There's a Kubernetes cluster out there right now burning ten thousand dollars a month on GPU nodes that sit idle sixty percent of the time. Why? Because the scheduler can't say "only schedule pods on nodes with MORE than four GPUs." It's 2026, and our scheduler still can't count. But that's about to change.

In this episode, we dive deep into two alpha features in Kubernetes 1.35 that represent a fundamental shift in how Kubernetes handles scheduling and storage:

**Extended Toleration Operators (KEP-5471)** - Finally, numeric threshold-based scheduling with taints. New Gt (greater than) and Lt (less than) operators let you express "I can tolerate risk up to 5%" or "schedule me on nodes with at least 4 GPUs."

**Mutable PersistentVolume Node Affinity (KEP-5381)** - Storage topology that adapts to reality. When you migrate volumes between availability zones, you no longer need to recreate pods and PVs - just update the nodeAffinity.

Plus platform engineering news: - OpenEverest: Percona's database platform goes open governance - GKE Agent Sandbox: Kernel-level isolation for AI agent code execution - MongoBleed (CVE-2025-14847): Critical vulnerability with 87,000 exposed servers - Predictive capacity planning and the shift from reactive to proactive infrastructure

This is Kubernetes evolving from reactive feedback loops to truly predictive infrastructure.

Listen on the web: https://platformengineering.org/podcasts/00087-kubernetes-upcoming-features-deep-dive