Recoleta
GitHub
Source note

UModel: An Agent-Ready Observability Data Modeling Method at Scale

Authors
Changhua Pei, Zheyuan Li, Zexin Wang, Hang Cui, Xiaohui Nie, Qi Zhou, et al.
Source
arXiv
Published
Collection
Software Intelligence
AiopsObservabilityRoot Cause AnalysisLLM AgentsKnowledge Graph
Open arXivMarkdown source

Summary

UModel organizes observability data as linked objects so LLM agents can query telemetry, topology, and operations knowledge during root cause analysis. The paper claims an 8% gain in root cause localization on the AIOps 2025 Challenge dataset and reports large-scale Alibaba Cloud deployment.

Problem

  • RCA for networked systems needs logs, metrics, traces, topology, and expert knowledge in one view; current observability stacks split these across tools such as Prometheus, Elasticsearch, and tracing systems.
  • LLM agents often lack schema meaning, units, query rules, and entity relationships, so they make invalid queries or miss causal paths during diagnosis.
  • The problem matters because over 40% of real industrial faults are reported as zero-shot failures, where supervised RCA models that rely on past incidents can fail.

Approach

  • UModel converts operational resources into objects: services, pods, hosts, metrics, logs, traces, events, runbooks, and tools.
  • It links these objects in a semantic graph, using EntitySet, DataSet, EntitySetLink, and DataLink to connect topology, telemetry, and knowledge.
  • U-SPL gives agents one pipeline-style query language for schema lookup, entity search, and graph traversal instead of separate PromQL, SQL, log, and trace queries.
  • USearch handles entity and telemetry retrieval with keyword search, filtering, fuzzy matching, ranking, and top-k control.
  • GSearch lets agents walk dependency graphs, such as service call paths, to inspect likely fault propagation paths.

Results

  • On the AIOps 2025 Challenge dataset, remodeling the data with UModel improved root cause localization precision by 8% over the naive agent-based approach; the excerpt does not provide the absolute precision values.
  • The introduction also describes this as an 8% RCA accuracy improvement, but the excerpt does not clarify whether the 8% is absolute percentage points or relative gain.
  • Alibaba Cloud has deployed UModel in production for more than 1 year.
  • The deployment has served tens of thousands of users and unified thousands of standardized models across subsystems including Kubernetes, network, container services, APM, and cloud infrastructure monitoring.
  • The system is reported to handle millions of operations per second with sub-second or second-level query latency for large data retrieval.

Link