A Graph Transformer-Based Framework for Multi-Modal Failure Diagnosis in Microservice Systems

Failure diagnosis in microservice systems is difficult due to the complex, multimodal nature of telemetry data. Existing methods use metrics, logs, and traces, but rely on message-passing graph neural networks with limited ability to model global context. This study introduces TransTVDiag, which replaces TVDiag's GraphSAGE encoder with a Graph Transformer enhanced with structural encodings for microservice correlation graphs. The study provides four main contributions: (1) adapting Graphormer to multimodal alert graphs with degree centrality and shortest-path encodings, (2) analyzing these encodings in microservice diagnostics, (3) quantifying the individual and joint impact of metrics, logs, and traces, and (4) demonstrating robustness to missing or noisy alerts. TransTVDiag improves root cause localization in hit ratio by 5.3%, in ranking quality by 3.5%, while reducing inference time by 83.8% over TVDiag. The study also outlines how model outputs can be made more actionable for operators, showing that Graph Transformers offer an accurate and efficient alternative for multimodal failure diagnosis.