Time Series Forecasting Based Proactive Kubernetes Auto scaler using Facebook Prophet and Long Short-Term Memory

Abstract

The progression of cloud computing technologies in the recent years increased its usage and involvement in application deployment. Kubernetes framework provides resilient deployment strategies for distributed systems by enabling auto-scaling and failover, rollout strategies, and more. The author conceived, designed, and tested a proactive hybrid autoscaler for Kubernetes using a hybrid time-series forecasting framework based on Facebook Prophet and Long Short-Term Memory (LSTM). The model combines with Monitor-Analyze-Plan-Execute (MAPE) loop to estimate incoming request volumes and predict the pods needed to manage workload. The algorithm accurately captures seasonal variations in the workload and predicts the required pod numbers based on the workload, allowing for rapid effective resource allocation.

The evaluation of the proposed hybrid model demonstrates superior performance compared to existing single-model proactive autoscalers, achieving up to 65%–90% higher accuracy on the NASA and FIFA 1998 World Cup datasets. The proposed autoscaler was tested against system-oriented elasticity metrics having 35% - 75% improvement in over-provisioning and a 100% increase in elasticity speedup.

The study enhances the fields of autoscaling, Kubernetes deployments and container orchestration by creating novel autoscaling mechanism which increase the resource utilization, optimize availability and lowering costs. The study opens discussion for further exploration in integration multi-model systems to proactive autoscaling, integrate vertical autoscaling and adaptive deployment architectures.