Reproduce “Dependency Aware Incident Linking in Large Cloud Systems” on OpenResearch

About this paper

Supriyo Ghosh, Karish Grover, Jimmy Wong, Chetan Bansal, Rakesh Namineni, Mohit Verma, Saravan Rajmohan

A dependency-aware incident linking framework, DiLink, improves the identification of related incidents in large cloud systems by integrating textual data with service dependency graphs using an orthogonal Procrustes alignment. The system, developed by Microsoft, achieved up to a 14% gain in F1-Score and 15% gain in Accuracy over text-only baselines across 610 services within 5 Microsoft workloads.

Key insights

• Simply concatenating multi-modal embeddings without alignment can degrade incident linking performance, as evidenced by lower scores compared to text-only baselines.
• Explicitly incorporating service dependency graph information alongside textual data substantially enhances incident linking accuracy, particularly for complex cross-service and cross-workload scenarios.
• The completeness of the service dependency graph significantly impacts linking accuracy, with performance improving as more dependency edges are included.

Summary by alphaXiv.