.jpg)
The evolution of Kubernetes in resource management has been constant, but the arrival of Dynamic Resource Allocation (DRA) in version 1.35 represents a qualitative leap. This feature, which has reached general availability (GA) status, allows cluster administrators and developers to request and assign hardware devices —such as GPUs, FPGAs, or network units— in a much more flexible and granular way than the old Device Plugin. Instead of relying on node labels and affinity rules, DRA introduces a model based on DeviceClass, ResourceSlice, and ResourceClaim, similar to the PersistentVolumeClaim (PVC) system for storage. This abstraction allows expressing complex needs: for example, requesting a GPU with at least 20 GiB of memory, or specifying a preference for a specific architecture, such as Ampere or Turing. Additionally, the selection logic can be defined using the Common Expression Language (CEL), opening up an almost infinite range of combinations.
For companies working with artificial intelligence and machine learning workloads, DRA greatly simplifies the management of heterogeneous hardware. In a cluster with GPUs from different generations, a data science team can deploy a model requiring a specific GPU without having to manually label nodes or write nodeSelector rules. Simply create a ResourceClaimTemplate that includes a CEL selector looking for, for example, a productName equal to 'NVIDIA RTX A5000' and, as a fallback plan, allow falling back to a Tesla T10 if none are available. This is achieved with the firstAvailable operator, which respects the order of preference. The result is a much more efficient and automated allocation, ideal for environments where resources are limited and demands vary constantly.
Another notable capability is Time Slicing under DRA, which allows multiple workloads to share the same physical GPU by segmenting usage into time intervals. Unlike the previous approach, it is not necessary to predefine how many 'slices' to divide; simply configure the sharing strategy and the interval in the ResourceClaim parameters. This is especially useful in development phases or testing environments where a full GPU per container is not required. Furthermore, starting from Kubernetes v1.36, device health reporting will be incorporated, allowing distinguishing whether a failure comes from hardware or the application, improving debugging and cluster robustness.
In this context, having a technology partner that understands both infrastructure and business needs is essential. At Q2BSTUDIO we offer development of custom applications and custom software that integrate perfectly with platforms like Kubernetes, enhancing automation and scalability. Our AWS and Azure cloud services allow deploying clusters optimized for artificial intelligence workloads, while our Power BI and business intelligence services help visualize the performance of these resources. Additionally, we implement artificial intelligence for businesses and AI agents that directly benefit from precise hardware allocations, and we reinforce security with advanced cybersecurity. All with the goal of enabling organizations to extract maximum value from their technology investments.
In short, Dynamic Resource Allocation not only improves operational efficiency but also opens the door to a new generation of cloud-native applications that demand specialized hardware. As clusters grow and workloads diversify, having a solid orchestration foundation and expert allies like Q2BSTUDIO makes the difference between having infrastructure and truly leveraging it.
