Detecting an operation that deviates from the regular operation is crucial in reducing the system's downtime and increasing its overall availability. A data-driven approach to anomaly detection is based on training a particular class of neural networks called autoencoders. Traditionally, these neural networks (autoencoders) are trained semi-supervised, necessitating an accurately labeled training set. Accurately training sets (precise data about anomalies) are seldom available and represent a significant barrier to deploying data-driven anomaly detection. RUAD approach, developed by researchers from the University of Bologna and ETH, is the first modeling approach that has the potential to be trained even if the data about the anomalies is not available (in an unsupervised manner). In this presentation, we will show how we have taken the RUAD approach and adopted it in monitoring online retail platforms running on the cloud (MS Azure). To our knowledge, this is the first commercial deployment of data-driven, unsupervised anomaly detection for cloud applications. We have developed and deployed a system that automatically monitors the overall health of an online retail platform and automatically alerts the system administrators in the case of reduced availability. Data collected from MS Azure is submitted to RUAD, which analyses the data in real time and provides feedback to the system administrators. RUAD network was trained and is deployed on MS Azure. Deployment of the anomaly detection solution, which was trained without an accurate dataset of anomalous events, significantly reduced the time from the occurrence of the anomalies to the response from the support personnel. This, in turn, reduced the overall downtime and increased the platform's availability. The presented methodology applies to many cloud applications beyond online retail.