Kubernetes Horizontal Pod Autoscaler (HPA) is reactive, scaling only when resource usage crosses predefined thresholds. This approach leads to delayed scaling decisions and resource inefficiencies.
This project introduces a predictive autoscaling model that leverages machine learning on historical resource usage data to proactively determine the optimal number of replicas for a deployment. By forecasting future CPU and memory consumption, this model ensures efficient autoscaling while reducing the risk of over-provisioning and under-provisioning.
Predictive Scaling – Anticipates resource demands instead of reacting late
Machine Learning Model – Uses Linear Regression to estimate the required number of replicas
Real-time Kubernetes Integration – Fetches live metrics and adjusts scaling dynamically
Improved Resource Utilization – Reduces unnecessary replica allocation and cost overhead
Minimal Latency – Quick decision-making for optimal autoscaling
A detailed research paper accompanying this project explores the design, implementation, and evaluation of the predictive autoscaling model. Although the paper is not formally published yet, it is available for review:
👉 Read the Research Paper on Google Docs
Alternatively, you can download the PDF version from this repository.
Below is the flow diagram illustrating how metrics are collected, processed, and used for predictive scaling:
