RESEARCH NOTEAUTHORS · Imad Buljic · Delaida Muminovic · Adnan Dzelihodzic · BHAAAS 2026 · Sarajevo · 2026

Reducing p95 Latency and SLO Violations with ML-Driven Autoscaling

A lightweight predictive autoscaler for Kubernetes that uses short-horizon traffic forecasts to act before tail latency crosses an SLO.

Reactive autoscaling sees demand after it has already arrived. This work evaluates whether a small amount of forecasting can buy a latency-sensitive service enough time to scale before the next predictable transition.

The experiment compares a KEDA-style reactive baseline with linear regression and gradient-boosted tree predictors. A Go gateway and worker service expose Prometheus latency histograms; the controller uses the resulting p95 signal, a one-step-ahead request-rate forecast, and a reactive floor so prediction can add lead time without removing the safety net.

The local paper record reports repeated experiments across five traffic scenarios, with the strongest gains on structured ramps and calendar-aware patterns. The public page keeps the status conservative until an official proceedings or DOI link is available.