【AWS】2026/04/09 のアップデートまとめ

# ウォームプール対応のマネージドノードグループ作成例
$ eksctl create nodegroup \
  --cluster my-cluster \
  --name warm-pool-ng \
  --node-type m5.large \
  --nodes 3 \
  --nodes-min 1 \
  --nodes-max 10 \
  --asg-access \
  --warm-pool-enabled \
  --warm-pool-min-size 2 \
  --warm-pool-max-prepared-capacity 5

resource "aws_autoscaling_group" "eks_warm_pool" {
  name                = "${var.cluster_name}-warm-pool-asg"
  vpc_zone_identifier = var.subnet_ids
  
  min_size         = 1
  max_size         = 10
  desired_capacity = 3

  launch_template {
    id      = aws_launch_template.eks_node.id
    version = "$Latest"
  }

  warm_pool {
    pool_state                  = "Stopped"
    min_size                    = 2
    max_group_prepared_capacity = 5
    
    instance_reuse_policy {
      reuse_on_scale_in = true
    }
  }

  tag {
    key                 = "kubernetes.io/cluster/${var.cluster_name}"
    value               = "owned"
    propagate_at_launch = true
  }
}

apiVersion: apps/v1
kind: Deployment
metadata:
  name: cluster-autoscaler
spec:
  template:
    spec:
      containers:
      - image: k8s.gcr.io/autoscaling/cluster-autoscaler:v1.21.0
        name: cluster-autoscaler
        command:
        - ./cluster-autoscaler
        - --v=4
        - --stderrthreshold=info
        - --cloud-provider=aws
        - --skip-nodes-with-local-storage=false
        - --expander=least-waste
        - --node-group-auto-discovery=asg:tag=k8s.io/cluster-autoscaler/enabled,k8s.io/cluster-autoscaler/my-cluster
        - --scale-down-enabled=true
        - --scale-down-delay-after-add=1m
        - --scale-down-unneeded-time=2m

import boto3

sagemaker = boto3.client('sagemaker')

# HyperPodクラスター作成時のギャングスケジューリング設定
cluster_config = {
    'ClusterName': 'my-ml-cluster',
    'InstanceGroups': [
        {
            'InstanceGroupName': 'worker-group',
            'InstanceType': 'ml.p4d.24xlarge',
            'InstanceCount': 8,
            'LifeCycleConfig': {
                'SourceS3Uri': 's3://my-bucket/lifecycle-config.sh',
                'OnCreate': 'lifecycle-config.sh'
            }
        }
    ],
    'Orchestrator': {
        'Eks': {
            'ClusterArn': 'arn:aws:eks:us-west-2:123456789012:cluster/my-eks-cluster'
        }
    }
}

# ギャングスケジューリング有効化
training_job_config = {
    'RoleArn': 'arn:aws:iam::123456789012:role/SageMakerRole',
    'InputDataConfig': [{
        'ChannelName': 'training',
        'DataSource': {
            'S3DataSource': {
                'S3DataType': 'S3Prefix',
                'S3Uri': 's3://my-bucket/training-data/'
            }
        }
    }],
    'AlgorithmSpecification': {
        'TrainingImage': '763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:1.12.0-gpu-py38',
        'TrainingInputMode': 'File'
    },
    'HyperParameters': {
        'gang-scheduling': 'enabled',
        'min-worker-nodes': '8',
        'coordination-timeout': '300'
    }
}

apiVersion: batch/v1
kind: Job
metadata:
  name: distributed-training-job
  annotations:
    scheduler.alpha.kubernetes.io/gang-scheduling: "true"
    scheduler.alpha.kubernetes.io/gang-min-available: "8"
spec:
  parallelism: 8
  completions: 8
  template:
    metadata:
      annotations:
        scheduler.alpha.kubernetes.io/gang-name: "distributed-training-job"
    spec:
      restartPolicy: Never
      containers:
      - name: pytorch-worker
        image: my-training-image:latest
        resources:
          requests:
            nvidia.com/gpu: 8
          limits:
            nvidia.com/gpu: 8