The 5 th generation (5G) mobile network is expected to meet huge, diverse customer demands and novel use cases with stringent QoS requirements and minimized costs. To meet these challenges, the 5G network adopts a cloud-native approach, which brings the potential for great flexibility such as dynamic scalability for its network functions. Over the years, Kubernetes became a desirable option for telecom operators as a container orchestration software for cloud-native network solutions thanks to its versatility.As it stands, Kubernetes’ default resource scaling solution HPA (Horizontal Pod Autoscaling) represents a significant step forward in addressing the need to dynamically adapt compute resources to current demands.However, this paper shows that in some cases, HPA genericity can be costly in terms of resources mobilized and QoS degradation. Experimental analysis performed on a volatile traffic profile demonstrates the limitations of HPA. We propose a deep learning-based proactive scaling solution to overcome them and balance the cost-QoS trade-off. Essentially, it predicts the future CPU load of the Containerized Network Function (CNF) and makes decisions based on a new scaling algorithm. We also embed our approach in ETSI ZSM (Zero touch network & Service Management) framework. We compare our solution with the standard Kubernetes HPA and another alternative approach found in the state of the art. Our results show that the proposed solution outperforms these scaling mechanisms in terms of maintaining QoS levels during scaling and reducing operational costs.