How RobustScaler Enables QoS‑Aware Autoscaling for Complex Kubernetes Workloads

Background

Alibaba Cloud Container Service (ACK) operates large‑scale Kubernetes clusters and provides the Container Intelligence Service (CIS) platform for automated operations.

Motivation

Enterprise traffic often exhibits pronounced peaks and troughs. Fixed‑instance configurations waste resources, while existing Kubernetes autoscalers such as Horizontal Pod Autoscaler (HPA) and CronHPA react slowly to workload changes, causing QoS degradation. A proactive, QoS‑aware scaling method that leverages historical time‑series data is needed.

RobustScaler Framework

RobustScaler models workload arrivals with a Non‑Homogeneous Poisson Process (NHPP) . The intensity function λ(t) captures time‑varying arrival rates and is learned from historical metrics (e.g., request count, CPU usage). Scaling is expressed as a stochastic constrained optimization problem:

min   E[resource usage]
subject to   QoS constraints (e.g., latency ≤ L)

To estimate λ(t) efficiently, the authors develop an Alternating Direction Method of Multipliers (ADMM) algorithm that iteratively updates the NHPP parameters while satisfying the stochastic constraints.

Integration into CIS – Adaptive HPA (AHPA)

RobustScaler is deployed in CIS as the Adaptive HPA (AHPA) component. AHPA combines three mechanisms:

Active prediction : uses the trained NHPP model and the optimization solution to forecast the required replica count for the next scaling interval.

Passive prediction : applies traditional HPA logic based on real‑time metrics (CPU, memory, QPS) to react to sudden spikes.

Protection strategy : enforces user‑defined lower and upper bounds on the replica count.

The effective replica count for a workload is the maximum of the three values:

replicas = max(active_prediction, passive_prediction, protection_bound)

The architecture of AHPA is illustrated below.

Experimental Evaluation

Extensive experiments on real‑world workloads—including applications with complex periodic patterns and bursty traffic—show that RobustScaler consistently outperforms common autoscaling strategies (HPA, CronHPA, etc.). The framework achieves higher resource utilization and lower latency, confirming the effectiveness of the NHPP‑based proactive scaling and the ADMM‑driven training process.