Red Hat is releasing the latest version of OpenShift, which is based on Kubernetes 1.34 and CRI-O 1.34. OpenShift 4.21’s new capabilities revolve around enabling companies to run AI training jobs, containerized microservices, and virtualized applications on the same infrastructure and with the same operational model.

The company gave the example of a financial company that needs to maintain legacy virtual machines while training AI models for fraud detection—both of which would live on different systems. “With OpenShift 4.21, you can simultaneously modernize existing IT infrastructure and accelerate AI innovation on a single, cost-efficient platform that scales automatically based on real-time business demand,” Red Hat wrote in a blog post.

This release includes a Red Hat build of Kueue v1.2, an open-source resource management system for Kubernetes. This addition helps streamline AI workloads in two major ways, according to Red Hat. First, support for KubeFlow Trainer v2 in Red Hat OpenShift AI 3.2 allows data scientists to work with a single TrainJob API instead of having to maintain separate resources for each machine learning framework. Second, the Visibility API allows users to see where in the queue their batch jobs are and allows admins to identify where specific resources are oversubscribed.

Additionally, OpenShift 4.21 features the general availability of the JobSet Operator, which allows users to orchestrate distributed workloads using existing GitOps workflows, role-based access control policies, and monitoring tool

Several new capabilities were also added to the Dynamic Resource Allocation (DRA) operator to help with matching workloads to GPU hardware. Now, instead of requesting a specific GPU, users can request what they need, like a GPU with at least 40GB of VRAM, and the scheduler will query hardware attributes to match the user with the appropriate hardware. DRA also now supports namespace controlled admin access and the ability to define fallback strategies in a resource request, including their priority level.

In addition, hosted control planes can now dynamically adjust memory. Hosted control planes now integrate with VerticalPodAutoscaler (VPA), allowing components to scale to real-time memory consumption instead of estimates or node counts. They can now scale to zero during inactivity as well, during which time they will hibernate and preserve their configuration and state so they can be resumed when needed.

“Instead of predicting what you’ll need six months out, the platform observes what’s happening now and responds. Your control planes stay precisely sized for current workload. That means no performance degradation from under-provisioning, no wasted spend from over-provisioning,” Red Hat wrote.

Red Hat OpenShift Virtualization now lets organizations run VMs and containers on the same platform where they share the same networking, storage, and security layer. The company supports moving running VMs to different clusters without downtime, allowing operators to perform cluster maintenance, rebalance resources across regions, or migrate workloads to newer hardware without service interruption.

Red Hat is also introducing an IPv6-only control plane and secondary network support, which is beneficial for companies that are running out of IPv4 addresses.

“By supporting IPv6 across both the cluster’s core management layer and its secondary interfaces, OpenShift Virtualization ensures that high-density deployments can scale indefinitely while meeting strict government and telecommunications compliance mandates. This transition not only streamlines network routing and improves end-to-end security but also future-proofs infrastructure for the next generation of globally connected services,” the company wrote.

More information on Red Hat OpenShift 4.21 is available here.