RuntimeSlicer: Towards Generalizable Unified Runtime State Representation for Failure Management

ArXi:2603.21495v1 Announce Type: cross Modern software systems operate at unprecedented scale and complexity, where effective failure management is critical yet increasingly challenging. Metrics, traces, and logs provide complementary views of system runtime behavior, but existing failure management approaches typically rely on task-oriented pipelines that tightly couple modality-specific preprocessing, representation learning, and downstream models, resulting in limited generalization across tasks and systems.