MetaKube: An Experience-Aware LLM Framework for Kubernetes Failure Diagnosis Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates. Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates ISBN: 979-8-4007-2307-0/2026/04 DOI: 10.1145/3774904.3792631 CCS: Networks Network management
Wei Sun, Ting Wang, Xinran Tian, Wanshun Lan, Xuhan Feng, Haoyue Li, Fangxin Wang · 2026
Read MetaKube: An Experience-Aware LLM Framework for Kubernetes Failure Diagnosis Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates. Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates ISBN: 979-8-4007-2307-0/2026/04 DOI: 10.1145/3774904.3792631 CCS: Networks Network management by Wei Sun, Ting Wang, Xinran Tian, Wanshun Lan, Xuhan Feng, Haoyue Li, Fangxin Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-23580.txt.
Introduction
Sun and six co-authors present MetaKube, a Kubernetes failure diagnosis system that learns from past fixes. Older language model tools lean on fixed knowledge and never improve. MetaKube stores past resolutions as patterns in an episodic memory network. A meta-cognitive controller picks fast pattern matching or deep causal search by familiarity. Its local 8B model trains on 7,000 fault cases for privacy. On 1,873 real scenarios, scores jump from about 51 to over 90 points. Memory alone adds a clear share of that gain.
Catalog record
| CREATOR | Wei Sun, Ting Wang, Xinran Tian, Wanshun Lan, Xuhan Feng, Haoyue Li, Fangxin Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 57148 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.23580
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MetaKube: An Experience-Aware LLM Framework for Kubernetes Failure Diagnosis Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates. Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates ISBN: 979-8-4007-2307-0/2026/04 DOI: 10.1145/3774904.3792631 CCS: Networks Network management Wei Sun, Ting Wang, Xinran Tian, Wanshun Lan, Xuhan Feng, Haoyue Li, Fangxin Wang Abstract . Existing LLM-based Kubernetes diagnostic systems cannot learn from operational experience, operating on static knowledge bases without improving from past resolutions. We present MetaKube, an experience-aware LLM framework through three synergistic innovations: (1) an Episodic Pattern Memory Network (EPMN) that abstracts diagnostic patterns from historical resolutions and provides confidence-calibrated retrieval for both rapid pattern matching and guided causal exploration, (2) a meta-cognitive controller that dynamically routes between intuitive and analytical pathways based on problem familiarity, optimizing the trade-off between speed and depth, and (3) KubeLLM, a locally-deployable 8B model enhanced through domain-specific post-training on our 7,000-sample Kubernetes Fault Resolution Dataset. Evaluation on 1,873 real-world scenarios demonstrates MetaKube transforms Qwen3-8B from 50.9 to 90.5 points, approaching GPT-4.1 performance while ensuring complete data privacy. EPMN contributes 15.3% improvement through experiential learning, with continuous learning experiments showing progressive gains as the system accumulates operational knowledge. The source code and related resources are available at https://github.com/MetaKube-LLM-for-Kubernetes-Diagnosis/MetaKube. † † cc-license: by 1. Introduction Kubernetes ( Burns et al. 2016 ) has become the critical infrastructure orchestrating containerized applications across global cloud deployments. However, its operational complexity in production environments poses significant challenges for fault diagnosis ( Barletta et al. 2024 ) . Modern clusters contain thousands of containers with intricate dependencies across pods, services, and custom resources, making it extremely difficult for operators to trace failure propagation and identify root causes from scattered monitoring data ( Kannaiah 2024 ) . Large Language Models (LLMs) present transformative potential for Kubernetes fault diagnosis through their reasoning capabilities and knowledge integration ( Ghorab et al. 2025 ) . Recent work ( Guu et al. 2020 ) ( Zhang et al. 2024a ) explores retrieval-augmented generation (RAG) to enhance LLMs with domain-specific knowledge, enabling conversational troubleshooting interfaces that interpret error messages and recommend remediation strategies. These initial efforts demonstrate promising results in reducing the expertise threshold for cluster management ( Tippins et al. 2024 ) . However, deploying solutions based on LLMs in production Kubernetes environments faces three fundamental challenges ( Boateng et al. 2025 ) . First , existing RAG systems cannot learn from operational experience—they operate on static knowledge repositories without feedback mechanisms ( Li et al. 2024 ) , treating each diagnostic session in isolation and ignoring valuable insights from historical resolutions. Second , the Kubernetes ecosystem suffers from acute scarcity of high-quality diagnostic data ( Hrusto et al. 2025 ) , with troubleshooting knowledge fragmented across documentation, GitHub issues, and proprietary runbooks, severely constraining both model training and retrieval effectiveness. Third , enterprise deployments require strict data privacy, prohibiting external API usage for sensitive cluster information ( Gupta 2024 ) . This creates a difficult trade-off: large open-source LLMs (70B+ parameters) impose prohibitive computational overhead, while smaller LLMs (<10B parameters) lack the specialized reasoning capabilities needed for accurate diagnostics ( Savage et al. 2025 ; Almubark 2025 ) . To address these fundamental challenges, we present MetaKube, a cognitive architecture that synergistically integrates episodic memory networks, specialized language models, and causal knowledge graphs within a unified diagnostic framework. MetaKube implements dual reasoning pathways: an intuitive pathway leveraging our novel Episodic Pattern Memory Network (EPMN) for rapid pattern recognition, and an analytical pathway utilizing KubeGraph for systematic causal exploration. A meta-cognitive controller employing confidence-calibrated assessment dynamically orchestrates pathway selection based on problem complexity and memory similarity scores. Our contributions are threefold: • First , we propose the MetaKube architecture to address the critical deficiency of learning from operational experience. MetaKube integrates three synergistic components: (1) an Episodic Pattern Memory Network (EPMN) that continuously abstracts and refines diagnostic patterns from resolution outcomes, (2) a specialized language model (KubeLLM) that adapts through post-training, and (3) a dynamic knowledge graph (KubeGraph) that expands its causal structure through discovered relationships, collectively enabling the system to improve diagnostic accuracy through operational feedback. • Second , to overcome the extreme scarcity of high-quality diagnostic data in the Kubernetes ecosystem, we construct comprehensive diagnostic resources by aggregating and structuring knowledge from diverse online sources including documentation, forums, and issue trackers. This effort produces both a Kubernetes knowledge graph encompassing causal relationships and operational constraints, and a curated fault resolution dataset covering common failure patterns. • Third , considering the stringent data privacy requirements in production Kubernetes environments, we adapt the locally deployable Qwen3-8B ( Yang et al. 2025 ) model as MetaKube’s backbone, employing advanced post-training techniques—specifically domain-specific supervised fine-tuning on our large-scale curated Kubernetes Fault Resolution Dataset—to significantly elevate its diagnostic reasoning and resolution capabilities to levels comparable to commercial API services. This approach ensures complete on-premise deployment, keeping sensitive cluster data securely within organizational boundaries while delivering state-of-the-art diagnostic performance. Experiments on KubeFault, our dataset of 1,873 Kubernetes fault scenarios, validate MetaKube’s effectiveness: it improves Qwen3-8B by 40.6 points, approaching GPT-4.1 performance while ensuring data privacy. 2. Related Work 2.1. Kubernetes Fault Diagnosis Traditional Kubernetes fault diagnosis approaches range from rule-based systems to machine learning techniques ( Kaul 2020 ; Boluda-Prieto et al. 2025 ) . Statistical methods like CloudRanger ( Wang et al. 2018 ) correlate metrics for root cause analysis, while graph-based approaches including MicroRCA ( Wu et al. 2020 ) trace anomaly propagation patterns. Commercial platforms (Datadog, Dynatrace) and open-source tools (Prometheus, Jaeger) provide comprehensive monitoring but lack semantic understanding of failure causality. Recent AIOps systems ( Dang et al. 2019 ) combine multiple signals yet remain fundamentally reactive, requiring extensive manual feature engineering and struggling with novel failure modes. These approaches collectively fail to capture the adaptive learning capabilities demonstrated by expert operators who continuously refine their diagnostic techniques based on prior experiences ( Fritzsche et al. 2024 ) . MetaKube transcends these limitations through episodic memory that continuously learns diagnostic patterns from operational experience, eliminating predefined rules while adapting to emerging failure modes. 2.2. LLMs for System Operations Recent work explores LLMs for system administration, with OpsGPT 1 1 1 Broadridge Financial Solutions, Inc., “Broadridge Launches GenAI-Powered OpsGPT to Transform and Optimize Trading in a T+1 Environment,” press release, January 11, 2024, https://www.broadridge.com/press-release/2024/broadridge-launches-genai-powered-opsgpt . automating incident mitigation and X-Lifecycle ( Goel et al. 2024 ) enabling conversational cloud management. Studies by ( Maeno et al. 2024 ) reveal both promise and limitations in operational contexts. Advanced prompting techniques like chain-of-thought ( Wei et al. 2022 ) and ReAct ( Yao et al. 2023 ) improve diagnostic reasoning but cannot accumulate experience across sessions. While these general-purpose models generate semantically plausible solutions, they cannot guarantee operational feasibility in production environments, potentially suggesting configurations incompatible with specific Kubernetes versions or cluster constraints ( Carrión 2022 ; Furnadzhiev et al. 2025 ) . Additionally, concerns about data privacy arise when sensitive system logs and error messages are processed by external LLM services, posing significant risks for enterprise deployments ( MALIK 2024 ; Rathod et al. 2025 ) . These privacy concerns are particularly acute in enterprise