Skip to content

MLflow Tracking

mlflow-experience integrates with MLflow via MLflowTracker (src/core/trackers/mlflow_tracker.py) to give you complete visibility into every experiment run.

1. What Gets Tracked Automatically?

For every pipeline run, MLflowTracker records:

  1. Parameters:
  2. Model type and hyperparameters.
  3. Dataset paths and splitting ratios.
  4. Metrics:
  5. Training and testing scores (Accuracy, F1 Score).
  6. Artifacts:
  7. Full console and file execution logs (logs/<run_name>.log).
  8. Trained model binaries.
  9. Tags & Metadata:
  10. Project name, owner, execution environment.
  11. Git commit hash and active branch.

2. Viewing Results in the MLflow UI

  1. Start your MLflow server (either via Docker or locally).
  2. Open http://localhost:5000 in your browser.
  3. You will see:
  4. Experiment List: Listed on the left (e.g. customer_churn_experiments).
  5. Run History: Table of every execution, including timestamps, durations, and metric scores.
  6. Comparison View: Select multiple runs and click Compare to see side-by-side metric charts and hyperparameter differences.
  7. Artifact Browser: Click any run and select the Artifacts tab to inspect the saved session log file.

3. Running MLflow Locally (Without Docker)

You don't need Docker to use MLflow! You can run a lightweight local MLflow tracking server directly with uv:

uv run mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./mlruns --port 5000

This creates: - mlflow.db: A local SQLite file that holds all your experiment metadata. - ./mlruns: A local directory where artifacts and models are stored.

Or use Direct File Tracking

You can even configure your YAML to write directly to a local directory without starting any server:

mlflow:
  experiment_name: customer_churn_experiments
  tracking_uri: file:./mlruns