ECS Managed Instances のイメージキャッシュでタスク起動はどれだけ速くなるのか Fargate と比較してみた

resource "aws_ecs_capacity_provider" "mi" {
  name    = "${var.name_prefix}-mi-cp"
  cluster = aws_ecs_cluster.mi.name

  managed_instances_provider {
    infrastructure_role_arn = aws_iam_role.ecs_infrastructure.arn
    propagate_tags          = "CAPACITY_PROVIDER"

    # アイドルインスタンスの自動回収を無効化。
    # 計測中にインスタンスが入れ替わるとイメージキャッシュが消えるため必須
    infrastructure_optimization {
      scale_in_after = -1
    }

    instance_launch_template {
      ec2_instance_profile_arn = aws_iam_instance_profile.ecs_instance.arn
      capacity_option_type     = "ON_DEMAND"

      network_configuration {
        subnets         = [aws_subnet.public.id]
        security_groups = [aws_security_group.task.id]
      }

      storage_configuration {
        storage_size_gib = 50
      }

      # インスタンスタイプ固定(タイプ差のノイズ排除)。m7i.large = 2 vCPU / 8 GiB
      instance_requirements {
        allowed_instance_types = [var.mi_instance_type]

        vcpu_count {
          min = 2
          max = 2
        }

        memory_mib {
          min = 8192
          max = 8192
        }
      }
    }
  }
}

locals {
  container_definitions = jsonencode([{
    name      = "app"
    image     = var.container_image
    essential = true
    command   = ["sleep", "3600"]
  }])
}

resource "aws_ecs_task_definition" "fargate" {
  family                   = "${var.name_prefix}-fargate"
  requires_compatibilities = ["FARGATE"]
  network_mode             = "awsvpc"
  cpu                      = "1024"
  memory                   = "2048"
  container_definitions    = local.container_definitions
}

resource "aws_ecs_task_definition" "mi" {
  family                   = "${var.name_prefix}-mi"
  requires_compatibilities = ["MANAGED_INSTANCES"]
  network_mode             = "awsvpc"
  cpu                      = "1024"
  memory                   = "2048"
  container_definitions    = local.container_definitions
}

def measure_once(ecs, args, trial):
    run_task_called_at = datetime.now(timezone.utc)
    resp = ecs.run_task(**build_run_task_params(args))

    if resp.get("failures"):
        raise RuntimeError(f"RunTask failed: {resp['failures']}")
    task_arn = resp["tasks"][0]["taskArn"]

    try:
        task = wait_for_task_status(
            ecs, args.cluster, task_arn, "RUNNING", RUNNING_TIMEOUT_SEC
        )
    finally:
        ecs.stop_task(cluster=args.cluster, task=task_arn, reason="measurement done")

    row = {
        "total_sec": diff_sec(task.get("startedAt"), run_task_called_at),
        "provisioning_sec": diff_sec(task.get("pullStartedAt"), task.get("createdAt")),
        "pull_sec": diff_sec(task.get("pullStoppedAt"), task.get("pullStartedAt")),
        # MI でインスタンス入替(=キャッシュ消失)がないことの確認用
        "container_instance_arn": task.get("containerInstanceArn", ""),
        # ...
    }

# Fargate: 10回
python3 measure.py --mode fargate --cluster cm-ta-fargate \
  --task-definition <taskdef-arn> --subnet <subnet-id> --security-group <sg-id> \
  --runs 10 --csv results_fargate.csv

# Managed Instances: 11回(1回目 = キャッシュなし参考値)
python3 measure.py --mode mi --cluster cm-ta-mi \
  --task-definition <taskdef-arn> --capacity-provider cm-ta-mi-cp \
  --subnet <subnet-id> --security-group <sg-id> \
  --runs 11 --csv results_mi.csv