AI
TAG Workloads Foundation: Mission and Scope
TAG Workloads Foundation is one of five CNCF technical advisory groups, reduced from eight in a restructuring last year. The mission: define and advance practices for running fundamental cloud-native workloads, from execution environments to life cycle management. Scope spans containers, virtual machines, special OS runtimes, CI/CD, scheduling, and WebAssembly. End-user activities and anything non-cloud-native are excluded. The TAG is fully volunteer-run, augmenting the small CNCF TOC by handling technical reviews and cross-project coordination at scale.
Batch Scheduling Sub-Project and Utilization Benchmarking
The TAG runs one formal sub-project: batch systems. Meetings are bi-weekly on Tuesdays at 8am PST. The group covers data locality, cluster topology, and node topology. A workloads API ships in the current KubeCon release. Two active initiatives: a white paper on scheduling for cloud-native AI (awaiting final TOC review) and a cluster utilization benchmarking project. The benchmarking work targets near-100% scheduler utilization. A working group on agentic systems is also forming and seeking contributors.
CNCF Graduation: Requirements and Key Projects
Graduation requires wide community adoption, diverse maintainer governance (no single-company control), and security compliance. Kubernetes was CNCF’s first graduated project. Recent graduates include Knative (zero-to-one autoscaling) and Crossplane (infrastructure provisioning without writing code), both from last year. For container registries, Harbor stores images in OCI format. Dragonfly handles distribution. The two now collaborate to package and distribute large AI models as OCI artifacts, addressing the pain of shipping multi-gigabyte models across clusters.
Incubating and Sandbox Projects
Incubating projects need several active users and a visible roadmap showing maintainer growth. Current incubating projects include Kubeflow, Volcano, Kata Containers, Lima (joined last year), and KServe. Sandbox projects represent early experiments. Four highlighted at this session: an AMD project shown in the KubeCon keynote, KAI Scheduler for AI workloads, Kube-Elasticity (focused on scale-to-zero rather than the more common zero-to-one), and OpenCHIO (a developer platform combining CI/CD, GitOps, and observability).
How to Get Involved in TAG Workloads Foundation
Entry points: bi-weekly TAG meetings, Slack channels, and a mailing list. Contributors can join the batch sub-project, write white papers, or review projects applying for CNCF sandbox or incubation. Project reviews happen in dedicated sub-project review meetings open to anyone. The TAG has roughly 30 active initiatives. Projects seeking sandbox status open a ticket in the TOC GitHub repo and enter a review queue. Community groups can form informally around any specific topic without a formal process.
Q&A
Why was the Kyros sandbox project missing from the TAG’s project overview? The slides were assembled that same week, so entries were missed; the TAG invited any maintainer to contribute content to future sessions. ▶ 20:04
If a sandbox project ships a new Linux distribution under its existing repositories, does it need a separate CNCF application? Sub-projects added loosely likely require no separate application, but the addition factors into the incubation review later; for logos and IP, contact CNCF sandbox onboarding staff directly. ▶ 21:03
Is there a formal process for aligning CI/CD and AI inference tools like KServe across projects? Collaboration is project-driven today, but shared components like the Gateway API inference extension are being moved into standard APIs to make them reusable across vendor-specific projects. ▶ 25:01
Notable Quotes
You want as close to 100 as you can get. Marlow Warnicke · ▶ 09:08
the AI model is is very I mean painful Yuan Tang · ▶ 14:10
Key Takeaways
- TAG Workloads Foundation consolidated from eight CNCF TAGs to five in a restructuring last year.
- Harbor and Dragonfly now jointly handle OCI-packaged AI model storage and cluster-wide distribution.
- Cross-project API alignment, such as the Gateway API inference extension, is project-driven with no formal TAG mandate yet.
About the Speakers
Stephen Rust is a Principal Architect at Akamai Cloud with over 20 years of experience in operating systems, storage, and open source work with containers and Kubernetes. He leads cloud-native architecture at Akamai and serves as tech lead for TAG Workloads Foundation.
Yuan Tang is a Senior Principal Software Engineer at Red Hat AI and co-chair of TAG Workloads Foundation. He holds leadership positions across Argo, Kubeflow, KServe, Kubernetes, and CNCF, and is a maintainer and author of numerous open source projects.
Marlow Warnicke is a Senior Engineer at Nvidia and chair of the CNCF Batch Working Group. She leads the Slinky project and brings expertise in resource management, AI/ML Kubernetes compute, high-performance compute systems, kernel drivers, and security.
Kante Yin is an R&D Engineer at HivergeAI in Cambridge, UK, focused on AI agents and model inference systems. He is the founder of InftyAI and serves as tech lead for TAG Workloads Foundation.