Skip to content

The Model Registry: Versioning Models Like Code

What was run

The real model from Level 1's capstone run registered for the first time with mlflow.register_model(), creating a real, versioned entry in MLflow's real Model Registry — confirmed afterward by querying the registry's own API directly, not by trusting the script's printed output.

A real registration, and a real confirmation of last level's finding

RUN_ID = "bf0a9dabb13d4e23ad29e692492fa126"
MODEL_URI = f"runs:/{RUN_ID}/model"
result = mlflow.register_model(MODEL_URI, "breast-cancer-classifier")

Successfully registered model 'breast-cancer-classifier'.
WARNING mlflow.tracking._model_registry.fluent: Run with id bf0a9dab... has no artifacts at artifact
path 'model', registering model based on models:/m-b0dbb16731a44c0f8b4c94ea17d2ff36 instead
Created version '1' of model 'breast-cancer-classifier'.
registered: name=breast-cancer-classifier version=1 source=models:/m-b0dbb16731a44c0f8b4c94ea17d2ff36

That real warning is MLflow's own code independently confirming Level 1, Lesson 1's own real finding from a different angle: the old-style runs:/<run_id>/model path genuinely has no artifacts there anymore in MLflow 3.x — the registry itself has to redirect to the real, new Logged Models URI (models:/m-b0dbb16731a44c0f8b4c94ea17d2ff36) to find the actual model. This isn't a workaround or an error; it's MLflow's own registration logic handling its own real architecture change transparently.

The real registration, confirmed independently

curl -s ".../registered-models/get?name=breast-cancer-classifier"

{"registered_model": {"name": "breast-cancer-classifier", "latest_versions": [
  {"version": "1", "current_stage": "None", "status": "READY",
   "source": "models:/m-b0dbb16731a44c0f8b4c94ea17d2ff36", "run_id": "bf0a9dab..."}
]}}

A real, queryable registry entry: version 1, status READY, with an unbroken real chain back to the exact source run that produced it — not a file copied somewhere and renamed, but a tracked, verifiable lineage.

How It Actually Works

The Model Registry adds a layer on top of experiment tracking that solves a different real problem: tracking answers "what happened during training" (every run from Level 1); the registry answers "which of these trained models is actually a candidate for production, and what version is it?" A registered model (like breast-cancer-classifier here) is a named, stable identity that can accumulate real versions over time — version 1 today, version 2 next month after retraining — the same way a package name on a real package registry accumulates releases, with each version traceable back to its exact real source (here, directly back to the specific run and its exact data hash, params, and metrics from Level 1).

This is a genuinely different concern from a raw MLflow run, which is cheap, numerous, and often exploratory — Level 1 logged 25 real runs in under 2 seconds, and most of them were never meant to go anywhere near production. Registering a model is a deliberate, real act of promotion: out of potentially hundreds of real tracked runs, this specific one is being marked as a real candidate worth naming, versioning, and — in the next lesson — moving through a real stage lifecycle toward production.

Common mistakes

  • Treating every tracked run as automatically "the model" without a real, deliberate registration step — tracking and registration answer different real questions, and conflating them means production deployment decisions get made from the same noisy pool as every exploratory experiment.
  • Registering a model under a vague or inconsistent name across different real training scripts, fragmenting what should be one real, continuous version history into several disconnected ones.
  • Assuming the source field in a registered model's real metadata is cosmetic — as this lesson's own real investigation showed, it's the actual, functioning link the registry uses to locate the real model files, not just a label.

Exercise

Register a second real model version: retrain with max_depth=10 (reusing Level 1's capstone script with that change), then register the resulting new run's model under the same real name, breast-cancer-classifier. Query .../registered-models/get again and confirm it now shows a real version: 2 alongside version 1, each with its own distinct, correct run_id and source — a real, growing version history for one named model.