We are very excited to announce that Kubeflow 1.1 and a new version of MiniKF have been released!
Kubeflow 1.1 brings ML workflow automation with Fairing and Kale. The latter enables you to work on your notebook, write your ML code, define the pipeline steps using an intuitive UI, and then build a multi-step pipeline automatically.
The latest release extends isolation and security through the delivery of multi-user pipelines. User authorization in the various Kubeflow web apps is enforced using subject access review.
Finally, the installation and operations of Kubeflow have been enhanced to support GitOps methodologies.
The latest MiniKF comes with brand-new support for HP tuning with Kubeflow Pipelines, Katib, and Kale. The same way you are using Kale to convert a Notebook to a Kubeflow Pipeline, you can now use it to optimize a model with the click of a button. Kale provides you with an intuitive UI to configure the parameters, search algorithm, and search objective. Then, it orchestrates Katib and Kubeflow Pipelines so that every Katib Trial is an individual Run on KFP. Kale groups all these runs under a single KFP experiment.
When performing hyperparameter tuning, some steps of the pipeline produce the same output every time, as they don’t include parameters that change. This is why it’s important to cache these steps and use your resources wisely to speed up the computation time. Rok caches individual pipeline steps for enhanced performance when performing hyperparameter tuning.
This ML workflow completely unifies Katib with Kubeflow Pipelines, providing full visibility, and reproducibility for each step of the hyperparameter tuning process.
Moreover, Kubeflow Pipelines now run in the same, isolated namespace as your notebooks, and have access to the same resources. This is a great feature, especially when deploying Kubeflow and Rok on a multi-user environment.
After interacting with a lot of MiniKF users, as well as with our customers, who are using Arrikto’s curated version of Kubeflow on public clouds, we realized that the Kubeflow Central Dashboard needed some enhancements. That’s why in the latest version of MiniKF, we redesigned Kubeflow Central Dashboard sidebar, making it more intuitive and tailored to the needs of data scientists.
Another great feature is that we now expose the MiniKF dashboard at a standard `https://<minikf_name>.<project>.cloud.goog` URL for seamless HTTPS support.
If you want to experience a seamless data science workflow, where you will optimize a model using hyperparameter tuning with the click of a button, then follow this tutorial: https://codelabs.developers.google.com/codelabs/cloud-kubeflow-minikf-kale-katib/#0
The Kubeflow Community is pretty active and welcoming. We’d love you to join us!