ML Models
Manage machine learning experiments, model registry, and inference through MLflow integration.
Overview
FacilFlow integrates with MLflow to provide a complete ML lifecycle:
- Track experiments and runs
- Register and version models
- Manage staging transitions with approval workflows
- Run inference in pipelines via the ML Inference node
Experiments
Navigate to Build > ML Models to see experiments.
Creating an Experiment
- Click New Experiment
- Provide a name and optional description
- Select the artifact storage location
Viewing Runs
Each experiment contains runs. A run tracks:
| Field | Description |
|---|---|
| Parameters | Hyperparameters used for training |
| Metrics | Evaluation metrics (accuracy, RMSE, etc.) |
| Artifacts | Model files, plots, logs |
| Tags | Custom metadata |
| Duration | Training time |
Compare runs side-by-side to find the best performing model.
Model Registry
Promote a run’s model to the registry for versioning and deployment.
Registering a Model
- Open a run with a logged model
- Click Register Model
- Choose an existing registered model name or create a new one
- The model is added as a new version
Model Versions
Each registered model tracks versions with:
- Version number (auto-incremented)
- Source run link
- Stage (see below)
- Description
Staging Workflow
Models move through stages:
| Stage | Description |
|---|---|
| None | Newly registered, not yet reviewed |
| Staging | Under evaluation and testing |
| Production | Approved for production inference |
| Archived | Retired from active use |
Transitioning Stages
- Open a model version
- Click Transition Stage
- Select the target stage
- Add a comment explaining the transition
Approval Workflows
Transitions to Production require approval:
- A user requests the transition
- Designated approvers are notified
- An approver reviews the model metrics and approves or rejects
- On approval, the model moves to Production
Feature Engineering
The Feature Engineer pipeline node computes derived features for ML:
| Feature Type | Examples |
|---|---|
| Rolling statistics | Mean, std, min, max over a window |
| Lag features | Value at t-1, t-5, t-60 |
| Ratios | Feature A / Feature B |
| Time-based | Hour of day, day of week, is_weekend |
| Aggregates | Count, sum over grouped windows |
Configure feature engineering in the pipeline builder by adding a Feature Engineer node between your source and ML Inference nodes.
Inference
Use registered models for real-time scoring in pipelines:
- Add an ML Inference node to your pipeline
- Select the registered model and version (or use “Production” to auto-select)
- Map input features from upstream nodes to model inputs
- The node outputs predictions downstream
The inference node loads the model from MLflow and applies it to each incoming data batch.