本提案的主题是探索利用大语言模型（LLMs）进行Kubernetes集群管理的可能性。我们的目标是希望将使集群用户能够使用自然语言与集群进行交互，提高操作效率，并允许SRE使用AI来识别和解决集群问题。本次主题分享将讲述我们真实的探索经历，比如利用大语言模型帮助用户实现常规的集群查询“kubectl get”。当然，我们也将讨论到大模型目前的能力缺陷和瓶颈，基于我们的工作职能，在一个强确定性的环境中，如何能够去消除模型本身的不确定性。希望通过这次分享，能够让更多的人辩证的看待大模型的应用场景，也期待能给参会者全新的启发。

This proposal outlines our efforts to operate Kubernetes clusters using large language models (LLMs). This will enable cluster users to interact with clusters using natural language, improve operation efficiency, and allow SREs to use AI to identify and resolve cluster issues. Our design principles: - Start with replacing "kubectl get" - Use local LLM models to avoid data leaks - Iterate quickly to gather user feedback and empower the LLMs. We implemented the proposal by: - Designing training data to perform supervised fine-tuning, allowing the LLMs to learn to call our APIs to query cluster data. - Using a checklist before deploying LLM bots to multiple internal channels for production use. - By combining LLM with traditional AIOps techniques, we enabled the LLMs to detect cluster issues and facilitated cluster admins to resolve them. Finally, we share our progressive report of using LLMs with Kubernetes and propose a few open-questions for future discussions.