List issues","sessionId":"smoke-tool"}'
11

12
# Test 4: Session persistence
13
curl -s -X POST https://staging.agent.yourdomain.com/agent \
14
  -d '{"prompt":"My name is TestAgent","sessionId":"session-1"}'
15
curl -s -X POST https://staging.agent.yourdomain.com/agent \
16
  -d '{"prompt":"What is my name?","sessionId":"session-1"}'
17

18
# Test 5: Cache hit latency
19
# Gọi 2 lần → lần 2 phải nhanh hơn (<500ms)

Step 3: Blue/Green Deployment#

CodeDeploy quản lý blue/green cho ECS:

1
[Blue - Current] ── 100% traffic
2
[Green - New]    ── 0% traffic
3

4
1. Tạo Green task set với image mới
5
2. Route test traffic vào Green
6
3. Lambda validate agent response
7
4. Shift traffic: 10% → 50% → 100%
8
5. Nếu error rate cao → auto-rollback
9
6. Blue drain + terminate

Canary: CodeDeployDefault.ECSLinear10PercentEvery1Minute

Step 4: Pipeline (CodePipeline)#

7 stages:

Stage	Action
1. Source	GitHub main branch
2. Build	CodeBuild → Docker → ECR
3. Deploy Staging	ECS deploy to staging
4. Smoke Tests	Agent health + LLM + tools
5. Manual Approval	Human checks the smoke test results
6. Deploy us-east-1	Blue/green production
7. Deploy eu-west-1	Blue/green DR region
8. Invalidate CloudFront	Clear edge cache

Step 5: Rollback#

1
# Auto-rollback khi error rate > threshold
2
aws deploy stop-deployment --deployment-id d-EXAMPLE --auto-rollback-enabled
3

4
# Manual: rollback task definition
5
aws ecs update-service --cluster mcp-server-cluster --service github-issue-mcp --task-definition github-issue-mcp:v42

Step 6: Config Versioning#

Agent không chỉ code — còn model, prompts, tools. Version control tất cả:

1
config/
2
├── models/production.yml     # Model ID, temperature
3
├── prompts/system-v2.txt     # System prompt
4
├── cache/policies.yml        # Similarity thresholds
5
└── tools/github.yml          # Tool definitions

Chi phí#

Component	Monthly
CodePipeline	~$2
CodeBuild	~$5
ECR	~$1
CodeDeploy	Free
CloudWatch alarms	~$3
Total CI/CD	~$12/tháng

Kết thúc Series 🎉#

Day	Chủ đề	AWS Services
1	Deploy MCP Server lên ECS Fargate	ECS, ECR, ALB, Secrets Manager
2	Agent State với DynamoDB	DynamoDB Global Tables, DAX
3	LLM Caching với ElastiCache + Bedrock	ElastiCache, Bedrock
4	Serverless Agent với Lambda + Bedrock	Lambda, API Gateway, Bedrock
5	Multi-Region Agent Routing với Route53	Route53, CloudFront, Global Accelerator
6	CI/CD cho AI Agents	CodePipeline, CodeBuild, CodeDeploy

Bây giờ bạn có đủ infrastructure để chạy AI agents trên AWS: compute, state, cache, serverless, routing, và automation.

Series: AWS cho AI/Agent Developers. Day 6 (finale): CI/CD với CodePipeline, blue/green, canary, smoke tests, rollback. Git push lên production.