Scholarlink

Kubernetes’ built-in auto-scalers (HPA, VPA, KEDA) rely on fixed thresholds or simple metrics, which often fail to satisfy complex service-level objectives (SLOs) under dynamic cloud workloads. This paper proposes a Reinforcement Learning (RL)-based multidimensional auto-scaler that simultaneously performs horizontal and vertical scaling while optimizing both latency and cost objectives. Unlike prior work [1, 2] that focused primarily on microservices or single-dimension scaling, the proposed approach introduces a tunable reward formulation balancing SLO adherence and resource efficiency, enabling deployment across diverse workload types. A Kubernetes-native architecture was designed that integrates Prometheus metrics, a Multidimensional Pod Auto-scaler (MPA), and an RL agent trained using PPO and DDPG. The system was evaluated on industry-scale workloads, including the Spark TPC-DS benchmark (1 TB, 104 queries) and latency-sensitive microservices, deployed on a 9-node Amazon EKS cluster. Results show that this RL auto-scaler achieves up to 30% higher CPU utilization, 15–20% lower 90th percentile latency, and ~20% cost savings compared to HPA, VPA, and KEDA. Statistical analysis across 20 runs confirms these improvements are significant. Safe exploration strategies were also discussed, including bootstrapping with HPA to ensure SLA protection during early learning, and challenges of applying RL in production-scale Kubernetes clusters. This study demonstrates that RL provides a practical and extensible path toward intelligent autoscaling for cloud-native applications, bridging the gap between academic proposals and enterprise FinOps practices.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reinforcement Learning-Based Autoscaling for Cost and Performance Optimization in Kubernetes Clusters

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vaibhav Pandey

摘要

Kubernetes’ built-in auto-scalers (HPA, VPA, KEDA) rely on fixed thresholds or simple metrics, which often fail to satisfy complex service-level objectives (SLOs) under dynamic cloud workloads. This paper proposes a Reinforcement Learning (RL)-based multidimensional auto-scaler that simultaneously performs horizontal and vertical scaling while optimizing both latency and cost objectives. Unlike prior work [1, 2] that focused primarily on microservices or single-dimension scaling, the proposed approach introduces a tunable reward formulation balancing SLO adherence and resource efficiency, enabling deployment across diverse workload types. A Kubernetes-native architecture was designed that integrates Prometheus metrics, a Multidimensional Pod Auto-scaler (MPA), and an RL agent trained using PPO and DDPG. The system was evaluated on industry-scale workloads, including the Spark TPC-DS benchmark (1 TB, 104 queries) and latency-sensitive microservices, deployed on a 9-node Amazon EKS cluster. Results show that this RL auto-scaler achieves up to 30% higher CPU utilization, 15–20% lower 90th percentile latency, and ~20% cost savings compared to HPA, VPA, and KEDA. Statistical analysis across 20 runs confirms these improvements are significant. Safe exploration strategies were also discussed, including bootstrapping with HPA to ensure SLA protection during early learning, and challenges of applying RL in production-scale Kubernetes clusters. This study demonstrates that RL provides a practical and extensible path toward intelligent autoscaling for cloud-native applications, bridging the gap between academic proposals and enterprise FinOps practices.