NEW! Now with Kubeflow 1.4!
Create a PyTorch distributed job from inside your notebook using Kale and Kubeflow. Distributed training made easier than ever.
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 1.4
- Rok 1.4-rc8-11-g47325593f
New Features
- Upgrade Rok to the latest stable version.
- Upgrade Kubeflow to version 1.4.
- Upgrade Kubernetes to version 1.19.15.
- Upgrade Minikube to version 1.23.2.
- Upgrade Linux Kernel to version 5.4.151.
- Updated Kubeflow UIs for an enhanced data science experience.
- Brand new MiniKF dashboard.
- Ability to expose Kubernetes metadata, resources, and spec in the Kale SDK. Users are now able to set limits, requests, labels, annotations, or use the nodeSelector using the Kale SDK.
- Allow setting environment variables in Kale step using the Kale SDK.
- Allow users to configure the size of the Kale marshal volume.
- Allow users to define a container and entrypoint for a pipeline step, built using Kale.
- Integrate Kale with the PyTorch distributed training operator to enable users to run distributed PyTorch training jobs.
- Extend Kale to support running conditionals with the outputs of the pipeline steps.
- Enable users to make predictions using an existing KF Serving Inference Service via the Kale API.
- Support a default automatic snapshot policy for notebooks so data scientists don’t accidentally lose work.
- Ability to view notebook servers across all namespaces in the Notebooks UI.
- Monitor the last activity of the notebook servers.
- Configurable way to stop idle notebook servers automatically.
- Enable users to mount an existing volume to a notebook server.
- Automatic log gathering process.
- Ability to present a notebook programmatically.
- Fully automated process for snapshotting all notebooks in a Kubeflow cluster and publishing them to Rok Registry.
- Fully automated process for restoring all notebooks of a Rok bucket.
- Support for ReadWriteMany (RWX) volumes.
- Protect critical Rok data from EBS failure or user error.
- Support Istio authorization to EKF resources based on groups inherited from the identity provider.
- Introduce an option to disable the auto-profile creation so that users don’t have their own namespace, but only be members of a shared namespace.
- Enable admins to apply a skeleton of Kubernetes resources for every user namespace in an automated manner.
- Extend the official K8s autoscaler to support scale-in when using local volumes.
You can find more information about bugfixes and improvements here: https://docs.arrikto.com/release-1.4/NEWS.html