Multi-source KPIs’ root cause localization in online service systems

Root cause localization is challenging because of the large number of monitoring metrics and the many types of faults in an online service system extended by a microservices architecture. Previous work has located root causes by constructing fault dependency graphs, but most faults occur repeatedly, there are many system indicators, and there is an imbalance in fault data. Aiming at the problem of a large number of system key performance indicators (KPIs) and imbalanced fault data, this paper proposes a multi-source KPI root cause locating framework ST-RF. Firstly, a multi-index fusion feature vector (MFFV) is constructed based on Tsfresh feature extraction. Secondly, random forest is used to classify MFFV, and SMOTE is used to generate new samples using existing samples to solve the problem of data imbalance. Evaluations on semi-synthetic and real data sets show that ST-RF achieves better Top-k accuracy and fewer average searches.