Due to the growth of Big Data processing technologies and cloudcomputing services, it is common to have multiple tenants share thesame computing resources, which may cause performance anom-alies. There is an urgent need for an effective performance anomalydetection method that can be used within the production environ-ment to avoid any late detection of unexpected system failures.To address this challenge, we introduce, TRACK, a new black-boxtraining workload configurationoptimization with a neural net-work driven methodology to identify anomalous performance inan in-memory Big Data Spark streaming platform. The proposedmethodology revolves around using Bayesian optimization to findthe optimal training dataset size and configuration parameters totrain the model efficiently. TRACK is validated on a real ApacheSpark streaming system and the results show that the TRACKachieves the highest performance (95% for F-score) and reducesthe training time by 80% to efficiently train the proposed anomalydetection model in the in-memory streaming platform.