Containers, as a lightweight technology for virtualizing applications, have recently revolutionized the management of cloud applications. Containers can be scaled up or down rapidly and easily depending on the workload. Predicting workload is crucial for efficiently auto-scaling of resources in cloud environments. Accurate predictions help estimate the number of needed containers, which reduces costs and ensures optimal resource utilization. However, workloads for web applications often vary across different applications and timeframes. A single prediction model struggles to capture these diverse workload patterns, highlighting the need for more specialized or adaptive models to efficiently handle the dynamic nature of cloud application workloads. To address this limitation, this paper introduces a Monitor–Train–Test–Deploy (MTTD) framework, a closed-loop workload-prediction system that continuously retrains, evaluates, and selects the most suitable predictive model during runtime based on observed workload behavior and recent prediction errors. Instead of relying on one static predictor, MTTD dynamically switches among multiple learning models to maintain stable forecasting quality under changing workload conditions. Experiments conducted on containerized web-application workloads show that the proposed framework improves prediction accuracy by up to 15% compared with individual fixed models, while also reducing performance degradation during model transitions, leading to more reliable container provisioning and lower resource waste.
Nashaat et al. (Wed,) studied this question.