Kubernetes clusters are often over-provisioned at the pod level, yet some workloads still experience CPU throttling. This usually comes from static CPU requests that are sized for worst-case behavior and rarely revisited once workloads are running. Recommendation-based systems exist, but applying frequent changes in production has traditionally been difficult due to restarts and operational risk. In-place pod resize makes it possible to update CPU requests without recreating pods, and PSI (Pressure Stall Indicator) metrics provide a clearer signal of real CPU contention. Together, these enable a different approach: making small, frequent, just-in-time corrections instead of relying on long-term prediction. In this talk, we share how we built CruiseKube, a runtime CPU optimization system that continuously right-sizes pods in place and takes node context into account. We discuss the key design choices, tradeoffs, and lessons learned from running this approach in real clusters.
Raman Tehlan is a Cloud Native Consultant at Zurich Lab with over five years of experience building and operating distributed systems at scale. His work focuses on Kubernetes, cloud infrastructure, and developer platforms, with an emphasis on practical, production-ready solutions. He is also an active community contributor and an organizer of PyData Delhi, where he helps foster knowledge sharing around data, systems, and open-source technologies.
Shubham Rai is a Staff Software Engineer at TrueFoundry with 10 years of experience building Kubernetes and distributed systems. A contributor to KubeElasti, he focuses on reliability, cost efficiency, and DR at scale, sharing practical insights from real production systems.
