Archetype-Aware Predictive Autoscaling with Uncertainty Quantification for Serverless Workloads on Kubernetes

2025-07-08

AI Summary

To address the challenges of high dynamism and stringent SLO guarantees in Kubernetes-based serverless workloads, this paper proposes the first predictive autoscaling framework leveraging weakly supervised learning. Our method automatically clusters over 300,000 workload time windows using weak supervision—identifying, for the first time, four canonical patterns: periodic, bursty, ramp-up, and steady-noise—and integrates uncertainty quantification to enable perception-aware scaling decisions. The framework unifies time-series feature extraction, weakly supervised classification, and predictive modeling, and is natively embedded into the Kubernetes controller. Evaluated on real-world Azure Functions traces, it reduces SLO violations by 50%, shortens average response time by 40%, and incurs only a 2–8× increase in resource consumption during peak periods—demonstrating significant improvements in both performance and efficiency.

Technology Category

• Machine Learning: Scalability of ML Systems
• Planning, Routing, and Scheduling: Learning for Planning and Scheduling
• Data Mining & Knowledge Management: Scalability, Parallel & Distributed Systems

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

High-performance extreme computing (HPEC) platforms increasingly adopt serverless paradigms, yet face challenges in efficiently managing highly dynamic workloads while maintaining service-level objectives (SLOs). We propose **AAPA**, an archetype-aware predictive autoscaling system that leverages weak supervision to automatically classify 300,000,+ workload windows into four archetypes (PERIODIC, SPIKE, RAMP, STATIONARY_NOISY) with 99.8% accuracy. Evaluation on publicly available Azure Functions traces shows that AAPA reduces SLO violations by up to 50%, improves response time by 40%, albeit with a 2--8,$ imes$ increase in resource cost under spike-heavy loads.

Problem

• Efficiently managing dynamic serverless workloads on Kubernetes
• Classifying workload archetypes accurately for predictive autoscaling
• Reducing SLO violations while optimizing resource costs

Innovation

• Archetype-aware predictive autoscaling for serverless workloads
• Weak supervision classifies 300,000+ workload windows
• Reduces SLO violations by 50% and improves response time