What Do You Actually Federate? Notes from the 4th NFCS AI Container Registry Workshop

Notes from the 4th NFCS AI Container Registry Workshop, which circled a deceptively simple question all day: what should a federated AI container registry actually distribute, and who should it serve? The event surfaced governance and architectural disagreements more than technical implementation detail.

The Core Problem

Christian Kniep (QNIB Solutions) reframed container registries as no longer simple artifact stores. Via OCI/ORAS standards, a registry can now distribute almost anything — code, model weights, configurations, prompts — which makes “what to federate” a genuinely hard question. End users, resource providers, and software managers have conflicting priorities, and every choice about the unit of versioning directly expands or contracts the scope of governance.

Human Factors Matter

Rich Gilham’s framing: “HPC is the means, not the end.” Containers solve real reproducibility problems, but at a real cost — teams need dedicated expertise to manage image bloat, GPU-driver coupling, and security constraints, and containerization does not automatically guarantee reproducibility no matter how convenient it feels.

A Metadata-First Approach

Marc Dillon (AMD) offered the most concrete alternative: federate metadata and validation pipelines, not container blobs. A user asks for an application (“I need a chat model”), and an intelligent resolver selects the optimal implementation across whatever hardware and software stack is available — AMD, NVIDIA, Intel — so “the correct, optimised container appears.” This moves the burden of hardware optimization off researchers and onto platform engineers.

Governance as Foundation

Adam Huffman (Oxford/EESSI) reframed the whole effort as fundamentally a governance challenge, tracing EasyBuild’s evolution from a single-maintainer project to one with formal steering committees, explicit lifecycle policies, and structured decision-making — a proven model for sustainability.

MVP Consensus

Participants converged on a few concrete principles:

• Federate validated metadata descriptions, not multi-gigabyte container images
• Use OCI and ORAS standards rather than building new infrastructure
• Enforce profile validation and linting rather than relying on convention
• Put everything — recipes, metadata, policies — into version control and CI/CD
• Focus on small, real use cases with measurable value

Cross-Cutting Insights

Recurring tensions ran through the day: users want applications, not infrastructure; governance and funding decisions have to precede technical ones; metadata and recipes matter more than the built artifacts themselves; and treating build/teardown as versioned, first-class operations is what actually unlocks reproducibility at scale.

Takeaways Beyond HPC

Separate environment recipes from built images. Treat deployment as a versioned CI step. Move governance into the pull request pipeline instead of a separate process. Make vulnerability visibility explicit rather than assumed. EasyBuild’s reproducibility model is worth studying even outside HPC contexts.

Read the full article on Medium: What Do You Actually Federate? Notes from the 4th NFCS AI Container Registry Workshop