New arXiv paper maps future of predictive autoscaling with Kubernetes and FL
Researchers introduce Autoscaling Drift Index and uncertainty-aware strategies for edge-cloud efficiency.
A team of researchers from Varanasi and the University of Melbourne (Kumar, Verma, Buyya) published a comprehensive survey on arXiv that systematically breaks down the state and future of predictive autoscaling in cloud-native and federated cloud-edge systems. The paper introduces a taxonomy organized around triggers (reactive vs. proactive), targets (CPU, memory, latency), prediction models (ML, time-series, reinforcement learning), and evaluation metrics. It dives deep into Kubernetes Custom Resource Definitions (CRDs) and operators that implement Monitor-Analyze-Plan-Execute (MAPE) loops, enabling fully autonomous scaling decisions.
The authors also tackle two emerging areas: federated learning (FL) autoscaling and drift-aware scaling. In FL settings, they categorize reactive and proactive strategies while highlighting privacy-preserving container-level isolation. A key contribution is the Autoscaling Drift Index (ADI), a metric for detecting when scaling models become stale due to workload shifts, paired with feedback-driven correction loops. The paper concludes with open challenges – including stability control under heterogeneous workloads, cross-edge coordination, and energy-aware scaling – laying out a research agenda for next-generation intelligent autoscaling.
- Introduces a taxonomy of autoscaling techniques based on triggers, targets, prediction models, and evaluation metrics.
- Proposes the Autoscaling Drift Index (ADI) for detecting and correcting model staleness in production environments.
- Covers privacy-preserving container-level isolation and MAPE-based Kubernetes operators for federated learning workloads.
Why It Matters
Provides a research blueprint for building self-driving, latency-aware autoscaling systems in hybrid cloud-edge deployments.