Home › Key concepts › Versioning data, code and models
Key concept · Walk (intermediate) · the same on every platform
Versioning data, code and models
Be able to say exactly which data and code made which model.
What it is
Reproducing a model needs three things: the code version, the data version and the settings. Experiment tracking records them for each run; a model registry stores the resulting models with versions and approvals.
Why it matters
When a model misbehaves you must be able to rebuild it, compare it and roll it back.
An example
Model v14 was trained on the 1 October snapshot with commit a3f9; rolling back to v13 takes one click.
The common mistake
Saving models as model_final_v2_REAL.pkl on a laptop.
Where you meet it on each platform
The idea stays the same; only the product name changes.
| Job | AWS | Azure | Google Cloud | Databricks | Open source |
|---|---|---|---|---|---|
| Experiment trackingremember which run, data and settings gave which score | SageMaker MLflow | Azure ML (MLflow) | Vertex AI Experiments | MLflow | MLflow |
| Model registryversions, approvals and rollback for models | SageMaker Model Registry | Azure ML registry | Vertex AI Model Registry | Models in Unity Catalog | MLflow registry |
Related ideas
Where it shows up
Part of the Walk level on these paths: ML Engineering (production ml).