Andreea Munteanu
on 1 August 2024
Kubeflow 1.9. Kubeflow is a fully open source MLOps platform, part of the Cloud Native Computing Foundation (CNCF) as an incubation project. Canonical is one of the project’s official distributors, ensuring security maintenance of the container images, tight integration with the wider AI/ML ecosystem, and delivering enterprise support and managed services. The latest release offers new features related to GenAI, enabling organisations to run any new initiative in production.
Kubeflow in CNCF
Kubeflow has been part of CNCF since 2023, which has helped the project to evolve, drive adoption and attract contributors. Since the last release, the community has focused not only on new features for the product but also on ensuring an easier way to welcome new members. For instance, the community moved its communication to the Cloud Native Computing Foundation (CNCF) Slack channel and voted on its steering committee.
Upstream developments
Kubeflow is an open source project started by developers for developers. Contributors have always looked to bring new enhancements that keep the MLOps platform relevant in the market. In the latest release, the Kubeflow community started a new working group that focuses on model registry. This is a crucial step of the ML lifecycle that ensures that models are stored, which was previously achieved by integrating with other tooling. For now, the feature is available in beta, but the community is committed to work on improving it and adding it as part of the default features. Additionally, integration with the Spark Operator is a cross-community effort that has been progressing between the two projects. While it isn’t part of the default deployment,it is for now available in the official documentation.
One other challenge that Kubeflow users reported in the past was the need for advanced machine learning skills. Integrations with low code platforms such as Elyra have been done in the past, but in the last 12 months, the project has just benefited from Kale’s donation. It is a project that simplifies the experience of deploying Kubeflow Pipelines workflows by providing a simple user interface to define them.
“Kubeflow desperately needs to elevate the Data Science user experience and lower the barrier to entry to more Machine Learning practitioners. The Kale project will lay the foundations to build new usage patterns and integrations with popular Machine Learning developer tools. I cannot wait for the donation to complete and get exciting new tools into the hands of our Kubeflow users.” says Stefano Fioravanzo, Kubeflow upstream contributor.
What’s new in Charmed Kubeflow 1.9?
Alongside these community-driven developments, Charmed Kubeflow 1.9 also brings a number of additional features focused on security and integrations for enterprise use cases.
Secure MLOps platform
Canonical’s promise to provide secure open source software does not fall short when it comes to Kubeflow. In this release, Charmed Kubeflow comes with Istio CNI enabled by default so that users can configure the network interfaces in their Linux containers. It protects users from possible attacks and enables larger teams to better manage access to different ML projects.
Often, MLOps projects work with highly sensitive data, so organisations are looking to run their platforms on confidential virtual machines. Charmed Kubeflow can run on confidential VMs on Azure, opening the way for highly regulated industries to innovate at speed with open source.
These features add to Charmed Kubeflow’s existing security capabilities, such as CVE patching of the images, integrations with different identity providers and user management. The platform also runs in air-gapped environments, enabling highly regulated organisations to push through their AI/ML projects without worrying about the infrastructure.
Integration with the wider AI/ML ecosystem
Canonical and NVIDIA have collaborated for many years now, with exciting integrations starting with the hardware and up the software stack. After having both Canonical Kubernetes and Charmed Kubeflow certified as part of the NVIDIA DGX-Ready Software program, new integrations are now available. You can now develop your GenAI apps on Canonical’s MLOps platform, Charmed Kubeflow, and deploy them using NVIDIA NIM or NVIDIA Triton Inference Server, which are both integrated with KServe, Kubeflow’s component for deployment. Read more about it in the blog published last month.
In addition to the integrations with NVIDIA, Canonical has enabled the latest release of Charmed Kubeflow on Google Kubernetes Engine, and worked on integrating it with Azure Blob Storage. Last but not least, Canonical has released the Data Science Stack, a ready-to-run ML environment that enables users to rapidly develop and optimise models on any Ubuntu AI workstation.
