Huawei's Log-Insight diagnoses microservice incidents 1,000x faster

Huawei researchers have introduced Log-Insight, a production-deployed system that automates root cause analysis (RCA) for microservice incidents. The system tackles a fundamental scaling problem: a single 30-minute incident can generate over two million log lines (1.2 billion characters), far exceeding the context windows of standard LLMs. Existing approaches—template-based parsers, deep-learning detectors, or raw LLM pipelines—either lack semantic reasoning, offer black-box signals, or suffer from context overflow and domain hallucination.

Log-Insight's core innovation is a six-stage symbolic pipeline that mirrors how a skilled SRE manually triages: sampling, schema understanding, pattern clustering, statistical anomaly ranking, and more. This reduces millions of raw events by 1,000–7,000x while preserving statistically significant failure signals. The pre-ranked evidence dossier is then handed to an LLM, which synthesizes a hypothesis report. Evaluated on 11 historical production incidents (110 runs, SRE-validated ground truth), Log-Insight achieves MRR=0.790 and returns the correct root cause within the top-3 hypotheses in over 90% of runs, all in under a minute of latency. Operators especially valued the Forensic Evidence section—exact log templates and skew statistics—which shifted the system's role from opaque oracle to investigative assistant.