MiniKF 20210428.0.0

NEW! Now with Kubeflow 1.3!
Create an AutoML workflow with the click of a button. Start from a dataset, define a task, then discover, train, and optimize a model from inside your notebook automagically.

New MiniKF release including

• Kubeflow 1.3
• Rok 1.2-pre-306

New Features

• Upgrade Rok to the latest pre-release version.
• Upgrade Kubeflow to version 1.3.
• Upgrade NVIDIA drivers to version 460.
• Upgrade Linux Mainline Kernel to version 5.4.104.
• Upgrade Istio to version 1.9.0.
• Brand new TensorBoard UI for logging of training jobs.
• Updated Katib UI for a better hyperparameter tuning experience.
• Updated Volume Manager UI for a better experience when managing Kubernetes PVCs.
• Updated Kubeflow Pipelines UI with the ability to manage recurring runs.
• Updated Kale UI for an enhanced data science experience.

Improvements

• Support fine-grained authorization in Rok with Kubernetes RBAC.
• Remove custom kernel modules, use Arrikto dm-clone module from the upstream Linux kernel.
• Use SubjectAccessReview to authorize requests performed inside the Kubernetes cluster.
• Improve authentication mechanism in the Rok API and client.
• Use kustomizations for all applied manifests.
• Use upstream NGINX controller instead of Minikube’s addon.
• Improve pre-pulling of Docker images to auto-generate the list of images based on the kustomizations to be applied.
• Support token-based authentication as the sole method of programmatic authentication in the Rok client and CLI.
• Introduce pagination in the Rok UI.
• Feature a new landing page for MiniKF on Vagrant.
• Produce a smaller Vagrant box by excluding non-critical images from the pre-pull list.

Fixed Issues

• Fix a bug that could lead to data corruption when snapshotting a filesystem that needs recovery.
• Fix an issue with the Go runtime which could lead to CSI pods crashing due to max locked memory.
• Fix an issue with Rok becoming ready after an unscheduled node reboot.
• Fix pre-clone verification steps that could lead to errors.
• Fix a KFP API server issue causing pod crashes if the default parameter values are empty.
• Reduce the amount of time that the provisioning script might wait for APT in an air-gapped environment.

Known issues

• The OpenVaccine Codelab does not currently run end-to-end inside an airgapped environment, due to an issue with the Knative Serving controller.
• Under heavy load Kubernetes might delay starting pods even if Rok has successfully provisioned the underlying storage.