KubeCon + CloudNativeCon China 2025

A variety of applications in Kubernetes typically require higher memory or compute resources during startup—such as Java, .NET, and Node.js applications, as well as those utilizing large data processing frameworks or machine learning models—due to the need to load substantial dependencies and perform complex initialization tasks. To prevent startup failures from resource contention, these applications typically have their resource requests set based on peak startup demands. However, this often leads to resource waste after startup is complete.
To address this challenge, this session presents a queue-based approach using Karpenter. This method allows applications set resource requests based on typical usage instead of peak startup needs. It temporarily spreads applications across multiple smaller nodes during startup, preventing single-node overload. After startup, it smoothly consolidates them onto fewer but larger nodes to optimize resource usage while maintaining service stability.