Kubernetes 1.34 and the Slow Arrival of Real GPU Scheduling

Kubernetes has always scheduled CPU and memory well and specialized hardware badly. That gap mattered little when clusters ran web services. It matters a great deal now that the same clusters are expected to run inference.

Dynamic Resource Allocation

DRA gives Kubernetes a richer API for allocating specialized hardware — GPUs, accelerators, high-performance NICs — rather than treating them as opaque countable resources. For teams sharing expensive accelerators across workloads, this is the difference between careful manual partitioning and something the scheduler can reason about.

VolumeAttributesClass to GA

Modifying volume characteristics without recreating the volume graduated to stable. Anyone who has resized or re-tiered storage under a running stateful workload knows why that matters.

The upgrade planning point

Support windows move faster than most teams plan for. Treat cluster version upgrades as a scheduled quarterly obligation with a named owner, not an event triggered by an end-of-life notice. The teams that fall badly behind are rarely the ones who decided not to upgrade — they are the ones for whom it was nobody’s job.