Although Kubernetes–the de facto standard for container orchestration in CPSs–was originally designed for centralized cloud environments, it is now adapting to manage heterogeneous nodes in distributed settings. Its native reactive mechanisms, such as the Horizontal Pod Autoscaler (HPA), often lead to resource inefficiency or SLA violations. To overcome these limitations, Deep Learning (DL) techniques are increasingly being used to enable proactive autoscaling by forecasting dynamic resource demands. This paper presents a systematic review of 64 primary studies (2021–2025) investigating which DL architectures are employed to predict resource usage in Kubernetes workloads, and how these models support automatic scaling. Results show that recurrent neural networks, particularly LSTMs, dominate time series forecasting, while hybrid approaches combining DL with Reinforcement Learning are emerging for scaling decisions. Notably, 59.4% of studies validated their solutions in real Kubernetes environments. However, critical gaps remain, including a lack of standardized operational metrics and insufficient evaluation of model inference costs. These findings highlight the urgent need for comprehensive benchmarks that balance predictive accuracy, operational efficiency, and SLA compliance in cloud-native systems.