MLflow

Summary

MLflow is used for experiment tracking for both training and inference runs.

Local setup

Run UI locally:

make mlflow_ui

Default endpoint:

  • http://localhost:5000

Experiments

  • Training experiment: anomaly_detection (configurable)
  • Inference experiment: anomaly_detection_inference (configurable)

What is logged

Training logs:

  • model and hyperparameter settings
  • evaluation metrics (point, changepoint, NAB)
  • drift metrics
  • model artifacts and metadata files

Inference logs:

  • source/model metadata
  • anomaly_rate, avg_score
  • inference drift metrics
  • predictions artifact and drift report

Environment configuration

Configure in .env:

  • MLFLOW_TRACKING_URI
  • MLFLOW_EXPERIMENT_NAME
  • MLFLOW_INFERENCE_EXPERIMENT_NAME
  • MLFLOW_REGISTERED_MODEL_PREFIX

Model registry

Training can register models and set the champion alias when backend supports registry operations.