AWS ECS Go Auto Scaling

验证：未记录

适用版本: 未记录
验证范围: 尚未登记验证记录。

本文缩写（英文全称与中文含义）

缩写	英文全称／名称	中文含义
ALB	Application Load Balancer	应用负载均衡器
API	Application Programming Interface	应用程序编程接口
CPU	Central Processing Unit	中央处理器
DB	Database	数据库
ECS	Elastic Container Service	AWS 容器编排服务
GC	Garbage Collection	垃圾回收
gRPC	gRPC Remote Procedure Calls	基于 HTTP/2 的远程过程调用框架
HTTP	Hypertext Transfer Protocol	超文本传输协议
OOM	Out Of Memory	内存不足
RPS	Requests Per Second	每秒请求数
RSS	Resident Set Size	进程常驻内存大小
SIGTERM	Termination Signal	请求进程正常终止的信号
SLO	Service Level Objective	服务等级目标
SQS	Simple Queue Service	AWS 消息队列服务
vCPU	Virtual CPU	虚拟处理器

Links#

1. Scaling Strategy#

Go 服务通常 CPU signal 很干净，但 HTTP 服务仍优先用 request rate per task，因为它更接近业务 demand。

Workload	Primary Scaling Metric	Target
HTTP API behind ALB	`ALBRequestCountPerTarget`	load test safe RPS per task 的 60-75%
CPU-bound API / gRPC	`ECSServiceAverageCPUUtilization`	60-70%
SQS / stream worker	custom `backlog_per_task`	backlog 在 SLO 内下降
memory-sensitive service	memory guard + Go runtime metrics	不用 RSS spike 直接推导 demand

Go guardrails:

p95 latency:
    <= service SLO

CPU:
    target load <= 70%
    sustained > 80% means scale-out or profile

goroutine:
    count stable under stable traffic
    no unbounded goroutine growth

GC:
    pause stays inside latency budget
    heap goal / live heap trend stable

memory:
    use GOMEMLIMIT below ECS memory limit
    keep container memory headroom >= 20%

Decision:

GOMEMLIMIT 是 Go runtime 的软内存限制，不是容器 RSS 硬上限，也不保证避免 OOM；为非 Go 内存和运行时开销留余量。Go 内存限制

web API:
    use ALBRequestCountPerTarget for predictable routing capacity
    use CPU policy when request cost varies heavily

worker:
    use backlog_per_task
    add oldest-message-age alert

runtime:
    set GOMAXPROCS intentionally when task vCPU is small
    set GOMEMLIMIT to avoid container OOM

2. ECS Policy#

aws application-autoscaling register-scalable-target --service-namespace ecs --scalable-dimension ecs:service:DesiredCount --resource-id service/prod-apps/order-api --min-capacity 2 --max-capacity 25 --region ap-east-1

ALB request count target tracking:

aws application-autoscaling put-scaling-policy --service-namespace ecs --scalable-dimension ecs:service:DesiredCount --resource-id service/prod-apps/order-api --policy-name order-api-go-rps-per-task --policy-type TargetTrackingScaling --target-tracking-scaling-policy-configuration '{
    "TargetValue": 600,
    "PredefinedMetricSpecification": {
      "PredefinedMetricType": "ALBRequestCountPerTarget",
      "ResourceLabel": "app/prod-public-alb/abc123/targetgroup/prod-order-api/def456"
    },
    "ScaleOutCooldown": 60,
    "ScaleInCooldown": 300
  }' --region ap-east-1

CPU target tracking:

aws application-autoscaling put-scaling-policy --service-namespace ecs --scalable-dimension ecs:service:DesiredCount --resource-id service/prod-apps/order-api --policy-name order-api-go-cpu --policy-type TargetTrackingScaling --target-tracking-scaling-policy-configuration '{
    "TargetValue": 65,
    "PredefinedMetricSpecification": {
      "PredefinedMetricType": "ECSServiceAverageCPUUtilization"
    },
    "ScaleOutCooldown": 60,
    "ScaleInCooldown": 300
  }' --region ap-east-1

3. Load Test For Threshold#

Goal: 找到 latency 还稳定时的 per-task throughput。

test setup:
    desiredCount = 1
    same task CPU/memory as production
    same GOMAXPROCS / GOMEMLIMIT as production
    expose pprof or runtime metrics in test env

increase load:
    100 RPS -> 300 -> 600 -> 900 -> 1200
    each step >= 10m

stop point:
    p95 latency > SLO
    5xx >= 1%
    CPU avg >= 80%
    goroutine count keeps growing
    GC pause or heap growth breaks baseline

Example result:

task size:
    1 vCPU / 512MB

observed:
    800 RPS: p95 160ms, CPU 62%, memory stable
    1000 RPS: p95 420ms, CPU 83%, goroutine count rising

safe RPS per task:
    800

autoscaling target:
    800 * 0.75 = 600
    choose ALBRequestCountPerTarget target = 600

Useful commands:

hey -z 10m -c 300 https://api.example.com/orders

go tool pprof -top http://localhost:6060/debug/pprof/profile?seconds=30

4. Readiness Checklist#

scaling:
    minCapacity >= 2
    maxCapacity checked against DB and downstream quotas
    scale-in cooldown slower than scale-out

go runtime:
    GOMAXPROCS matches task CPU allocation
    GOMEMLIMIT lower than ECS memory limit
    SIGTERM graceful shutdown implemented

observability:
    ALB RPS / latency / 5xx
    ECS CPU / memory
    goroutine count
    GC pause / heap allocation
    task restart count