Fast and Robust Localization of Multi-dimensional Root Causes in Online Service

The precise localization of root causes for multi-dimensional attributes in online service failures is crucial for effectively mitigating the manual operational challenges faced by service-operating enterprises. Existing research encounters suboptimal localization results due to an increase in prediction residuals. To address these challenges, this paper introduces AFS, a novel multi-dimensional root cause localization method. AFS ensures accurate identification of low-deviation as score anomalies and recognizes anomalies of varying orientations. Initially, a one-dimensional DPC clustering is employed for time-efficient processing, clusters of anomalies from noisy and outlier-ridden data. Subsequently, high and low-deviation score clusters are individually treated, and Occam’s razor principle is applied to determine the final set of root causes. Lastly, an enhanced explanatory power formula in the hierarchical search process allows for anomalies to be traced from different directions. Experimental results demonstrate that AFS surpasses the baseline model in terms of F1 score for the high prediction residual section in semi-synthetic dataset B, with an average localization time remaining within 10 seconds.