Model Registry & Training Artifacts¶

This page explains how trained models are managed in the Snowflake Model Registry.

What is the Model Registry?¶

The Snowflake Model Registry is a centralised store for ML models that provides:

Versioning: multiple versions of a model coexist.
Metadata: each version carries metrics, hyperparameters, and lineage.
Lifecycle management: models can be promoted through stages (development, staging, production).
Lineage: trace which data and features produced a model.

Model registration flow¶

graph TD
    A["Training DAG runs"] --> B["Train XGBoost model"]
    B --> C["Evaluate on validation set"]
    C --> D["Log model to Registry"]
    D --> D1["Model name: PUDO__XGBOOST_CAPACITY"]
    D --> D2["Version: auto-incremented"]
    D --> D3["Metrics: RMSE, MAE, R²"]
    D --> D4["Hyperparameters"]
    D --> D5["Lineage: dataset, feature_views"]

Model versioning¶

Each training run creates a new model version:

Attribute	Description
Name	`PUDO__<MODEL>` (project-prefixed).
Version	Auto-incremented (e.g., v1, v2, v3).
Metrics	RMSE, MAE, R² on the validation set.
Created	Timestamp of registration.
Lineage	References to the dataset and feature views used.

Version selection for inference¶

The inference pipeline loads a model version based on configuration:

# config/inference/base.yaml
model:
  name: PUDO__XGBOOST_CAPACITY
  version: latest   # or a specific version number

Using latest means inference always uses the most recently registered version. Pinning a specific version provides stability during evaluation periods.

Training artifacts¶

Beyond the model itself, the training pipeline produces:

Artifact	Where it lives	Purpose
Trained model	Model Registry	Used for inference.
Training metrics	Model Registry metadata	Model comparison and selection.
Datasets	Snowflake tables	Train/val/test splits for reproducibility.
Evaluation results	Project schema	Validation metrics per model version.

Lineage¶

Snowflake ML provides built-in lineage tracking:

# Load a model and explore its lineage
model = registry.load_model("PUDO__XGBOOST_CAPACITY", version=3)

# What data created this model?
lineage = model.lineage
print(lineage.upstream)   # Datasets, feature views
print(lineage.downstream) # Predictions, evaluations

Lineage is useful for:

Debugging: trace poor predictions back to the data and features that trained the model.
Compliance: demonstrate which data influenced a specific model version.
Impact analysis: understand what would be affected by changing a feature view.

Distributed training¶

The training pipeline uses Snowflake Container Services for distributed XGBoost training:

A compute pool provides the execution environment.
The training job runs in a container with XGBoost and dependencies.
Data is read from Snowflake stages. No data leaves the platform.
Training is orchestrated by the training DAG.

Configuration¶

Training parameters are controlled by YAML configuration:

# config/training/base.yaml
training:
  n_estimators: 500
  max_depth: 6
  learning_rate: 0.05
  compute_pool: PUDO_TRAINING_POOL

Environment overlays adjust these for dev (faster, smaller) vs. prod (full training).