Kubernetes & Amazon EKS Training

中文

한국어
English
日本語
Español

中文

한국어
English
日本語
Español

Appearance

Spot Instance 使用策略 ​

支持版本: EKS 1.29+, EKS Auto Mode GA 最后更新: February 19, 2026

本指南介绍在 EKS Auto Mode 中有效使用 Spot instances（竞价实例）的策略，包括混合容量配置、多样化和中断处理。

混合 Spot 和 On-Demand 策略 ​

混合策略可以同时实现稳定性和成本效率。

yaml
# spot-ondemand-mixed.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: mixed-capacity
spec:
  template:
    spec:
      requirements:
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["m", "c", "r"]
        - key: karpenter.k8s.aws/instance-generation
          operator: Gt
          values: ["5"]
        # Allow both Spot and On-Demand
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot", "on-demand"]
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 1m
---
# Configure Spot preference in Pod
apiVersion: apps/v1
kind: Deployment
metadata:
  name: spot-friendly-app
spec:
  replicas: 10
  selector:
    matchLabels:
      app: spot-friendly
  template:
    metadata:
      labels:
        app: spot-friendly
    spec:
      # Prefer Spot instances
      affinity:
        nodeAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              preference:
                matchExpressions:
                  - key: karpenter.sh/capacity-type
                    operator: In
                    values: ["spot"]
      # For critical workloads, require On-Demand
      # requiredDuringSchedulingIgnoredDuringExecution:
      #   nodeSelectorTerms:
      #     - matchExpressions:
      #         - key: karpenter.sh/capacity-type
      #           operator: In
      #           values: ["on-demand"]
      containers:
        - name: app
          image: my-app:latest
          resources:
            requests:
              cpu: 500m
              memory: 512Mi

容量类型选择指南 ​

工作负载类型推荐容量理由
无状态 Web services优先使用 Spot可以处理中断
Batch processing仅 Spot节省成本，可重试
Databases仅 On-Demand数据完整性
CI/CD runners优先使用 Spot节省成本，可重试
Machine learning inference混合平衡成本和可用性
Machine learning training混合并使用 checkpointing长时间运行，可设置 checkpoint

Spot Instance 多样化 ​

使用多种 instance types 来分散中断风险。

yaml
# diversified-spot.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: diversified-spot
spec:
  template:
    spec:
      requirements:
        # Various instance families
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["m", "c", "r", "i", "d"]
        # Various generations
        - key: karpenter.k8s.aws/instance-generation
          operator: In
          values: ["5", "6", "7"]
        # Various sizes
        - key: karpenter.k8s.aws/instance-size
          operator: In
          values: ["large", "xlarge", "2xlarge"]
        # Various architectures
        - key: kubernetes.io/arch
          operator: In
          values: ["amd64", "arm64"]
        # Use only Spot
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot"]
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
  # Fast re-provisioning on Spot interrupt
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: 30s

多样化最佳实践 ​

策略好处配置
多个 instance families减少相关性中断["m", "c", "r", "i", "d"]
多个 generations访问更大的容量池["5", "6", "7"]
多个 sizes提高预置灵活性["large", "xlarge", "2xlarge"]
多个 architectures2 倍容量池["amd64", "arm64"]

Spot 中断处理 ​

Disruption Budget 配置 ​

yaml
# spot-disruption-budget.yaml
apiVersion: karpenter.sh/v1
kind: NodePool
metadata:
  name: spot-with-disruption-budget
spec:
  template:
    spec:
      requirements:
        - key: karpenter.sh/capacity-type
          operator: In
          values: ["spot"]
        - key: karpenter.k8s.aws/instance-category
          operator: In
          values: ["m", "c", "r"]
      nodeClassRef:
        group: eks.amazonaws.com
        kind: NodeClass
        name: default
  disruption:
    consolidationPolicy: WhenEmptyOrUnderutilized
    consolidateAfter: 1m
    # Limit concurrent node disruptions
    budgets:
      - nodes: "10%"    # Only 10% of total nodes simultaneously
      - nodes: "3"      # Or maximum 3 nodes
      # Minimize disruptions during business hours
      - nodes: "0"
        schedule: "0 9-18 * * mon-fri"  # Weekdays 9-18
        duration: 9h

面向 Spot 感知的 Pod 配置 ​

yaml
# Configure Spot interrupt handling in Pod
apiVersion: apps/v1
kind: Deployment
metadata:
  name: spot-aware-app
spec:
  replicas: 5
  selector:
    matchLabels:
      app: spot-aware
  template:
    metadata:
      labels:
        app: spot-aware
    spec:
      # Allow time for graceful shutdown
      terminationGracePeriodSeconds: 120
      containers:
        - name: app
          image: my-app:latest
          lifecycle:
            preStop:
              exec:
                command: ["/bin/sh", "-c", "sleep 90"]
      # Spread across multiple AZs
      topologySpreadConstraints:
        - maxSkew: 1
          topologyKey: topology.kubernetes.io/zone
          whenUnsatisfiable: DoNotSchedule
          labelSelector:
            matchLabels:
              app: spot-aware

中断处理检查清单 ​

项目描述实现方式
优雅关闭正确处理 SIGTERMterminationGracePeriodSeconds
PreStop hook延迟终止lifecycle.preStop
Multi-AZ 分布在 AZ 范围中断中保持存活topologySpreadConstraints
多个 replicas无单点故障replicas > 1
状态外部化不要在本地存储状态使用外部 databases/caches

成本节省示例 ​

+-----------------------------------------------------------------------------+
|                       Spot Instance Cost Savings Examples                    |
+-----------------------------------------------------------------------------+
|                                                                              |
|  Workload Type          On-Demand Cost   Spot Cost      Savings             |
|  ---------------------------------------------------------------------------  |
|  Batch Processing        $1,000/month    $300/month     70%                 |
|  Dev/Test Environment    $2,000/month    $500/month     75%                 |
|  CI/CD Pipeline          $500/month      $150/month     70%                 |
|  Non-critical API Server $3,000/month    $1,200/month   60%                 |
|                                                                              |
|  * Spot instance prices vary based on supply and demand                      |
|                                                                              |
+-----------------------------------------------------------------------------+

计算 Spot 节省 ​

要估算你潜在的 Spot 节省：

  1. 识别适合 Spot 的工作负载：无状态、容错、灵活
  2. 检查 Spot 定价历史：使用 AWS Spot Instance Advisor
  3. 计算 On-Demand 基线成本：当前或预计支出
  4. 应用平均 Spot 折扣：通常比 On-Demand 低 60-90%
  5. 考虑中断开销：用于重新预置的额外计算资源

Spot 节省公式 ​

Estimated Monthly Savings =
    (On-Demand Hours * On-Demand Price) -
    (Spot Hours * Spot Price) -
    (Interrupt Overhead * On-Demand Price)

Where:
- Interrupt Overhead = Estimated interrupts * Recovery time * Instance count

Spot 最佳实践总结 ​

实践描述
多样化 instance types使用 10+ 种 instance types 以降低中断风险
使用多个 AZs跨 3+ 个 AZs 分布以提高可用性
设置合适的 grace periods为优雅关闭预留 120+ 秒
实现 health checks快速检测并替换不健康的 pods
使用 topology spread防止所有 replicas 位于同一个 Spot pool
外部化状态不要在 Spot nodes 上存储关键数据
设置 disruption budgets限制并发中断
监控 Spot metrics跟踪中断和节省

< 上一页：Scaling Behavior | 目录 | 下一页：Operations >