Search papers, labs, and topics across Lattice.
This paper evaluates the effectiveness of 24 existing root cause analysis (RCA) methods (metric-based, trace-based, and multi-source) when applied to a GPU-driven LLM inference deployment under controlled failure injections. The study reveals that multi-source RCA methods achieve the highest accuracy in LLM environments, while metric-based methods exhibit fault-type-dependent performance, and trace-based methods largely fail. The findings highlight the limitations of current RCA tools for LLM systems and motivate the development of tailored analysis techniques and enhanced observability strategies.
Off-the-shelf root cause analysis tools fall flat when applied to LLM inference stacks, demanding a new generation of observability techniques.
Large language model (LLM) services have become an integral part of search, assistance, and decision-making applications. However, unlike traditional web or microservices, the hardware and software stack enabling LLM inference deployment is of higher complexity and far less field-tested, making it more susceptible to failures that are difficult to resolve. Keeping outage costs and quality of service degradations in check depends on shortening mean time to repair, which in practice is gated by how quickly the fault is identified, located, and diagnosed. Automated root cause analysis (RCA) accelerates failure localization by identifying the system component that failed and tracing how the failure propagated. Numerous RCA methods have been developed for traditional services, using request path tracing, resource metric and log data analysis. Yet, existing RCA methods have not been designed for LLM deployments that present distinct runtime characteristics. In this study, we evaluate the effectiveness of RCA methods on a best-practice LLM inference deployment under controlled failure injections. Across 24 methods (20 metric-based, two trace-based, and two multi-source), we find that multi-source approaches achieve the highest accuracy, metric-based methods show fault-type-dependent performance, and trace-based methods largely fail. These results reveal that existing RCA tools do not generalize to LLM systems, motivating tailored analysis techniques and enhanced observability, for which we formulate guidelines.
