Here's a scenario: you've trained six versions of your model. Version 3 performed best in testing, but version 5 is in production. Someone asks which training data version 5 used. You have no idea. Sound familiar? That's model versioning chaos, and it will destroy your AI project if you let it.
What Needs Versioning
Everything that affects your model's behavior needs to be tracked:
- Model weights: The actual trained parameters
- Training code: The scripts used for training
- Training data: Which data version was used
- Hyperparameters: All the settings
- Dependencies: Library versions, Python version
- Environment: Docker image, hardware specs
Versioning Strategies
Semantic versioning: Model v1.2.3 - major.minor.patch. Clear but manual.
Hash-based: Version = hash of everything that went into training. Automatic but not human-readable.
Hybrid: Semantic for major changes, hash for reproducibility.
Tools That Help
MLflow: Tracks experiments, registers models, manages lifecycle. Open source and widely used.
Weights & Biases: Excellent experiment tracking. Great visualizations of training runs.
Neptune.ai: Another experiment tracking option with good integration support.
DVC (Data Version Control): Git for data. Version control for datasets and models.
Model Registry
A model registry is a centralized catalog of your models. It tracks:
- Model versions and their lineage
- Performance metrics
- Deployment status
- Who trained/approved each version
Every production model should be registered. Never deploy an unregistered model.
Practical Tips
- Start versioning from day one—it's harder to add later
- Automate as much as possible
- Log everything (hyperparameters, metrics, git commit)
- Tag production models clearly
- Keep previous versions deployable for fast rollback
The goal isn't perfection—it's reproducibility. When something goes wrong, you need to be able to recreate exactly what happened.