MLflow has four building blocks: Tracking (log params/metrics/artifacts per run), Projects (package reproducible runs), Models (standard packaging across frameworks), and the Model Registry (version, alias, and govern models on their way to production).
Distilled & cross-checked across: mlflow.org/docs/latest/ml · grafdavid.com/cheatsheets/mlflow · github.com/aishwaryaprabhat/MLflow
pip install mlflow★pip install mlflow[extras]mlflow.set_tracking_uri("file:./mlruns")os.environ["MLFLOW_TRACKING_URI"]mlflow.set_experiment("name")★with mlflow.start_run():★mlflow.start_run(run_name="x")mlflow.start_run(run_id=id)mlflow.start_run(nested=True)mlflow.end_run()mlflow.log_param("lr", 0.01)★mlflow.log_params(dict)mlflow.log_metric("acc", 0.95, step=e)★mlflow.log_metrics(dict)mlflow.log_artifact("plot.png")★mlflow.log_artifacts("./plots")mlflow.set_tag("k", "v")mlflow.autolog()★mlflow.sklearn.autolog()★mlflow.tensorflow.autolog()MLFlowLogger(...) PyTorch Lightningmlflow.sklearn.log_model(model, "model")★mlflow.pytorch.log_model(...)mlflow.transformers.log_model(...)mlflow.pyfunc.PythonModellog_model(..., registered_model_name="x")★mlflow.register_model(model_uri, "name")★client = MlflowClient()★client.set_registered_model_alias(name, "champion", v)★client.set_model_version_tag(...)client.transition_model_version_stage(...)deprecatedmlflow.pyfunc.load_model("models:/name@champion")★mlflow.pyfunc.load_model("models:/name/3")mlflow.flavor.load_model("runs:/id/model")mlflow models serve -m models:/name@champion -p 1234★curl -d '{"instances":[[..]]}' …/invocationsmlflow ui★mlflow ui --port 5001 --host 0.0.0.0mlflow ui --backend-store-uri postgresql://…mlflow experiments listmlflow runs list --experiment-id 1mlflow doctormlflow.search_runs(experiment_ids=["1"])★filter_string="metrics.accuracy > 0.9"★mlflow.get_run(run_id)client.search_model_versions("name='x'")MLproject★conda.yaml / python_env.yamlmlflow.run("uri#", parameters={…})mlflow run . -P lr=0.01mlflow.start_run(nested=True)★log_metric() inside a tight loopslowlog_artifact() not log_param()mlflow.set_tag("framework", …)runs:/<run_id>/modelmodels:/<name>/3models:/<name>@championmodels:/<name>/ProductiondeprecatedFrom a logged artifact to a governed, servable model — and how an MLflow Project folder is laid out. Diagrams render via Mermaid; based on the official MLflow Model Registry guide.
Stages (Staging/Production/Archived) are deprecated since MLflow 2.9. Aliases are flexible, mutable, and multiple can point at one version.
flowchart LR
A["mlflow.<flavor>.log_model()"] --> B["register_model() → v1, v2, v3 …"]
B --> C{"set_registered_model_alias"}
C -->|"@champion"| D["serves prod traffic"]
C -->|"@challenger"| E["A/B test candidate"]
D -->|"reassign on rollback"| B
Conceptually the same goal — label which version is "live" — but aliases drop the rigid 4-state machine.
flowchart TD
subgraph Deprecated["Stages (deprecated 2.9+)"]
direction LR
N["None"] --> S["Staging"] --> P["Production"] --> Ar["Archived"]
end
subgraph Current["Aliases (current)"]
direction LR
V1["v1 @staging"]
V2["v2 @champion"]
V3["v3 @challenger"]
end
Two files make any script reproducible on another machine: the entry-point spec and the environment definition.
flowchart LR
Root["MLflow_project/"] --> MLP["MLproject — entry points + params"]
Root --> ENV["conda.yaml / python_env.yaml — dependencies"]
Root --> SRC["train.py — your code"]
MLP -. "mlflow run ." .-> RUN["Reproducible run, anywhere"]