Are GNNs Actually Effective for Multimodal Fault Diagnosis in Microservice Systems?

Graph Neural Networks (GNNs) are widely used for fault diagnosis in microservice systems, but their true benefit is often conflated with that of complex preprocessing pipelines. To isolate the GNN's contribution, we propose DiagMLP, a minimal, topology-agnostic MLP baseline. We conduct an ablation study by replacing GNN modules with DiagMLP in existing state-of-the-art frameworks. Across five datasets, this simple baseline achieves performance parity with GNN-based methods in fault detection, localization, and classification. These findings challenge the assumption that GNNs are indispensable, suggesting their contribution is marginal and that performance is primarily driven by preprocessing that already encodes critical dependency information. Our work advocates for a systematic re-evaluation of model complexity and the adoption of rigorous baselines to validate future innovations.