Contrast enables multi-dimensional distributed trace comparison for faster diagnosis

Diagnosing anomalies in distributed systems often requires comparing traces across multiple dimensions, but existing tools struggle to capture structural, temporal, critical-path, and semantic differences simultaneously. Contrast, a new system from researchers Vaastav Anand, Rodrigo Fonseca, Jonathan Mace, and Antoine Kaufmann, addresses this with the Trace Projection Object (TPO)—a mergeable representation that preserves these properties while enabling efficient, on-the-fly construction of arbitrary comparison sets. By decoupling trace representation from comparison semantics, Contrast allows multiple interfaces to reason selectively about specific dimensions, avoiding the rigidity of fixed difference notions.

Contrast ships with two complementary interfaces: SpectroViz, a visual tool that highlights execution differences via critical-path analysis, and Parallax, which generates natural language explanations of trace differences using LLMs. The system was validated through controlled experiments on DeathStarBench and production traces from Uber, demonstrating both effectiveness and efficiency. This separation of concerns lets operators combine visual and textual insights, progressively refining hypotheses during anomaly investigation—a significant step forward for observability in large-scale distributed systems.