Design a centralized model registry for a large enterprise
Tests ML artifact governance at scale. Strong answers cover immutable versioned artifacts with dependency manifests, a framework-agnostic API, and pluggable deployment targets. Red flag: treating models as opaque files without environment reproducibility.
What's really being asked
This tests whether you can design a platform service that decouples model artifacts from training frameworks and deployment targets while enforcing reproducibility and access control at enterprise scale. Interviewers want to see that you understand the difference between a file server and a governance layer, and that you can reason about metadata schemas, dependency hell, and multi-tenant isolation.
The full answer
First, immutable storage and versioning. Propose a content-addressable store for large binaries with semantic or calendar versioning for human-readable aliases, plus a metadata database that tracks lineage, training metrics, and approval workflows. Second, dependency management. Require each registered model to include a manifest specifying library versions, CUDA dependencies, and a base container image digest rather than assuming host environment compatibility. Third, API design. Recommend a framework-agnostic REST or gRPC API with paginated list endpoints, strong filtering by tags and frameworks, and role-based access control so teams remain isolated within their own namespaces. Fourth, deployment abstraction. Introduce a target-agnostic interface where the registry emits containerized artifacts or compiled formats like ONNX and TensorRT so that Kubernetes, edge devices, and batch pipelines consume the same source of truth without framework-specific coupling.
The mistakes people make
A major red flag is suggesting a shared NFS folder or S3 bucket with loose naming conventions as the registry. Another is ignoring dependency management entirely and assuming the deployment environment will match the training environment. Proposing framework-specific APIs such as a PyTorch-only serving endpoint also signals a lack of platform thinking. Finally, omitting metadata search and lineage means teams cannot audit which production model came from which training run.
What usually comes next
Interviewers often ask how you would handle schema evolution for model metadata without breaking existing clients, how to enforce immutability when a model artifact is terabytes in size, or how to design a promotion pipeline from staging to production. They may also probe how you would support A-B testing with multiple active versions or how to garbage collect unreferenced artifacts while respecting compliance retention rules.
A concrete example
Imagine a computer vision team registers a model with version 2.3.1. The registry stores the ONNX binary in object storage under a SHA-256 key and writes a metadata record containing the training dataset ID, the Python 3.11 and CUDA 12.2 container digest, and approval signatures. When the deployment team requests the latest approved version for the edge fleet, the API returns a signed URL for the binary plus the exact container manifest, ensuring the edge compiler can reproduce the runtime without guessing dependencies.
Interview question
Which architectural decision most effectively prevents dependency hell when promoting models from training to heterogeneous deployment targets?
- a.Require each registered model to include a manifest specifying library versions, CUDA dependencies, and a base container image digest.Correct
- b.Serialize models into a framework-specific format like PyTorch TorchScript and standardize on that framework across the company.
- c.Store models in a content-addressable object store with immutable versioning and role-based access control.
- d.Enforce that all training and deployment nodes use the same base operating system image.
Why? this is the answer
A manifest with container digests explicitly captures the runtime environment, ensuring deployment targets reproduce dependencies without assuming host compatibility. While immutable content-addressable storage with access control is essential for governance, it secures artifacts and enforces immutability without resolving runtime dependency mismatches.
Just read this? Test yourself on what you have been reading.
- #mlops
- #system-design
- #model-registry
- #machine-learning
- #infrastructure
Put your scrolling time to good use
Learn one idea, try a quiz and save useful cards for revision. Tezvyn makes it easy to learn and stay current in your tech field, a few minutes at a time.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.
We are hiring for this. Open roles that interview on mlops — each one lists the topics its interview covers.
See open roles