TopolearnConsistent logging is what makes experiments comparable weeks later.
An experiment groups related runs, e.g. one per project or model family. mlflow.set_experiment("churn") selects it (creating it if needed). A run is one execution, opened with a context manager so it is closed even when the code raises:
with mlflow.start_run(run_name="rf-depth-8"):
...What to log:
- Parameters: inputs that define the run, logged once (
log_param,log_params(dict)). Values are stored as strings; logging the same key again with a different value raises an error. - Metrics: numbers that can change over time, with an optional
step(log_metric("loss", v, step=epoch),log_metrics(dict, step=...)). The UI plots them as curves. - Tags: free-form metadata such as git commit, dataset version or owner (
set_tag,set_tags). MLflow adds system tags like the source file and user automatically.
The run object exposes run.info.run_id, used to reload results or models later. Outside a with block, call mlflow.end_run() yourself; the run's status becomes FINISHED, or FAILED if the block raised an exception.
Log parameters for everything that affects results, including data version and random seed. A run without them can be seen but not reproduced.
Example
import mlflow
mlflow.set_experiment("churn-model")
with mlflow.start_run(run_name="gbm-lr-0.1") as run:
mlflow.log_params({"learning_rate": 0.1, "max_depth": 4, "seed": 42})
mlflow.set_tags({"data_version": "2026-10-01", "owner": "ml-team"})
for epoch in range(3):
mlflow.log_metric("val_loss", 0.9 / (epoch + 1), step=epoch)
mlflow.log_metric("val_auc", 0.87)
finished = mlflow.get_run(run.info.run_id)
print(finished.info.status) # FINISHED
print(finished.data.params) # {'learning_rate': '0.1', 'max_depth': '4', 'seed': '42'}
print(finished.data.metrics["val_loss"]) # 0.3 (latest value)Remember
- Use
with mlflow.start_run()so runs always close with the right status - Params are logged once as strings; metrics can have steps
- Tag runs with data version, commit and owner for later filtering
Watch out for
- Logging a changing value such as epoch loss with log_param, which fails on the second call; use log_metric.