Temporal Fusion Transformer Based Vertical Scaling Management for Kubernetes
Kemalcan Bora, Elli Kartsakli, Eduardo Quiñones Moreno · 2025
Language Models (LMs) deployed in Kubernetes face resource management challenges due to their variable computational demands. Traditional scaling methods, like the default Vertical Pod Autoscaler (VPA), struggle to adapt, often causing inefficient resource use and performance issues. We propose the Temporal Fusion Transformer Based Dynamic Vertical Pod Autoscaler (TFT-DVPA), which employs advanced time series forecasting to predict and adjust memory allocation. Tested with three Small Language Models (under 100 million parameters) under simulated fluctuating workloads, TFT-DVPA outperforms the default VPA, achieving lower prediction errors and improved resource efficiency.