International Journal of Science, Engineering and TechnologyAuthors: Sneha Iyer
Abstract: Predictive scaling represents a transformative shift in Kubernetes resource management, moving away from reactive thresholds toward proactive, data-driven orchestration. Traditional mechanisms, such as the Horizontal Pod Autoscaler (HPA), rely on observed metrics like CPU and memory utilization, which often results in a "lag" where resources are provisioned only after performance degradation has begun. By integrating machine learning (ML) models—including Time Series Analysis, Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks—Kubernetes clusters can now anticipate traffic surges and workload spikes before they occur. This review explores the architectural integration of ML providers with the Kubernetes Metrics API, the efficacy of various algorithmic approaches in reducing latency, and the cost-optimization benefits of predictive modeling. As cloud-native environments grow in complexity, predictive scaling emerges as a critical component for maintaining high availability while minimizing resource wastage in dynamic, large-scale microservices architectures.
DOI: https://doi.org/10.5281/zenodo.19481685
