Topo-MFP: Topology-aware multimodal propagation learning for fault detection and root cause localization in microservice systems
TL;DR
Topo-MFP is proposed, a topology-aware multimodal propagation learning framework for service-level fault detection and root cause localization that combines heterogeneous observability data with service dependency information in two stages and provides diagnostic information beyond random ranking, local anomaly scores, and raw invocation topology.
Abstract
Fault detection and root cause localization in microservice systems are challenging because runtime failures may propagate across dependent services and appear heterogeneously in metrics, logs, and traces. Existing approaches often rely on single-modality observations, local anomaly scores, or static service topology, which may be insufficient to distinguish root-cause services from downstream affected services. To address this problem, this paper proposes Topo-MFP, a topology-aware multimodal propagation learning framework for service-level fault detection and root cause localization. Topo-MFP integrates heterogeneous observability data with service dependency information in two stages. For fault detection, it constructs a semantic-transition log template graph by combining log template semantics with transition statistics, and learns topology-aware service representations from dynamic service dependency graphs through spatio-temporal graph attention. For root cause localization, it learns diagnosis-oriented multimodal propagation graphs from metrics, logs, and traces. The modality-specific propagation graphs are fused using cross-modal consistency and a topology-aware soft sparsity prior, where different penalties are assigned to topology-adjacent and topology-distant service pairs. Root-cause candidates are then ranked by symptom-to-source reverse random walk with restart over the learned propagation graph. Experiments are conducted on AIOps2022 and a TrainTicket-based TT dataset. Over five random seeds, Topo-MFP achieves mean F1-scores of 0.949 and 0.951 for fault detection, and mean Hit@1 scores of 0.872 and 0.901 for root cause localization on AIOps2022 and TT, respectively. Comparisons with Random, Local-Score, and Topology-RWR baselines show that the learned multimodal propagation graph provides diagnostic information beyond random ranking, local anomaly scores, and raw invocation topology. Ablation and sensitivity analyses further indicate the contributions of semantic log modeling, cross-modal consistency, and topology-aware regularization. The learned propagation graph is used as a diagnosis-oriented ranking representation rather than as a verified causal graph.