Restart-Free Vertical Scaling for Kubernetes Pods
summary
Design enables restart-free vertical scaling of pod CPU/memory by treating a resource delta as a new scheduling decision: API patch writes an annotation, scheduler checks node cache, kubelet revalidates, then the resources section is patched and containers continue running. Policy options include in-place only, in-place-preferred with eviction, and force restart; failures use event-driven retries and backoff. Demo shows live patch from 1Gi to 2Gi without restart; integration with VPA, metrics-server and Prometheus for recommendations; gene-sequencing customer cites ~50–60% resource savings.
AI-written summary · about
chapters (13)
- 0:00Introduction and motivation
- 1:47Customer scenario
- 3:52Kubernetes pod scheduling
- 6:04Designing the solution
- 8:20Implementation workflow
- 10:24Kubelet and failure handling
- 12:08Policy controls and restarts
- 14:23Event-based failure handling
- 15:22Integration with VPA
- 17:34Memory spikes and OOM
- 19:18Live demonstration
- 23:23Future work and conclusions
- 26:23Q&A session
related talks
34mPredictive autoscaling patterns in Kubernetes
36mTaming the Beast: Advanced Resource Management With Kubernetes
29mRight-Size Your Kubernetes Deployments
13mUsing CPU & Memory Hotplug for VM Vertical Scaling
31mFailing to Autoscale?
33mResize Your Pods In-Place With Deterministic eBPF Triggers
34mIntroduction to Autoscaling
08m⚡ Fast Restarts, Not Just Fast Starts: Accelerating Pod Recovery