ПОРІВНЯЛЬНИЙ АНАЛІЗ МОДЕЛЕЙ МАШИННОГО НАВЧАННЯ ДЛЯ ПРОГНОЗУВАННЯ СЕРВЕРНОГО НАВАНТАЖЕННЯ У МАСШТАБОВАНИХ ВЕБ-ДОДАТКАХ

Topicality. Scalable web applications operate under highly variable request streams, where sudden peaks, seasonal patterns and noisy resource metrics make static provisioning and purely threshold-based auto-scaling inefficient. The subject of study in the article is machine learning models for predicting server workload in cloud and containerized web systems. The purpose of the article is to compare the applicability of statistical, ensemble and deep learning models for workload prediction and to determine which model classes are most suitable for integration into adaptive server resource management. The following results were obtained. The main workload characteristics affecting prediction accuracy were identified, including trend, seasonality, burstiness, nonlinearity and metric correlation. ARIMA-based, SVR-based, XGBoost-based, LSTM/Bi-LSTM, GRU and Transformer-based approaches were compared according to prediction accuracy, ability to model nonlinear temporal dependencies, computational complexity, interpretability and suitability for real-time auto-scaling. A generalized evaluation scheme was proposed for selecting a prediction model before its integration into an adaptive resource management loop. Conclusion. The analysis shows that no single model is optimal for all workload patterns. ARIMA-type models are useful as interpretable baselines, XGBoost is effective for fast tabular prediction with engineered features, LSTM and GRU models better capture sequential dependencies, while Transformer-based models are promising for long-range and highly dynamic workloads but require more careful tuning and computational resources.