Tracking experiments

What MLflow is

What

An open-source platform for the machine learning life cycle: tracking experiments, packaging models, a model registry, serving, evaluation, and tracing for GenAI apps.

Why

It replaces spreadsheets of results and folders named model_final_v3.pkl with a searchable record of every run and model.

Experiments and runs

What

A run is one execution of training or evaluation code. Runs are grouped into experiments, usually one per problem or model family.

Why

Every result becomes traceable to the code, settings and data that produced it.

import mlflow

mlflow.set_tracking_uri("http://localhost:5000")
mlflow.set_experiment("churn-model")

with mlflow.start_run(run_name="rf-baseline") as run:
    mlflow.log_params({"n_estimators": 200, "max_depth": 8})
    clf.fit(X_train, y_train)
    mlflow.log_metric("val_auc", score_auc(clf, X_val, y_val))
print(run.info.run_id)

Parameters

Also called log_param, log_params

What

Inputs to a run, such as hyperparameters and settings, stored as strings. A parameter is written once per run and cannot be changed afterwards.

Why

Parameters are what you filter and compare on when asking "which settings gave the best score?".

Metrics

Also called log_metric, step

What

Numeric results logged over time. Each metric keeps its full history, with an optional step such as the epoch, and the UI plots it.

Why

Learning curves reveal overfitting and divergence that a single final number hides.

for epoch in range(epochs):
    train_loss = train_one_epoch(model, train_loader)
    val_loss = evaluate(model, val_loader)
    mlflow.log_metrics(
        {"train_loss": train_loss, "val_loss": val_loss},
        step=epoch,
    )

Tags

Also called set_tag

What

Free-form key-value labels on runs, experiments and models. MLflow also sets system tags, such as the source file, user and Git commit.

Why

Tags such as team, purpose or dataset_version make runs easy to find and filter later.

Artefacts

Also called artifacts, log_artifact

What

Output files attached to a run: plots, reports, feature lists, model files, sample predictions. They go to the artefact store, not the database.

Why

Everything needed to understand a run sits beside its metrics instead of on someone's laptop.

mlflow.log_artifact("reports/confusion_matrix.png")
mlflow.log_artifacts("reports/", artifact_path="reports")
mlflow.log_dict({"features": feature_names}, "features.json")
mlflow.log_text(classification_report(y, y_pred), "report.txt")
mlflow.log_figure(fig, "plots/roc.png")

Dataset tracking

Also called log_input, mlflow.data

What

mlflow.data wraps a DataFrame or table with its source, name, digest and schema; log_input records it on the run with a context such as training or evaluation.

Why

You can tell exactly which data version trained a model, and spot two runs that silently used different data.

dataset = mlflow.data.from_pandas(
    train_df,
    source="s3://data/churn/2026-09.parquet",
    name="churn-train",
    targets="churned",
)
mlflow.log_input(dataset, context="training")

System metrics

What

With log_system_metrics=True on a run (the psutil package is needed, plus nvidia-ml-py for GPUs), MLflow samples CPU, memory, disk and GPU use during training.

Why

You see whether a slow job was starved of GPU memory or stuck on data loading.