Source-linked AI summary
Hypothesize-Then-Verify: Speculative Root Cause Analysis for Microservices with Pathwise Parallelism
Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Pei Xiao, Ying Li
TL;DR
Microservice RCA must handle complex, cascading runtime anomalies, while existing LLM-based methods face limited exploration diversity and slow inference. SpecRCA uses broad hypothesis drafting, parallel verification, and diagnosis synthesis with models of different sizes. Preliminary AIOps 2022 experiments report higher accuracy and faster diagnosis than existing approaches, while further datasets and implementation refinement remain planned.
Problem
Existing LLM-based RCA methods have limited exploration diversity and depend heavily on large-scale models, constraining accuracy and inference speed.
Method
SpecRCA generates broad candidate root causes, independently verifies them in parallel, and synthesizes the verification results using models of different sizes.
Results
12.14% higher failure-localization accuracy and diagnosis reports within 20 seconds are reported on the AIOps 2022 dataset.
Takeaways & Limitations
SpecRCA shows potential for accurate and efficient RCA in complex microservice environments.
Takeaways & Limitations
The evaluation is preliminary, requires additional datasets, and identifies the Diagnosis Synthesizer as the primary inference-speed bottleneck.
Abstract
from arXiv · showhide
Microservice systems have become the backbone of cloud-native enterprise applications due to their resource elasticity, loosely coupled architecture, and lightweight deployment. Yet, the intrinsic complexity and dynamic runtime interactions of such systems inevitably give rise to anomalies. Ensuring system reliability therefore hinges on effective root cause analysis (RCA), which entails not only localizing the source of anomalies but also characterizing the underlying failures in a timely and interpretable manner. Recent advances in intelligent RCA techniques, particularly those powered by large language models (LLMs), have demonstrated promising capabilities, as LLMs reduce reliance on handcrafted features while offering cross-platform adaptability, task generalization, and flexibility. However, existing LLM-based methods still suffer from two critical limitations: (a) limited exploration diversity, which undermines accuracy, and (b) heavy dependence on large-scale LLMs, which results in slow inference. To overcome these challenges, we propose SpecRCA, a speculative root cause analysis framework for microservices that adopts a \textit{hypothesize-then-verify} paradigm. SpecRCA first leverages a hypothesis drafting module to rapidly generate candidate root causes, and then employs a parallel root cause verifier to efficiently validate them. Preliminary experiments on the AIOps 2022 dataset demonstrate that SpecRCA achieves superior accuracy and efficiency compared to existing approaches, highlighting its potential as a practical solution for scalable and interpretable RCA in complex microservice environments.
1 Introduction
Microservice complexity makes timely, interpretable RCA necessary, while existing LLM-based methods remain limited by insufficient exploration diversity and slow inference. SpecRCA addresses these challenges through hypothesize-then-verify processing, pathwise parallelism, and heterogeneous models, with preliminary results reporting higher accuracy and lower latency.
- Motivation: Microservice runtime interactions can cascade failures across dependent services, making timely and interpretable RCA essential for reliability.Manual troubleshooting becomes increasingly inadequate as systems scale and business scenarios diversify.
- Challenges: Existing LLM-based RCA methods often converge on similar reasoning paths, leaving alternative hypotheses underexplored and limiting accuracy.These methods include multi-agent voting and training-time exploration enhancement.
- Challenges: Current approaches also rely heavily on large-scale LLMs or multi-agent interaction, making inference impractically slow for real-time microservice RCA.The stated challenge concerns knowledge reasoning, long-context processing, and multi-round interaction costs.
- Approach: SpecRCA uses a broad hypothesis-drafting stage, independent root-cause verification, and diagnosis synthesis within a hypothesize-then-verify framework.Each candidate is treated as a hypothesis and evaluated separately before verification outcomes are integrated.
- Approach: SpecRCA combines a lightweight drafting model, parallel fine-tuned LLM verifiers below 3B parameters, and a larger synthesizer to balance efficiency and accuracy.Only the synthesis stage relies on a larger-scale LLM.
- Results: 12.14% higher failure-localization accuracy and complete diagnosis reports within 20 seconds are reported on the AIOps 2022 dataset.The preliminary comparison reports both improved diagnostic accuracy and reduced inference latency.
2 Methodology
SpecRCA builds candidate root causes from heterogeneous metrics, traces, and logs, then verifies hypotheses with a distilled lightweight verifier and synthesizes the results into a diagnosis.
- Hypothesis Drafting: The Hypothesis Drafting module processes metrics, traces, and logs to generate candidate root causes using modality-specific scores aligned to services and fused under topology constraints.It produces a comprehensive and prioritized service-level suspiciousness ranking.
- Hypothesis Drafting: Metric analysis combines anomalous deviations with Granger causality to identify suspicious services and refine their scores using inter-service dependencies.Abnormal-period metrics are compared with normal-operation baselines, while significant causal links adjust anomalousness scores.
- Hypothesis Drafting: Trace analysis extracts the critical path with the largest latency contribution and ranks suspicious services using residual latency and trace-weighted Service-Oriented PageRank.Residual latency is defined relative to each service’s baseline mean.
- Hypothesis Drafting: Log analysis combines sequence and frequency evidence into a unified service-level measure of anomalous patterns before topology-guided integration.The framework parses logs into templates and examines missing, emerging, rare, or bursty patterns.
- Root Cause Verifier: The Root Cause Verifier uses RCALite to evaluate each candidate through self-state, upstream, and downstream verification before consolidating evidence into a diagnosis.RCALite is distilled from a high-capacity Teacher LLM, trained with supervised fine-tuning, reward modeling, and reinforcement fine-tuning.
- Root Cause Verifier: During verification, RCALite checks whether parent or child services better explain the observed anomalies, helping distinguish a candidate service from neighboring root causes.The verifier independently examines upstream influences and downstream evidence around each candidate hypothesis.
3 Preliminary Evaluation
SpecRCA is evaluated preliminarily on AIOps 2022 against RCAgent and mABC, using accuracy metrics and per-query diagnosis speed. The evaluation reports stronger accuracy and faster inference for SpecRCA.
- Evaluation Setup: SpecRCA is evaluated on the AIOps 2022 dataset against RCAgent and mABC, both based on Qwen-2.5-Plus.RCALite is distilled from Claude-3.5 using Llama3.2-3B, and the subsequent RFT process has not yet been performed.
- Accuracy: SpecRCA significantly outperforms both baselines across all reported accuracy metrics: Recall@1 61.34%, Recall@3 75.72%, Recall@5 81.63%, and MRR 62.64%.Accuracy is measured by Recall@K and MRR.
- Efficiency: 9.89 seconds per query: SpecRCA generates a diagnosis considerably faster than the baseline approaches.Speed is measured in seconds per query (s/q).
4 Conclusion
The paper concludes by presenting SpecRCA as a microservice RCA framework designed to address limited exploration diversity and reliance on large-scale LLMs through pathwise parallelism and hypothesize-then-verify.
- Conclusion: SpecRCA addresses limited exploration diversity and overreliance on large-scale LLMs in current LLM-based RCA methods.The framework is presented as a speculative root cause analysis approach for microservices.
- Conclusion: SpecRCA leverages pathwise parallelism and follows a hypothesize-then-verify paradigm.Its effectiveness is demonstrated through preliminary evaluation on a prototype implementation.
- Conclusion: The paper reports preliminary evidence for the effectiveness of this approach through evaluation of a prototype implementation.
5 Future Plans
Future work focuses on refining SpecRCA's implementation and broadening its evaluation, while addressing the Diagnosis Synthesizer's inference-speed bottleneck.
- Implementation Refinement: The authors plan to apply RFT to RCALite, which has so far undergone only SFT distillation.
- Evaluation Expansion: The authors plan additional experiments across more datasets to evaluate SpecRCA's performance more comprehensively.
- Inference Speed: The Diagnosis Synthesizer is currently the primary inference-speed bottleneck.The authors plan to use diffusion-based large language models and parallel decoding to improve throughput.