For large-scale distributed systems, the knowledge component at the core of the MAPE-K loop remains elusive. In the context of end-to-end probing, fault monitoring can be re- casted as an inference problem in the space-time domain. We propose and evaluate Sequential Matrix Factorization (SMF), a fully spatio-temporal method that exploits both the recent advances in matrix factorization for the spatial information and a new heuristics based on historical information. Adaptivity oper- ates at two levels: algorithmically, as the exploration/exploitation tradeoff is controlled by a self-calibrating parameter; and at the policy level, as active learning is required for the most challenging cases of a real-world dataset.