This episode explores the challenges and strategies of running Kubernetes at scale, specifically focusing on LinkedIn's experience. Against the backdrop of LinkedIn's transition from a proprietary containerization system to Kubernetes, the discussion highlights the complexities of managing stateful workloads, including databases, on a bare-metal infrastructure. More significantly, the engineers detail their custom-built solutions, such as a generic stateful workload operator and an application control manager, which allow for granular control over maintenance and updates, even for stateful systems. For instance, the application control manager allows for communication with the database to coordinate maintenance without disrupting operations. As the discussion pivoted to infrastructure management, the team described their use of Kubernetes as an orchestration layer for their bare-metal servers, enabling programmatic management of their data center inventory. In contrast to using standard Kubernetes components like KubeDNS, LinkedIn leverages a custom networking stack for optimal performance. What this means for other organizations is that managing Kubernetes at extreme scale requires a deep understanding of the system and a willingness to build custom solutions to address specific needs and limitations of off-the-shelf components.
