A Hybrid SARIMA-LSTM Model for Predictive Scaling in Kubernetes : Improving Resource Efficiency in Cloud-Native Applications

Sammanfattning

As cloud-native applications continue to grow in both complexity and scale, effective auto-scaling strategies in containerized environments such as Kubernetes have become critical. Conventional reactive approaches, such as the Horizontal Pod Autoscaler (HPA), often struggle to allocate resources in time, leading to latency and service downtime. The current literature highlights possible ways of scaling systems proactively. However, the focus has mainly been on standalone models. In other fields of time-series forecasting, hybrid models have consistently demonstrated better performance. This thesis investigates a hybrid forecasting model that incorporates both the statistical model Seasonal Autoregressive Integrated Moving Average (SARIMA) and the neural network model Long Short-Term Memory (LSTM). A key challenge to predicting workload is that real-world time series typically contain both linear and non-linear components. By modeling linear components with SARIMA and non-linear components with LSTM this hybrid model can in theory produce better results. The model is evaluated on the Worldcup98 dataset and compared against standalone SARIMA, standalone LSTM, and two naive baseline models in terms of forecasting accuracy and auto-scaling performance in a real Kubernetes deployment. The results show that the hybrid model achieves the lowest Root Mean Squared Error (RMSE), confirming the ability of the LSTM components to model the residual structure left in SARIMA. When integrated into the predictive autoscaler as part of a real Kubernetes deployment, the hybrid model achieved fewer Service Level Agreement (SLA) violations and improved tail latencies compared to other models, indicating better SLA compliance and user experience. However, the marginal improvements over standalone SARIMA in auto-scaling tasks suggest that the benefits of the hybrid model may be more pronounced in workloads characterized by higher volatility or scale.