<p>To address the pervasive challenge of distribution shifts in microservice monitoring, this paper proposes the Spatio-Temporal Dual-Graph Frequency-Decoupled Framework (ST-DFF) for robust Key Performance Indicator forecasting. While accurate Key Performance Indicator prediction is critical for system stability, traditional models often falter under evolving workloads due to their reliance on static, Independent and Identically Distributed data assumptions. ST-DFF overcomes these limitations by decoupling the system’s stable dynamics from its variant ones. Spatially, the framework disentangles the system’s underlying structure from transient, context-driven correlations by concurrently learning a stable dependency graph and a dynamic association graph. Temporally, each Key Performance Indicator signal is decomposed into its low-frequency trend and high-frequency fluctuations. An enhanced Transformer encoder then fuses these representations using a novel dual-path graph convolutional network; one path is dedicated to trends, the other to perturbations. Both paths operate over the learned dual-graph structure, and a gating mechanism adaptively weighs their outputs. The resulting architecture can thus balance the robustness of stable, structural patterns against the flexibility required by contextual changes, leading to more resilient predictions. On real-world microservice datasets, extensive experiments confirm that ST-DFF outperforms contemporary state-of-the-art methods, establishing its competitive accuracy and robustness in environments prone to distribution shifts.</p>