Andreea Munteanu
on 22 November 2023
Canonical, the publisher of Ubuntu, announced today the general availability of Charmed Kubeflow 1.8. Charmed Kubeflow is an open source, end-to-end MLOps platform that enables professionals to easily develop and deploy AI/ML models. It runs on any cloud, including hybrid cloud or multi-cloud scenarios. This latest release also offers the ability to run AI/ML workloads in air-gapped environments.
Run AI workloads in air-gapped environments
MLOps platforms often need to be connected to the network, which can be problematic for organisations that have strict cybersecurity and compliance policies. Charmed Kubeflow allows users to run their workloads offline in an air-gapped environment, in addition to public clouds and on-prem data centres. With the inclusion of this feature, Charmed Kubeflow provides additional security for organisations in highly regulated industries or projects that handle sensitive data. This allows them to complete most of the machine learning workflow within one tool, and avoid time spent on connecting tools and ensuring compatibility between them.
Enhanced capabilities to customise MLOps tools
Every AI project is different and so are the tools, frameworks and libraries that organisations use. While some might prefer traditional options such as Tensorflow or Pytorch, others might go with industry-specific ones such as Nvidia NeMo. Charmed Kubeflow 1.8 brings new enhancements that allow end users to customise their MLOps platform.
Users can add any image within their Jupyter Notebook. This gives professionals freedom to choose their preferred tools and libraries and focus on developing machine learning models rather than maintaining their tooling. Users can plug tools or components in and out depending on the use case to work efficiently.
This capability differentiates Charmed Kubeflow from the upstream project. Organisations are more likely to be able to move beyond experimentation using Canonical’s supported solution, since they can add their own Notebook images and develop models using them.
Build models for production. Reproduce experiments with ease.
Kubeflow as a project was designed to run AI at scale. It is able to run the entire machine learning lifecycle within one tool. Kubeflow Pipelines are the heart of the project, since they are specialised in automating machine learning workloads. This is one of the reasons organisations that are looking to scale AI projects prefer Kubeflow over its alternatives.
Charmed Kubeflow benefits from the upstream project’s newly introduced Kubeflow Pipelines 2.0, which further simplifies the automation process. Some capabilities such as the directed acyclic graphs (DAG) have been available for a while in beta, but other capabilities have been introduced in Kubeflow 1.8. For example, Kubeflow now abstracts the pipeline representation format, so that it can run on any MLOps platform. This translates into smoother migrations from the upstream project to distributions or tools that can offer enterprise support, security patching or timeline bug fixes.
“I’m thrilled to be part of the upstream community’s Kubeflow 1.8 release and proud of the Charmed Kubeflow team for driving the release as well as providing feedback along the way”, said Kimonas Sotirchos, Working Group Lead in the Kubeflow Community. “Charmed Kubeflow 1.8 is a great way for newcomers and experienced users to try out all the latest and greatest features in Kubeflow, like KFP V2 and PVC browsing”, he added.
Innovate at speed with Canonical MLOps
Charmed Kubeflow is the foundation of a growing ecosystem that addresses different needs for AI projects. The MLOps platform is integrated with leading open source tools. For example, Charmed Kubeflow integrates with Charmed MLflow to facilitate experiment tracking and model registry. MLFlow is a lightweight machine learning platform that enables professionals to quickly get started locally or on the public cloud and then easily migrate to an open source fully-integrated solution. Charmed Kubeflow can also be integrated with KServe and Seldon for model serving.
