Model Registry

Centralized versioned store for trained models — the boundary between research and production. Covers staging lifecycle, lineage, model cards, and the 2026 registry landscape (MLflow as default OSS, W&B Models, Vertex/SageMaker registries, Hugging Face Hub for foundation models).

Overview

What a registry stores

A model registry is a versioned catalog of trained model artifacts, the formal boundary between research and the production deployment pipeline. Without one, "the production model" is whatever pickle file someone last uploaded to S3 with a timestamped filename.

  • Artifact reference — Pointer to weights in object storage plus a cryptographic hash
  • Version number — Monotonically increasing and immutable
  • Lifecycle stage — None, Staging, Production, Archived
  • Model signature — Input and output schema, enforced at deploy time
  • Lineage — Training run ID, dataset version, git SHA, hyperparameters, metrics, environment digest
  • Model card — Intended use, training data, performance by slice, limitations

The promotion lifecycle

  • None to Staging: signature verification, holdout evaluation and behavioral tests pass.
  • Staging to Production: a human approval or an automated gate; serving polls for the current version or receives a webhook.
  • Production to Archived: the previous version is retained for rollback but not served.

> [!TIP] Rollback is a registry operation: promote the previous Archived version back to Production. The serving layer picks it up automatically, with no manual file swaps.

MLflow in depth

MLflow is the dominant open-source registry. It originated at Databricks in 2018 and was donated to LF AI and Data in 2020.

  • Tracking — Experiments, runs, parameters, metrics, artifacts
  • Models — The MLmodel format that packages weights with their loading flavor
  • Model Registry — Versioned catalog with stages or aliases
  • Projects — Reproducible code packaging (less widely adopted)

> [!NOTE] The pyfunc flavor is the lowest common denominator: any MLflow model can be called as a Python predict function, whatever the framework.

Stages vs aliases

MLflow 2.x deprecated hard-coded stages in favor of aliases such as champion, challenger and production-blue, which allow several concurrent production versions. MLflow 3.0 added Prompts and Evaluation as first-class objects and deepened autologging for LangChain, LlamaIndex and Anthropic.

The 2026 landscape

  • MLflow — Default open-source, self-hosted registry
  • Weights and Biases Models — Deep-learning workflows: lineage graphs, visual comparison, native W&B integration
  • Vertex AI, SageMaker, Azure ML registries — Pragmatic default for cloud-committed teams
  • Hugging Face Hub — De facto registry for foundation model weights: open and gated models, standardized model cards, dataset lineage

> [!IMPORTANT] The registry is governance infrastructure. You should be able to answer "what is in production, why, and what would we roll back to?" in seconds.

Interview questions

  • What does a model registry store and why is it the critical boundary between research and production?
  • Walk through an MLflow workflow from training to production promotion.
  • How do you handle model versioning for a team releasing multiple models daily?
  • When would you choose W&B Models over MLflow? When would you use Hugging Face Hub?
  • What are the most common model registry anti-patterns and how do you fix them?

Sign up free to read the full topic

More MLOps & LLMOps topics

Start free on Cariara · Coding problems · System design studio