Intelligent video analysis from sources such as urban surveillance cameras is a prolific research area today. Multiple types of computer architectures offer a wide range of possibilities when addressing the needs of computer vision technologies. When it comes to real time processing for high level and complex event detections, however, some limitations may arise, such as the computing power in the edge or the cost of sending real time video to the cloud for running advanced algorithms. In this paper, we present a functional architecture of a complete video surveillance solution and we focus on the metadata-processing event engine which takes care of the high-level video processing that is decoupled from a low-level video analysis. The low-level video analysis running in the edge generates and publishes a flow of JSON messages structure containing the details of bounding boxes detected in each frame into an asynchronous messaging service. The metadata event engine is running in a remote cloud, far from the camera locations. We present the performance evaluation of this event engine under different circumstances simulating data coming simultaneously from multiple cameras, in order to study the best strategies when deploying and partitioning distributed processing tasks.