Track ML runs, manage model registry stages, configure tracking servers, log artefacts, and use autologging with scikit-learn and PyTorch.
0 / 5 completed
1 / 5
What is an MLflow run and what information does it capture?
MLflow run: created with mlflow.start_run() or auto-generated during a training loop. Each run stores log_param("lr", 0.001), log_metric("accuracy", 0.95, step=100), and file artefacts. Runs are grouped into experiments and can be compared in the MLflow UI or queried via the Python API.
2 / 5
What does the MLflow Model Registry provide beyond simple artefact storage?
Model Registry: you register a run's model artefact as a named model version. Teams move versions through stages with client.transition_model_version_stage("MyModel", 3, "Production"). Serving infrastructure reads the "Production" alias, so promoting a new version automatically updates what gets served without code changes.
3 / 5
What is the purpose of the MLflow tracking server versus logging locally?
Tracking server: without a server, runs are written to the local ./mlruns folder. A tracking server with a PostgreSQL backend and S3 artefact store allows every experiment to be logged centrally. Set MLFLOW_TRACKING_URI=https://mlflow.company.com and all runs from any machine or CI pipeline flow into the same shared history.
4 / 5
What are MLflow artefacts and how are they used in the model lifecycle?
Artefacts:mlflow.log_artifact("confusion_matrix.png") or mlflow.sklearn.log_model(model, "model") attach files to the run. The logged model artefact includes a MLmodel file with flavour metadata, enabling mlflow models serve -m runs:/<run_id>/model to launch a REST endpoint without writing serving code.
5 / 5
What is MLflow autologging and which frameworks support it?
Autologging: call mlflow.autolog() before training. MLflow patches the framework's fit/train methods to capture hyperparameters from the estimator constructor, training metrics at each epoch or iteration, and the final model — all without any additional instrumentation. This significantly reduces boilerplate in experiment scripts.