The scaling challenge for a system which constitutes multiple clients, which address application servers deployed on the cloud, becomes more complicate once the applications' nature imply consistent communication, e.g., video streaming. The effective scaling solution in this case is such that it will assure an acceptable client quality of experience (QoE), typically measured by video delay. In this paper, we provide a solution to the auto-scaling for cloud provider by means of analyzing the impact of various system parameters. The parameters which may impact the QoE on the client side include, but not limited to, average memory consumption, transmission and reception frequency, average CPU consumption on the side of the cloud provider. We perform Principal Component Analysis (PCA) in order to find a projection of the parameters, resulting in a set of features which can be sorted by their measure of impact. Next, we introduce scaling decision mechanism based on Random Forest (RF). Only most influencing features are employed for that, which renders the training process of the RF to be computationally effective. The proposed approach is novel in the sense that the scaling decisions found by the RF are in the projected space found by PCA (instead of having threshold derived directly from the original parameters), which is not necessarily intuitive. However, these features are numerically approved to be the most influencing. Moreover, as long as the features in the projected space are uncorrelated, it allows us to base the RF on only small subset of them, which would be ineffective in the original measurements space, where the correlation is high. In our Kubernetes-based implementation which employs this method, the resulting auto-scaler performs better than the default auto-scaler.