CNCF-hosted Co-located Events Europe 2025

Let's hear from the Release Team all updates about the recent releases and new components to the project that is growing in users and contributions. We'll talk about the features the Release Team helped bring in Kubeflow 1.10, and discuss the roadmap for the next releases.

ML workloads require repetitive access to data for model training. This repetitive access can be both slow and costly in cloud environments further slowing down model training and leaving GPU resources idle waiting for data to load. As datasets and training workloads become larger and more sophisticated in the era of GenAI, efficient data access is crucial to improving training workload speed and efficiency. In this talk, we will discuss optimized data caching for ML workloads using Apache Iceberg, Apache Arrow Flight, and Kubernetes. We will demonstrate a distributed in-memory cache of an Iceberg table across a fleet of Kubernetes pods used to load data more efficiently into Kubeflow training workloads.

As the demand for scalable machine learning (ML) workloads increases, efficient training in distributed environments has become crucial. This talk will delve into Kubeflow innovations that advance distributed training on Kubernetes with JAX and automate hyperparameter optimization for Large Language Models (LLMs).
JAX, known for high-performance large-scale computations, requires Kubernetes integration for efficient scaling. Additionally, hyperparameter optimization for LLMs has been manual and time-intensive, with existing tools lacking seamless Kubernetes integration.
To address these gaps, we extended Kubeflow to support distributed JAX workloads and developed a high-level API to automate LLM hyperparameter optimization. These advancements make complex, resource-intensive training more efficient. The speakers will highlight how these capabilities streamline end-to-end ML workloads, establishing Kubeflow as a powerful platform for modern AI development.

Competitive challenges in machine-learning serve as the central point for researchers to interact with their community. Nowadays popular services like Kaggle are used to share, exchange and compete on such challenges. But they are bound by resource constraints that block scalable model training by its participants, and are not suited for setups where data is kept internal or internal tooling is needed. Running your own infrastructure can help tackle these problems but requires management and scalable orchestration of workloads. As KubeFlow is the ideal tool for orchestration and distributed training on Kubernetes, it can be leveraged for running submissions. In this setup, user code is executed as a pipeline where data loading, distributed training and scoring is managed, allowing participants to focus solely on their model code. A case study for running particle physics based challenges at CERN will show how this framework is set up and which challenges were faced during development.

Edge AI leverages local data processing and millisecond-level response times, unlocking vast application potential. With cloud-native advancements, it is evolving into edge-cloud collaborative AI, enabling flexible AI task deployment across cloud and edge via coordination algorithms to meet diverse demands for real-time performance, accuracy, and privacy.
KubeEdge has introduced the distributed edge-cloud collaborative AI framework, Sedna, which supports seamlessly deploying existing AI applications to the edge. To address the management and scheduling challenges of distributed edge-cloud collaborative AI applications, this presentation will demonstrate how to integrate the KubeFlow training-operator into KubeEdge's Sedna framework, extending distributed training capabilities to the edge. Using training-operator's group scheduling, tasks are dynamically allocated across cloud and edge, optimizing resource use and enhancing edge-cloud AI efficiency.

Kubeflow exemplifies the power of community by bringing different unique components together to provide a simplified user experience to build, train, and deploy AI/ML at scale. Combining complementary roles and skills is critical for any community. We all play different roles, sometimes more than one. The community is heterogeneous, open to newcomers, and growing quickly. This session will discuss how different skills and roles can be valuable to any community. Every contribution matters, from updating a code snippet in the documentation to managing a release, building a course training, or contributing to the source code of any of the Kubeflow ecosystem's components. The speakers will share their unique experiences from lessons learned to develop a new skill and how this connects to the Kubeflow ecosystem. How contributing to an open source can bring career growth, opportunities to explore and learn a new role, and how to get started from technical to non-technical contributions.

Managing users in platforms on top of Kubernetes, like Kubeflow, is always challenging.

The source of truth is usually defined in an OIDC Provider, yet necessary changes need to be applied in the cluster to reflect the corresponding permissions.

In this talk we'll explore how in Canonical we managed to bridge the gap between defining users in one place, an OIDC Provider, and reflecting the corresponding changes to Kubeflow's Profiles and their contributors.

Lastly we'll also cover how the above solution should be generalised and be a more Kubeflow-native implementation and further establish best-practices and reduce the moving pieces. From Istio and K8s RBAC resources, all the way to performing efficient group support in Kubeflow.