Kubeflow Summit, October 2022

The 2022 Kubeflow Summit 2022 is dedicated to bringing together Kubeflow users, developers, contributors, data scientists, ML engineers, MLOps professionals and more who rely on Kubeflow as a product and its community as a critical element to their MLOps workflow.

This time we've broken the program into two separate days: Day 1 will focus on the user, with presentations from Kubeflow experts and hands-on workshops. Day 2 is dedicated to our Kubeflow Contributors and Community with discussions focused on the operational and administrative aspects of the Kubeflow Community.

Please note the schedule is subject to change.

**TI MES ARE PACIFIC TIME.**

AGENDA:

Day 1, October 18: Dedicated to Kubeflow Users (presentations from Kubeflow experts)

Join us at Kubeflow Summit to learn best practices for managing end-to-end machine learning workflows from training all the way to deployment.

We'll look at data science use cases for Kubeflow, the devops challenges those present, and some of the cutting-edge tools evolving in the Kubeflow ecosystem to address them.

Register for In-Person Attendance (will be added later this week)

8:00 am - 9:00 am - Registration

9:00 am - 9:30 am - Keynote / Panel Discussion - David Aronchick, Jeremy Lewi, Kubeflow Project Founders

9:30 am - 9:55 am - How Aurora uses Kubeflow Pipelines to Accelerate ML Model Development for Autonomous Vehicles - presented by Vinay Anantharaman

Aurora builds self driving technology and at the heart of its product development is a data engine for improving ML models. A key component to the data engine is the ML pipelines we have developed for the perception system of the driver.

In this talk we will discuss the ML Models in Aurora's perception system and dive deep into one of the Kubeflow pipelines for the perception system. Then we will discuss the open source and custom Kubeflow components of the pipeline and how the pipeline is integrated into developer workflows. Finally, we’ll discuss the approach we’ve used to migrate ML pipelines across the organization to Kubeflow.

10:00 am - 10:25 am - Roblox User Story presented by Denis Goupil

In this talk, I will drive you through the ML@Roblox journey. From the project kick off in August 2020 until 2 years after, I will describe the strategy we adopted to build the machine learning platform for Roblox, where we failed and how we adapt. You will also learn how we deploy and use Kubeflow and drove adoption up to 130 monthly users. Finally, I will highlight few of our current initiatives and what we want to tackle next.

10:25 am - 10:40 am - Break

10:40 am - 11:00 am - Lightning Talks (10 Mins each)

Nested Workflows in Kubeflow Pipelines presented by Greg Larmore

Kubeflow Pipelines (KFP) provides an excellent platform for building and deploying portable, scalable machine learning (ML) workflows. However, as ML problems become more complex and varied, so do their workflows.

In this session we demonstrate a method for nesting workflows to provide encapsulation and reuse. Integral to this method we create ad hoc components using the Jinja templating engine. Finally, we add data validation to pipelines in the KFP SDK v2 using Python type hints and pydantic.

Enterprise MLOps using Kubeflow with DKube present by AjayTyagi

In this lighting talk the attendees will learn how DKube, an enterprise MLOps platform optimized for on-prem, hybrid, and multi-cloud is built on top of Kubeflow and MLFLow as well as what additional capabilities it supports such as model monitoring. DKube is being used in pharma, biotech, retail analytics, and aerospace/defense applications and is a good example of successful use of Kubeflow in enterprise environments.

11:00 am - 11:25 am - The Infrastructure that powers ML Training at Pinterest presented by Karthik Anantha Padmanabhan and Chia-Wei Chen

"Training Compute Platform" powers 100% of Pinterest Training workloads and has been the underlying infrastructure on which ML innovations are iterated on and productionized. We have worked very closely with our partners in Ads and Organic Surfaces to power step function innovations that have made ML training faster and more cost efficient.

We enable more than 400 ML engineers to seamlessly iterate and experiment on their ML training without ML engineers having to worry about the operational details like Nvidia driver upgrades, cuda drivers, dependency management, host provisioning, instance upgrades etc. By providing a platform which centralizes and solves the above problems we are enabling ML engineers to focus on their strengths and provide massive leverage. Reducing the ops burden speeds ML developer velocity tremendously and allows ML engineers to iterate on next generation models thereby enabling rapid innovation at Pinterest.

We solved this problem by building of top of Kubernetes and the training operators provided by the Kubeflow framework which has allowed us to implement Distributed Training on Kubernetes. Using these compute framework we have also been able to power features like Multi Node Multi GPU Training, Hyperparameter Tuning and data privacy aware ephemeral developer instances on GPUS.

11:30 am - 11:55 am - Kubeflow for Enterprise - Samsung Case presented by Joonwon Lee

We have developed MLOps Portal based on the Kubeflow and Kubernetes since 2020 in our company, Samsung SDS. We also deployed our Portal with Kubeflow to our client sites as their AI Platform.

I'd like to explain what we added to the Kubeflow in order to provide the MLOps Portal, what we're preparing to upgrade it, how we used the Kubeflow, and what we expect for the future of Kubeflow. For example, we developed the functions and UIs of project and user management, resource quota management, training job submission and its monitoring, serving, and so on.

12:00 pm - 1:00 pm - Lunch

1:00 pm - 1:25 pm - Building a Unified ML Platform with Kubeflow and Kuberay presented by Richard Liu

Kubeflow is an open source, extensible ML platform aimed at simplifying the ML stack on Kubernetes. In this talk, we will demonstrate how Kubeflow can be enriched with other OSS technologies like Ray to provide a unified ML experience.

Audience will learn how to exercise the end-to-end machine learning user journey (building, training, and deploying ML models) using Kubeflow and Kuberay. Features include:

- Integration with Jupyter notebooks

- Authentication with Istio

- Collaborative options with Kubeflow profiles

- Build/train/deploy with RayTrain and RayServe

1:30 pm - 1:55 pm - Model Monitoring for model trained and served on Kubeflow presented by Subra Ongole

In this session the attendees will learn how model monitoring for columnar and image datasets with DKube Monitor can be implemented with data prep, training, serving being done with Kubeflow pipelines and KServe in an on-prem, cloud environment, or multi-cloud, multi-cluster environment. As models go into production model monitoring has now become the most requested capability from the industry. However, it has the least open source innovations available for people using Kueflow to take advantage of. Attendees will learn how DKube Monitor can be integrated with KF installations for model monitoring AI models built for columnar or image datasets.

2:00 pm - 2:25 pm - Building an ML Application Platform from the Ground Up presented by Sean Sheng

BentoML users leverage Kubeflow to achieve continuous model retraining but the value of an ML model is not realized until it is deployed and served in production.

Building an ML application is more challenging compared to a traditional application due to the added complexities from models and data in addition to the application code. Using web serving frameworks (e.g. FastAPI) can work for the simple cases but falls short for performance and efficiency. Alternatively, using pre-packaged models servers (e.g. Triton Inference Server) can be ideal for low-latency serving and resource utilization but lacks flexibility in defining custom logic and dependency. BentoML abstracts the complexities by creating separate runtimes for IO-intensive preprocessing logic and compute-intensive model inference logic. Simultaneously, BentoML offers an intuitive and flexible Python-first SDK for defining custom preprocessing logic, orchestrating multi-model inference, and integrating with other frameworks in the MLOps ecosystem.

2:30 pm - 2:55 pm - Serverless Magic for ML Orchestration using Kubeflow presented by Rishit Dagli and Shivay Lamba

Kubeflow is an awesome tool for writing, orchestrating, and running end-to-end machine learning pipelines. What if I told you that you could orchestrate your ML workflow in a serverless manner, while not having to leave the comfort of using the **Kubeflow Pipelines DSL package to create your pipelines.**

In this talk, we show how you can have pipelines built with Kubeflow’s Pipelines DSL and orchestrate it with Vertex AI pipelines. We focus on demonstrating how some unique features of Vertex AI could be integrated pretty well with Kubeflow including storing and managing artifacts with Vertex ML metadata and interacting with other cloud services from pipelines. In this talk, we also show how Google Cloud Pipeline Components can integrate well in such a setting and also show how it could very well help optimize your machine learning pipelines in terms of: resource usage, lower component startup times as well as better artifact and debugging support.

To top this off, we show demos of doing so and how the audience could include Vertex AI in their existing Kubeflow Pipelines.

3:00 pm - 3:25 pm - Concurrent-AI: A Parallelized AI Execution Engine that brings MLFlow and Kubernetes together presented by Jitendra Nath Pandey and Jagane Sundar

AI workloads are emerging as one of the largest use cases for kubernetes deployment at scale and require a variety of hardware utilizing a range of computational resources. At the same time, AI workloads require a strong MLOps platform managing the AI experiments, models and data.

In this paper we present the Concurrent-AI platform that bridges the gap between MLFlow (the most popular AI Experiment Tracking and Model Management platform) and Kubernetes. Concurrent schedules AI workloads on kubernetes for large scale parallel execution. Concurrent platform provides dynamic data-partitioning of a workload across containers for an efficient and scaleout execution. Concurrent-AI platform can also schedule tasks across multiple kubernetes clusters on public and private clouds. The builtin integration with MLFlow ensures every experiment run and data version is tracked for visibility and reproducibility at every stage of processing. Concurrent is fully open source, and we are looking forward to a fruitful collaboration with the KubeFlow community.

In this presentation we demonstrate the capabilities of Concurrent and provide an overview of its architecture and APIs.

3:30 pm - 3:55 pm - Converting Kaggle Competitions into Kubeflow Pipelines and Open Source Contributions presented by Jimmy Guerrero

In January of 2022, Arrikto sponsored an open source contribution initiative that involved the sponsorship of several community members to help make Kaggle competitions run as Kubeflow Pipelines. The idea was to make it easy for the Kaggle community to experience the capabilities of Kubeflow through popular competitions. To date, twelve Kaggle competitions have been made compatible with Kubeflow. The resulting notebooks, source code and documentation have all been open sourced and can be found in the kubeflow/examples directory.

In this talk, we'll walk you through the typical development workflow for making Kaggle competitions compatible with Kubeflow, including running the pipelines themselves. We'll explore environment preparation, authentication, building components with Python functions and Docker images, interfacing with the KFP SDK to create YAML configurations and generating submission files. Finally, we'll discuss what's next for the program in 2023.

Day 2, October 19: Dedicated to the Operational and Administration of the Kubeflow Project and Community (contributors, project operations/organization and feedback from the Kubeflow user survey)

Join us on Day 2 where we will hear from project leaders, discuss important community topics, and meet with Working Groups and Special Interest Groups to collaborate in-person and online. All Kubeflow contributors or those interested in contributing are encouraged to attend.

Day 2 Structure (Schedule will be updated as speakers and topics are confirmed)

9:00 am - 9:15 am - Opening Remarks and overview of the day

9:15 am - 9:50 am - Breakout into Working Groups to identify top challenges and Roadmaps Items

9:50 am - 10:00 am - Transition

10:00 am - 10:30 am - Katib and Training Operator

10:30 am - 11:15 am - Pipelines

11:15 am - 11:30 am - Break

11:30 am - 11:50 am - Notebooks

11:50 am - 12:10 pm - KServe

12:10 pm - 1:00 pm - Lunch

1:00 pm - 1:30 pm - Manifests

1:30 pm - 2:00 pm - Components (How do we take care non core components)

2:00pm - 2:15 pm - Break

2:15pm - 2:45 pm - Release Team Update (1.7 and Roadmap)

2:45 pm - 3:15pm - Improving the Kubeflow Project Organization (especially Testing, Documentation, Approvals)

3:15 pm - 3:30 pm - Onboarding New Community Contributions

3:30 pm - 4:00 pm - Summary Day, Open Discussion and Feedback

Event FAQ

How much does it cost to attend?

Kubeflow Summit both in-person and virtual is free to attend but registration is required.

Event Location (for in-person attendance)

We will announce the in-person venue (will be near the SFO airport) this week and will add that information ASAP.

Code of Conduct

All speakers, attendees, and event staff are expected to abide by the Kubeflow Project Code of Conduct at Kubeflow Summit and any related events.

Get In Touch

If you have any questions about the Kubeflow Summit please email akgraner at akgraner@arrikto.com or via eventbrite.

Registrations are closed

2022 Kubeflow Summit - Virtual Registration