Service dependency modeling and failure propagation prediction in distributed systems based on graph neural networks

The widespread adoption of microservices architecture has introduced unprecedented complexity in service dependencies within distributed systems, where traditional monitoring and prediction methods struggle to effectively address dynamic topology changes and cascading failure risks. This paper proposes a comprehensive graph neural network-based framework for accurate service dependency modeling and failure propagation prediction. The framework adopts a layered architectural design comprising data collection layer, graph construction module, GNN processing engine, and prediction output interface. The core innovation lies in a specialized multi-layer graph neural network architecture that integrates attention mechanisms, temporal modeling, and message passing algorithms to capture complex interaction patterns and dynamic dependency evolution among services. The failure propagation prediction algorithm employs stochastic process approaches to model cascade events, utilizing dynamic programming techniques to calculate propagation path probabilities and enable real-time risk assessment and impact quantification. Experimental validation encompasses diverse real-world distributed system environments with service scales ranging from 50 to 1500 components. Results demonstrate significant superiority over existing baseline methods, achieving 94.1% precision, 92.3% recall, and 93.2% F1-score, while reducing failure detection time by 63.2% and processing latency by 64.5%. Preliminary pilot deployment studies in controlled enterprise environments demonstrate the framework’s potential applicability, with observed 47% reduction in unplanned downtime and system availability increase from 99.91% to 99.97% (+ 0.06% points) during limited deployment periods. These results represent observational findings under specific operational conditions and warrant further systematic validation.