In Greek mythology, it took an arduous voyage of 10 years for the mighty Odysseus to finally return home to his wife Penelope, who waited patiently all the while suffering pretenders to the throne.
And it seemed fitting that on this 10th year of its own development, Kubernetes has gotten a logo featuring a statue of Penelope, to commemorate the many trials and tribulations of creating an open source container management platform.
The logo accompanies the release Wednesday of Kubernetes 1.32, which marks the completion of a number of fierce challenges, according to release lead Frederico Muñoz, who also designed the logo.
After years of testing, this release is the first release to pass, with a 100% score, the whole suite of conformance testing. It is “a huge milestone,” Muñoz said, as the work completes a large backlog of tests that guarantee that the Kubernetes API all work in the way they are described in the specifications.
The version of Kubernetes being released also includes a revamped Dynamic Resource Allocation (DRA), a much-needed feature of the Kubernetes resource management system in this day of AI and edge workloads. The Beta release quickly addresses some limitations that surfaces after the previous DRA implementation.
Dynamic Resource Allocation Redux
When Kubernetes was launched 10 years ago, the data center servers were largely built on x86-compatible processors. These days, however, the demand for alternative architectures is skyrocketing, thanks to the use of GPUs for AI, which organizations would like to manage with Kubernetes
The idea with DRA is to make it easier to match workloads to specialized hardware, not only GPUs but also FPGAs and network adapters favored in the high-performance computing community.
DRA was initially introduced in Kubernetes 1.26 with the KEP (Kubernetes Enhancement Proposal) 3063. The implementation proved problematic, however, and is being replaced entirely with KEP 4381 in this new version of Kubernetes.
A cross-functional team of members from SIG Node, SIG Scheduling, and SIG Autoscaling did the new work.
The first implementation was incompatible with the K8s cluster autoscaling capability. It provided no insight into allocated resources and required a third-party driver to determine if the resources were available.
The work in KEP 4381 substitutes the functionality with a structured parameter model. Now, either the kube-scheduler and Cluster Autoscaler can simulate claim allocation directly, eliminating the need for a third-party driver.
One bonus to the new enhancement is that it also includes a resource health status in the Pod status, making troubleshooting easier.
“This removal will allow Kubernetes to handle new hardware requirements and resource claims more predictably, bypassing the complexities of back and forth API calls to the kube-apiserver,” the docs read.
Conformance Testing Completed
Maintained by SIG Architecture, the Kubernetes Conformance test suite verifies the core set of interoperable features that Kubernetes clusters must support, covering APIs, networking, storage, and scheduling.
The PR that made it to 100% was merged live during the Kubernetes Contributor Summit in KubeCon North America last month.
There are other Kubernetes endpoints (RBAC, for instance) that have not been tested, but this group of tests covers everything that was deemed essential — as was determined around the Kubernetes 1.27 release.
The benefits of the conformance tests are twofold. First, it verifies that Kubernetes does what its APIs were designed to do, with full fidelity.
The test suite will also verify that a Kubernetes distribution is fully compatible with the Kubernetes specification itself, Muñoz said.
“It’s a matter of being able to guarantee that with the growing popularity of Kubernetes, and the many different possible Kubernetes-based solutions out there, that we get a consistent Kubernetes behavior,” Muñoz said.
Kubernetes Conformance Testing, from API Snoop.
“Contributing to conformance testing is akin to the task of “washing the dishes” — it may not be highly visible, but it remains incredibly important. It necessitates a strong understanding of Kubernetes, particularly in the areas where the endpoints need to be tested,” Muñoz told an interviewer on the Kubernetes Contributors blog.
Other Features of Note
Overall, Kubernetes 1.32 comes with 44 enhancements. Of these, 13 have graduated to Stable, 12 are entering Beta, and 19 have entered in Alpha.
Other significant new features include:
- Windows nodes in Kubernetes clusters can now be “gracefully” shut down (KEP 4802).
- Environmental names can now be made from all printable ASCII characters (except “=”) (KEP 4369).
- Kubernetes be made of aware of the Load Balancer behavior (KEP 1860).
- Name conflicts can be resolved through automatic retries (KEP 4420).
Muñoz also pointed to a couple of features that he himself is particularly enthusiastic about.
The VolumeGroupSnapshot API (KEP 3476), from SIG Storage, allows users to take a snapshot of multiple volumes together, ensuring the data consistency across the volumes.
Another is pod-level resource allocations, as put forth in KEP 2837 by SIG Node. Creating a shared pool that all the containers in a pod can use, this feature “improves the ability to set up resource requests and limits at the pod level,” Muñoz said.
On Thursday, Jan. 9, 2025, at 5 p.m. (UTC), members of the Kubernetes v1.32 release team will hold an online discussion to go over the changes brought about by Kubernetes 1.32 in more detail.
The post Kubernetes 1.32 Aces API Conformance Testing appeared first on The New Stack.
K8s APIs do what they say on the tin; also Dynamic Research Allocation (DRA) has been revamped for better GPU mastery.