Beyond Reactive HPA: Designing a Predictive Autoscaler with KEDA and Time-Series Forecasting

Kubernetes scaling relies predominantly on the Horizontal Pod Autoscaler (HPA), a robust feedback loop that adjusts capacity based on observed metric saturation. While reliable for steady-state traffic, HPA is inherently reactive, it mitigates resource exhaustion only after it has begun. For workloads with steep, predictable traffic ramps (such as morning log-in spikes or scheduled synchronization jobs), this reactive lag guarantees a period of transient performance degradation. To achieve strict Service Level Objectives (SLOs) during these ramps, infrastructure must shift from reacting to current load to anticipating future demand. This article details a feed-forward architecture using time-series forecasting (Prophet) and Kubernetes Event-Driven Autoscaling (KEDA) to provision capacity before the demand arrives.