Work

From Symptoms to Root Causes: Uncovering Failure Propagation Paths from Multimodal Observability Data in Microservices

Year: 2026
Type: conference-paper
Abstract: Identifying root causes in modern microservice architectures is a major challenge: a fault in one service propagates through dependent services before surfacing as a symptom, requiring operators to examine large volumes of metrics, logs, and traces to identify its root cause. We present UPROOT, a multimodal root cause analysis (RCA) pipeline that combines metric-level causal discovery, trace-derived topology constraints, semantic log-to-metric mapping, and propagation-chain extraction to produce both a ranked root-cause hypothesis and an interpretable explanation connecting the observed symptom to the candidate root cause. UPROOT supports both statistical and LLM-based causal discovery backbones, enabling the first headto-head comparison of the two paradigms on a fine-grained microservice RCA task. Our experimental results on the RCAEval benchmark show that the two backbones achieve comparable performance within the UPROOT pipeline. Moreover, UPROOT outperforms existing RCA approaches for several fault types, particularly the challenging DELAY fault, with gains of +0.15 over TORAI and +0.30 over BARO. In addition, it generates compact failure propagation chains ( l ≈ 2, depth max ≤ 3) that facilitate root cause analysis for operators.
Source: HAL
Language: English
Open Access status: green
OpenAlex ID: W7215179216
FWCI:
Cites:
Cited by:
Related to: 10