SLO-Scaler's uncertainty-aware autoscaling slashes Kubernetes SLO violations by 56%

SLO-Scaler, a new research framework from Shuo Wang and colleagues at (affiliation not disclosed), targets a persistent problem in microservices: autoscaling that reacts too slowly to traffic spikes. Standard Kubernetes Horizontal Pod Autoscaler (HPA) relies on CPU/memory thresholds, which lag behind bursty workloads. Predictive methods improve responsiveness but produce point forecasts that ignore uncertainty, leading to over-provisioning or oscillatory replica counts. SLO-Scaler instead uses a Bayesian LSTM to generate short-horizon predictions of request rates, tail latency, and SLO violation probability, then makes scaling decisions based on confidence intervals. Its dependency graph analysis localizes bottleneck services, ensuring only the needed services scale rather than the entire request chain.

Evaluated on the DeathStarBench Social Network benchmark deployed on Kubernetes under periodic, bursty, and long-tail traffic, SLO-Scaler delivers major gains. Under bursty traffic, it reduces SLO violation rate by 29-56%, lowers average replica count by 18-33%, and decreases scaling event frequency by 38-59% compared to baselines, all while improving tail latency. The uncertainty-aware design prevents both over-provisioning and thrashing, making it a promising approach for production systems that need to balance cost and performance during unpredictable traffic patterns.