Kubernetes 1.35 landed. It brings in-place pod resize and AI-optimized scheduling. These are not small features. They change how you manage resources and plan...
Kubernetes 1.35: In-Place Pod Resize and AI-Optimized Scheduling — What Cluster Admins Must Do
Kubernetes 1.35: In-Place Pod Resize and AI-Optimized Scheduling — What Cluster Admins Must Do
Kubernetes 1.35 landed. It brings in-place pod resize and AI-optimized scheduling. These are not small features. They change how you manage resources and plan upgrades.
Why this matters
Kubernetes evolves fast. Every new release asks you to rethink cost, latency, and reliability. In-place pod resize saves restarts. AI-optimized scheduling helps pack and spread workloads smarter. Together, they let clusters run leaner and faster.
Key new features in Kubernetes 1.35
- In-place pod resize: change CPU and memory for a running pod without recreating it.
- AI-optimized scheduling: scheduler plugins use machine learning signals to place pods for cost and performance.
- Better observability hooks: metrics and events for resizing and AI decisions.
- Safety and policy controls: limits, policies, and rollbacks for both features.
In-place pod resize: simple, but powerful
Before 1.35, resizing often meant deleting and recreating pods. That caused downtime. Now you can scale CPU or memory in place when the container runtime and the app support it.
Why it matters:
- Less churn. Pods stay running.
- Faster autoscaling responses.
- Simpler day-two operations.
Gotchas and limits:
- Not all workloads can accept in-place changes. Some apps must reinitialize memory maps.
- The node and runtime must support the feature gate and kernel behaviors.
- You still need resource quotas and limits to avoid noisy neighbors.
AI-optimized scheduling: a new mind for the scheduler
Kubernetes 1.35 introduces scheduler plugins that use AI signals. They look at historical metrics, topology, and cost signals. Then they suggest better node placements.
Why it matters:
- Better bin-packing reduces cloud bill.
- Smoother latency for latency-sensitive apps.
- Smarter spread for stateful sets and GPUs.
What to watch for:
- Training data quality. Bad signals make bad placements.
- Explainability. Know why the scheduler moved a pod.
- Integration with existing affinity and taints.
Upgrade checklist for cluster admins
- Read the release notes. Always.
- Test in staging. Use traffic that mirrors production.
- Validate your runtimes and kernels for in-place resize.
- Audit policies for resize and scheduler plugins.
- Monitor metrics and alerts for unexpected resource shifts.
- Have a rollback plan. Keep the previous control plane accessible.
A quick story from my time in the trenches
At Microsoft, I ran a fleet with bursty jobs. We chased lower latency and lower cost. We tried in-place resize in a canary pool. It cut pod churn by half. It also exposed an app that reallocated memory poorly. We fixed the app and then saved compute. I remember the team cheering. I was Ernie Castro then, hair grayer, coffee in hand. 😊
Tuning AI scheduling felt like teaching a young apprentice. The model learned bad habits at first. We gave it better data. Then it began to place pods where engineers expected. Reliability improved.
Practical tips
- Start small. Enable features in a subset of nodes.
- Use metrics: CPU steal, memory pressure, and scheduler decision latency.
- Keep quotas tight. Let in-place resize work within bounds.
- Log scheduler decisions. It helps debugging and audits.
SEO keywords to keep in mind
Kubernetes 1.35, in-place pod resize, AI-optimized scheduling, K8s upgrade, cluster admin, Kubernetes release, scheduler plugins, Kubernetes resize, cloud cost optimization.
Final thoughts
Kubernetes 1.35 gives you tools to run faster and cheaper. But tools do not replace judgment. Test. Observe. Roll out slowly. Pray, if you need to. I often do—quietly. The work is hard. The gains are real.
— Ermenildo V Castro (Ernie Castro) 😊