Dynamic Resource Allocation (DRA) reaches General Availability in Red Hat OpenShift 4.21, built on Kubernetes 1.34. DRA replaces the legacy device plug-in model with a structured, CEL expression-driven framework that lets pods request GPUs and accelerators by attributes like model, memory, MIG profile, and compute capability rather than simple counts. Three GA capabilities are covered with live NVIDIA A100 examples: attribute-based GPU allocation (including MIG slice targeting), prioritized device alternatives with automatic fallback, and namespace-controlled admin access to in-use devices for monitoring without disrupting workloads. The post includes full YAML manifests and real cluster output demonstrating each feature.
Table of contents
The problem: Why device plug-ins fall shortWhat is dynamic resource allocation?See it in practiceThe road to GAWhat's GA in OpenShift 4.211. Attribute-based GPU allocation2. Prioritized alternatives in device requests3. Namespace-controlled admin accessWhat's nextSort: