[AWS] Auto Scaling: EC2 ASG, target tracking, predictive

Auto Scaling Group, ASG, Target Tracking, Step Scaling, Predictive Scaling, Scheduled Scaling

정의

Auto Scaling Group (ASG) = EC2 instance 자동 증감. desired / min / max + scaling policy.

구조

flowchart TB
    ASG[Auto Scaling Group<br/>min=2, max=20, desired=5]
    ASG --> LT["Launch Template<br/>(instance type, AMI, SG, ...)"]
    ASG --> Tg["Target Group<br/>(ALB/NLB)"]
    ASG --> Az[Multi-AZ<br/>az-a, az-b, az-c]
    ASG --> Policy[Scaling Policy]
    Policy --> TT[Target Tracking]
    Policy --> Step[Step]
    Policy --> Sched[Scheduled]
    Policy --> Pred[Predictive]

Scaling Policy 4종

1. Target Tracking (권장)

TargetTrackingConfiguration:
  PredefinedMetricSpecification:
    PredefinedMetricType: ASGAverageCPUUtilization
  TargetValue: 50.0

“CPU 평균 50% 유지하도록 알아서 증감”.

2. Step Scaling

CPU 70-80% → +1
CPU 80-90% → +2
CPU 90%+   → +5

세밀 제어. 복잡.

3. Scheduled

# 매일 오전 9시에 desired=10
aws autoscaling put-scheduled-update-group-action \
  --auto-scaling-group-name web \
  --schedule "cron(0 9 * * ? *)" \
  --desired-capacity 10

예측 가능한 트래픽 (예: 직장인 사용 패턴, 새벽 light).

4. Predictive Scaling

머신러닝으로 과거 패턴 예측 → 미리 증가. AWS managed.

Lifecycle

sequenceDiagram
    ASG->>EC2: Launch
    EC2->>EC2: bootstrap (UserData)
    EC2->>ALB: 등록
    Note over ALB,EC2: ALB health check
    ALB-->>ASG: in service
    Note over ASG: 정상 동작
    ASG->>EC2: Terminate (scale-in)
    EC2->>ALB: deregister
    Note over EC2: graceful shutdown<br/>(connection draining)
    EC2->>EC2: stop

Lifecycle Hooks

LifecycleHook:
  LifecycleTransition: autoscaling:EC2_INSTANCE_LAUNCHING
  HeartbeatTimeout: 300
  NotificationTargetARN: arn:aws:sns:...

• Launching: instance 시작 직후, 트래픽 받기 전. 데이터 워밍업, init.
• Terminating: terminate 직전. log flush, draining.

Cooldown vs Warmup

	Cooldown	Warmup
의미	다음 scaling 까지 대기	instance 가 준비되기 까지
기본	300s	0 (Target Tracking 자동)

ASG + Spot

MixedInstancesPolicy:
  InstancesDistribution:
    OnDemandBaseCapacity: 2          # 최소 on-demand
    OnDemandPercentageAboveBaseCapacity: 25
    SpotAllocationStrategy: capacity-optimized-prioritized
  LaunchTemplate:
    Overrides:
      - InstanceType: m6i.large
      - InstanceType: m6a.large
      - InstanceType: m6g.large

2 on-demand 보장 + 나머지 70% spot 비용 절감.

EC2 Auto Scaling vs Application Auto Scaling

	EC2 ASG	App Auto Scaling
대상	EC2	ECS, DynamoDB, Aurora, …
정책	동일	동일
단위	instance	task / capacity unit

흔한 함정

관련 위키

• aws-ec2
• aws-alb-nlb
• k8s-hpa-vpa (K8s 버전)
• Load Balancer