Wherefore Art Thou? Provenance-Guided Automatic Online Debugging with Lumos

A team of researchers from Princeton and other institutions has introduced Lumos, a novel framework designed to tackle the notoriously difficult problem of debugging complex, in-production distributed systems. Modern systems, with their myriad concurrent components, often exhibit non-deterministic bugs that evade offline testing. When these bugs surface, their symptoms (like a failed API call) can be far removed from their root cause (like a corrupted database entry hours earlier). Existing tools fail to automatically capture the necessary evidence without imposing significant performance penalties, forcing developers into a manual and arduous evidence-collection process.

Lumos addresses this by automatically exposing application-level bug provenance—the complete chain of computational events linking a symptom back to its origin. The key innovation is its use of static analysis to perform dependency-guided instrumentation. This allows Lumos to intelligently identify and selectively record only the program state relevant to a bug's history. This on-demand, lightweight recording approach provides developers with the evidence needed to pinpoint root causes while maintaining low runtime overhead, and it can work effectively after observing just a few occurrences of a bug.