eIRWR paper pinpoints microservice faults 2.8x faster with 25ms runtimes

Root cause analysis (RCA) in microservice architectures is notoriously hard: when one service fails, it cascades through hundreds or thousands of dependencies, and the originating fault is often quieter than the downstream victims. Existing graph-based random-walk methods restart from the raw anomaly vector, which gets dominated by loud downstream services, leaving most of the localization signal unused. In a new arXiv preprint, researchers Saiful Khan and Afrah Farea tackle this with eIRWR (Enhanced Iterative Random Walk with Restart). The method introduces three key innovations: power-law teleportation sharpening to concentrate restart mass on the most suspicious nodes, augmented transition matrices with self-loops and backward edges so probability accumulates at cascade sources, and an outer loop that refines beliefs iteratively.

Notably, the authors also show that the "resilience damping" commonly applied to transition matrices is mathematically equivalent to raising the restart probability, so they benchmark against a properly tuned Personalized PageRank (PPR) rather than a default configuration—raising the bar for fair comparison. On three large-scale topologies from the Alibaba Microservice Trace Dataset (12K-25K nodes), eIRWR achieves a Mean Reciprocal Rank (MRR) of 0.75 at moderate root-cause visibility, a 2.8x improvement over the best aggregate-metric baseline and well above the tuned PPR. At high visibility it reaches MRR=0.94, while running in under 25ms on graphs with 17,000 nodes. That performance makes eIRWR suitable for online deployment, where incident responders need to pinpoint the faulty service before the cascade spirals further.