| Opening/Intro | |
| AI Workspaces for the Modern Era – Kubeflow Notebooks v2- Mathew Wicks, NVIDIA. Siyuan Zhang, Google | Kubeflow Notebooks runs interactive development environments for AI, ML, and Data workloads on Kubernetes. In this talk, we are pleased to announce that Kubeflow Notebooks 2.0 is now available in beta!
Kubeflow Notebooks v2 is a ground-up rewrite with a new dual-persona architecture, separating what admins control from what users choose. Project maintainers from Google, NVIDIA, and Red Hat will walk through everything that changed.
For Admins: WorkspaceKinds replace user-defined Pod specs with pre-built image and pod options that you define. Activity rules let you auto-pause workspaces to reclaim idle GPUs. Option redirects let you roll out updates to running Workspaces without coordinating downtime with users.
For Users: Workspaces mean you no longer need to understand Kubernetes to do your work. Better yet, you can now easily update an existing environment by choosing from the options defined by your admin (no deleting and re-creating necessary).
Join us for a live demo of the beta, and gain early insights into how migration from v1 will work!
|
| Agentic RAG on Kubeflow - Rohit Kumar, Oracle. Santhosh T, Zenoti | Agentic RAG extends traditional retrieval-augmented generation by enabling AI systems to plan, use tools, evaluate intermediate results, and refine responses through multi-step workflows. In this session, I will demonstrate how Kubeflow can orchestrate an Agentic RAG system as a production-oriented LLMOps pipeline.
The talk will cover a practical architecture using Kubeflow Pipelines for document ingestion, embedding generation, vector index updates, retrieval evaluation, agent testing, and deployment. It will also show how agent workflows can connect with tools and MCP-style services while Kubeflow provides reproducibility, versioning, experiment tracking, and continuous evaluation.
Attendees will learn how to move from a RAG prototype to an observable, testable, cloud-native Agentic RAG workflow using open source MLOps practices
|
| Automatic Resource Configuration for Kubeflow TrainJobs - Vassilis Vassiliadis, IBM Research | Currently, Kubeflow TrainJob requires users to manually specify GPU, CPU, and memory requests before submission. Getting the resource requirements right is crucial:
- Under-requesting causes GPU out of memory exceptions, and
- Over-requesting uses up valuable and limited resources that other workloads could have benefited from.
In this session we talk about an upcoming feature that automatically right-sizes TrainJobs using an opt-in plugin system.
We're building a protocol in Kubeflow Trainer that enables TrainJobs to opt into automatic resource configuration. The protocol is quite simple. An external Kubernetes controller generates resource recommendations while Kubeflow Trainer applies guardrails to those recommendations and integrates with platform components like Kubernetes and Kueue. This plugin-based architecture enables platform teams to deploy custom recommenders without the need to modify Kubeflow Trainer.
This session covers:
- the current status of the feature
- a high level description of the protocol
- the safety mechanisms including timeouts and quota enforcement
- an example plugin that automatically configures the resource requirements of jobs that finetune LLMs
|
| Kubeflow Community Distribution - Julius von Kohout, DHL Data & AI | The goal of this session is to first introduce the new Kubeflow community distribution working group and committee. Who is involved, how to participate. Furthermore I want to explore the economics behind it and which companies are consuming it, whether as individual components or multi-tenant platform. This talk is from a maintainer of the community distribution and member of the Kubeflow Steering Committee. |
| Break | |
| Interactive Spark at Your Fingertips: Integrating SparkConnect into Kubeflow Notebooks - Vikas Saxena, RAICS.AI | Background
Apache Spark is indispensable for large-scale data processing and feature engineering in ML pipelines. However, running Spark interactively within a Kubeflow Notebooks environment has historically relied on tools like Jupyter Enterprise Gateway and Apache Toree — both of which are no longer actively maintained, making them a security risk, a source of operational debt, and unlikely to support future Spark versions.
Why SparkConnect (and Almond for Scala users)?
SparkConnect is a Spark-native, actively maintained alternative which, combined with the Spark Operator, makes Spark a true native service inside Kubeflow — no third-party gateway, no external process management.
Because the connection is interactive, data scientists can run exploratory analysis and quick proof-of-concepts against large datasets without spinning up dedicated jobs — iterating cell by cell, at scale, just like working with a local DataFrame.
What we'll cover
A live walkthrough of:
Deploying SparkConnect as a Kubernetes service
Connecting to the deployed service using PySpark
Connecting to the deployed service using the Scala API
|
| Closing the Observability Gap: Building a Kubeflow Plugin for Headlamp - Alok Dangre, NIT Rourkela | Managing ML workloads on Kubernetes means constant context-switching toggling between Headlamp and Kubeflow's UI just to understand what's running in your cluster. The Kubeflow Headlamp plugin closes that gap.
Built as part of an LFX Mentorship project, the plugin surfaces Kubeflow CRDs directly in Headlamp's sidebar: Pipelines, Pipeline Versions, Runs, Recurring Runs, Experiments, Notebook servers, Katib experiments, Trainer resources (TrainJob, TrainingRuntime, ClusterTrainingRuntime), and Spark applications each with status, timing, and a full detail view.
This session covers what was built and how: the plugin architecture in TypeScript using the Headlamp Plugin SDK, and the guiding principle of "complement, not duplicate" enriching cluster-level observability without reimplementing Kubeflow's native UI. We'll demo the one-click deep-link handoff to Kubeflow's own interfaces for detailed workflow management.
Attendees leave knowing how to install and use the plugin today, and how Headlamp's extension model can surface any CNCF project's custom resources in a unified Kubernetes UI.
|
| Stop AI Agents from Guessing SQL with Semantic Operator on Kubernetes, Manabu McCloskey - AWS | If you are building AI agents or BI applications that help analysts and business leaders answer questions from enterprise databases, we would value your feedback on a new open source project called Semantic Operator. AI agents can generate plausible SQL while misunderstanding business metrics, choosing incorrect joins, or bypassing governance rules. Semantic Operator addresses this by running Apache Ossie semantic models as Kubernetes resources. It provides agents and applications with certified metrics and dimensions, then generates deterministic, governed SQL for StarRocks or Trino. In this live demo, we will show how to manage semantic models through a Kubernetes operator, validate schemas and detect physical drift, enforce row, column, and metric governance, plan deterministic queries, and expose MCP and REST access for AI agents and applications. We will also walk through the end to end architecture and potential integration points. |
| Contributing to the Kubeflow Ecosystem - Evolving from beginner to pro OSS contributor - Christian Heusel, Contractor / Self Employed | Finding an entrypoint for your first contribution to Kubeflow can be hard, but it doesn't have to be!
This talk aims to point out ideas for new and experienced developers that want to help out in the Kubeflow ecosystem but don't know where to start in the context and beyond the information mentioned in the contribution guide: https://www.kubeflow.org/docs/about/contributing/
During the talk I will highlight my own journey of getting started within the project, which is mostly centered around the Manifests and Notebooks Working Groups, but also abstract a few principles that apply to many of the other established open source projects around.
Additionally to the obvious (but non-trivial) way of contributing by means of writing code the talk has the goal of pointing out ways for people to help out our community who are not mainly Software Engineers but rather Data Scientists, Infrastructure Administrators or Project Managers.
|
| Closing Remarks | |
