The use of machine learning (ML) technologies to predict server workloads and proactively adjust the amount of computational resource to maximize the quality of services is an enormous challenge. In this study, we introduce an ITU-T Y.3177 compliant framework for autonomous resource control and management of virtualized network infrastructures. Based on this framework, we propose (1) an architecture for a data analytics system consisting of learning and prediction components, and (2) a two-stage intelligent model pipelining mechanism for the learning component that cascades two ML models, namely nonlinear regression and multiple regression, to understand the trends of the fluctuations in CPU usage of a network node and predict the peak CPU usage of the node in the time granularity of seconds. We evaluated the proposed mechanism in an experimental network that installed in-network caching nodes as network functions. We prove that our ML models are capable of performing agile data analytics in the time granularity of seconds and can reduce the prediction errors of peak CPU usage.