METHOD FOR WORKLOAD-BASED SELECTION OF LIGHTWEIGHT PREDICTION MODELS IN MICROSERVICE AUTOSCALING

The paper investigates the feasibility of using lightweight predictive models for proactive microservice autoscaling, addressing the limitations of the alternative approaches: reactive threshold-based scaling, causing potential delays in resource adjustment and response time, and the deep learning models, such as LSTM, requiring high computational resource allocation. Using Alibaba Cluster Trace dataset, microservice workloads are analyzed and classified into four distinct categories (Stable, Periodic, Spiky, Mixed) based on the coefficient of variation and peak-to-mean ratio. Coming from the considerations of simplicity in implementation, low level of computational complexity, and covering main methodological categories, six forecasting methods were selected for evaluation: simple moving average (SMA), exponential moving average (EMA), Holt-Winters smoothing, Kalman filter, autoregressive integrated moving average (ARIMA), and percentile-based estimation. Each method is tested for different forecast horizons in both vertical and horizontal scaling scenarios. The evaluation criteria were forecast accuracy (RMSE, MAE, MAPE), computational efficiency (execution time, amount of memory used), and model suitability for specific types of workloads. The results showed that lightweight approaches provide acceptable forecast accuracy (RMSE 0.0621–0.0846) with minimal computational costs (0.43–11.76 ms per forecast). Across predictive algorithms compared, SMA offers optimal efficiency for stable workloads, Holt-Winters is most effective for periodic patterns, Kalman filter excels in handling spiky and mixed workloads, while percentile-based estimation is advantageous for long-horizon volatile patterns. Aggregation at the service level significantly reduced errors for spiky workloads. Based on the findings, a method for workload-aware selection of lightweight prediction models was proposed, mapping workload type, scaling objective, and prediction horizon to the most suitable model and parameters, enabling resource-efficient autoscaling.