Streamlining Resilient Kubernetes Autoscaling with Multi-Agent Systems via an Automated Online Design Framework

TL;DR

Kubernetes autoscaling under adversarial conditions challenges traditional HPA methods. The authors propose KARMA, an automated multi-agent framework that decomposes operational resilience into failure-specific roles and missions, trained in a digital twin via MAPPO and transferred to real clusters. KARMA demonstrates superior resilience, adversarial robustness, and explainability compared with baselines, aided by an MLP-based transition model and structured organization constraints. The work advances practical, interpretable, and adaptable autoscaling for cloud-native deployments, with clear avenues for scaling to larger clusters and broader failure scenarios.

Abstract

In cloud-native systems, Kubernetes clusters with interdependent services often face challenges to their operational resilience due to poor workload management issues such as resource blocking, bottlenecks, or continuous pod crashes. These vulnerabilities are further amplified in adversarial scenarios, such as Distributed Denial-of-Service attacks (DDoS). Conventional Horizontal Pod Autoscaling (HPA) approaches struggle to address such dynamic conditions, while reinforcement learning-based methods, though more adaptable, typically optimize single goals like latency or resource usage, neglecting broader failure scenarios. We propose decomposing the overarching goal of maintaining operational resilience into failure-specific sub-goals delegated to collaborative agents, collectively forming an HPA Multi-Agent System (MAS). We introduce an automated, four-phase online framework for HPA MAS design: 1) modeling a digital twin built from cluster traces; 2) training agents in simulation using roles and missions tailored to failure contexts; 3) analyzing agent behaviors for explainability; and 4) transferring learned policies to the real cluster. Experimental results demonstrate that the generated HPA MASs outperform three state-of-the-art HPA systems in sustaining operational resilience under various adversarial conditions in a proposed complex cluster.

Figures (5)

• Figure 1: Overview of the KARMA framework in use with a Kubernetes cluster
• Figure 2: Example of inter-agent graph inferred from trajectories during a DDoS attack: The DDoS Manager holds authority over the other agents, coordinating their responses. Additional relations capture observed communication (e.g., Bottleneck and Resource Managers) and awareness links (e.g., acquaintance between Failure and Bottleneck Managers).
• Figure 3: A graph representation of a "Chained Services" cluster with four services
• Figure 4: Learning curves across baselines for the mixed scenario over 2000 episodes.
• Figure 5: Dendrogram obtained after hierarchical clustering of agent trajectories for role inference in the mixed scenario.