For over a decade, Kubernetes operated on two rigid scheduling assumptions: workloads were bound to fixed cgroup limits at birth, and kube-scheduler evaluated every Pod as an isolated island. For complex enterprise applications, from state-heavy Java services to distributed AI/ML batch jobs, this meant cold-boot penalties, aggressive OOM kills, fragmented bin-packing, and head-of-line blocking.

Those days are officially over.

With the GA graduation of In-Place Pod Vertical Scaling (IPPR), Pod-level resource updates, VPA CPU Startup Boost, and the introduction of native Workload and PodGroup APIs for gang scheduling, Kubernetes has undergone its most significant architectural shift in years.

In this talk, we bridge the gap between in-pod runtime elasticity and cluster-wide workload awareness. You will learn:

- How In-Place Pod Resizing directly mutates cgroups without Pod recreations and how JVM runtimes (Java 21+) and GC ergonomics adapt dynamically.
- How to eliminate startup CPU taxes using VPA CPUStartupBoost (AEP-7862).
- How the native Workload API and PodGroup primitives solve all-or-nothing gang scheduling without requiring third-party secondary schedulers.
- Practical architecture patterns to combine in-place vertical scaling with group-aware scheduling for maximum density and zero-downtime performance.