Multi-Granularity Dependency Modeling for Automated Fault Triage in High-Throughput Financial Transaction Systems

Abstract

High-throughput financial transaction systems constitute critical infrastructure in modern banking and financial services, processing millions of transactions per second while maintaining stringent requirements for reliability, consistency, and availability. The inherent complexity of these distributed systems, characterized by intricate interdependencies among microservices, databases, message queues, and external interfaces, presents substantial challenges in fault diagnosis and root cause analysis. This research proposes a novel multi-granularity dependency modeling framework that captures system behaviors across service-level, transaction-level, and resource-level abstractions to enable automated fault triage. The framework integrates real-time telemetry data including metrics, distributed traces, and transaction logs to construct dynamic dependency graphs that reflect evolving system states. A hierarchical fault propagation model is developed to distinguish between root causes and cascading failures, leveraging causal inference techniques and temporal correlation analysis. The proposed approach employs machine learning-based anomaly detection coupled with graph-based root cause localization algorithms to identify fault origins within seconds of symptom manifestation. Experimental evaluation on a production-scale financial transaction processing platform demonstrates that the multi-granularity approach achieves superior diagnostic accuracy compared to single-level analysis methods, reducing mean time to resolution by approximately 67% while maintaining a precision rate exceeding 92% for root cause identification. The framework provides actionable insights for automated incident response systems and contributes to the broader discourse on reliability engineering in mission-critical financial infrastructure.