Discovery of repeating temporal patterns and prediction based on time stamped data generated by IoT services raise important methodological issues. A typical predictive problem, addressed in this paper, consists in the early detection of change points or anomalies, that may be caused by a malicious use of the system for instance. Although online anomaly detection is now the subject of much attention in the data science literature, motivated by crucial industrial applications such as predictive maintenance and health monitoring of complex infrastructures, the rapidly changing environment inherent to most IoT applications makes this task even more challenging. Beyond the crucial control of the false alarm rate, the quasi real-time analysis must take into account the efficiency of computing resources and the possible security risks in data transfers over the network. We propose here an architecture for analyzing IoT datastreams data and a dedicated method for on-line anomaly/novelty detection, based on nonparametric (mean discrepancy) test statistics and multiple hypothesis testing techniques. Numerical results based on experiments involving synthetic datastreams and real energy consumption datastreams provides empirical evidence of the relevance of the methodology proposed.