The planned release date of September 27th arrived and passed but with some last-minute adjustments, machine learning on Kubernetes project Kubeflow has finally arrived.
Version 1.4 is the second bigger release of the year, which is why the Kubeflow team focused on simplifying operations and streamlining the ML workflows to keep the project maintainable and attract more users.
Kubeflow was taught to use additional metadata in its pipeline orchestration and model monitoring efforts with stabilized v2 protocols in the KFServing and KFPipeline components. The team also worked to remove redundant code.
Replacing redundant code involved replacing the training operators for PyTorch, MXNet, LightBGM, XGBoost, TensorFlow, and BytePS with a universal one and developing common code for the existing web apps.
Building processes have been evaluated as well and now have a higher degree of automation to make them faster and less error-prone. Improvements of a similar nature were to be included through refactoring Kubeflow’s manifest files. Soe PRs that were necessary to make this happen were still under review when the release was finalized.
Long-time users of Kubeflow will be quick to note that the project’s central dashboard now has a menu item for the Models web app and allow users to add new ones if a cluster-admin wants to add third-party apps into the navigation sidebar.
The team also fixed issues connected to autoscaling GPU node groups, limit calculations, and MountPath parsing in the Jupyter web apps. They also corrected things in the controller watches and WebSockets to handle notebooks, which should make the user experience smoother.
To find out more check the release notes here.