MLflow organises ML work around four components: Tracking (log params, metrics, artifacts every run), Projects (package code so any machine reproduces it), Models (a framework-agnostic format), and the Model Registry (version, alias, and promote to production). Learn the map once and the API stops being a list to memorise.
Synthesised from: mlflow.org/docs/latest/ml · grafdavid.com/cheatsheets/mlflow · github.com/aishwaryaprabhat/MLflow
pip install mlflow★
pip install mlflow[extras]
mlflow ui★
mlflow ui --port 5001 --host 0.0.0.0
mlflow doctor
mlflow.set_experiment("my-exp")★
with mlflow.start_run():★
mlflow.start_run(run_name="exp-1")
mlflow.start_run(run_id=id)
mlflow.start_run(nested=True)★
mlflow.active_run()
mlflow.end_run()
mlflow.log_param("lr", 0.01)★
mlflow.log_params({"lr": 0.01, "bs": 32})★
mlflow.log_metric("acc", 0.95)★
mlflow.log_metric("loss", v, step=epoch)★
mlflow.log_metrics({...}, step=e)
mlflow.set_tag("team", "nlp")
mlflow.set_tags({...})
mlflow.log_input(dataset, context="train")
mlflow.log_artifact("plot.png")★
mlflow.log_artifacts("./out/")
mlflow.<flavor>.log_model(model, "model")★
mlflow.pyfunc.log_model("model", python_model=…)
mlflow.pyfunc.load_model(uri)★
mlflow.artifacts.download_artifacts(uri)
mlflow.<flavor>.autolog()★
mlflow.autolog()★
mlflow.sklearn.autolog()
mlflow.pytorch.autolog()
mlflow.tensorflow.autolog()
mlflow.autolog(log_models=False)
mlflow.get_run(run_id)
mlflow.search_runs(experiment_ids=["1"])★
filter_string="metrics.acc > 0.9"★
order_by=["metrics.acc DESC"]
mlflow.search_logged_models(...)v3
mlflow.search_experiments()
mlflow.delete_run(run_id)soft
MLproject★
conda.yaml / python_env.yaml
mlflow run .★
mlflow run <git-url>
mlflow run . -P lr=0.01
mlflow run . --env-manager=local
mlflow.run(uri, parameters={...})
mlflow.<flav>.log_model(m, "model",
registered_model_name="Iris")★
mlflow.register_model("runs:/id/model", "Iris")★
client.set_registered_model_alias(
"Iris", "champion", version=3)★
client.transition_model_version_stage(
name="Iris", version=2, stage="Production")
client.set_model_version_tag("Iris", 2, "status", "ok")
for m in client.search_registered_models():
client.delete_model_version("Iris", 1)permanent
mlflow models serve -m models:/Iris/Production★
mlflow models serve -m runs:/<id>/model -p 1234
POST /invocations
Body: {"instances": [[5.1, 3.5, 1.4, 0.2]]}★
mlflow models build-docker -m x
mlflow deployments create -t sagemaker
mlflow.pyfunc.spark_udf(spark, uri)
mlflow server --backend-store-uri sqlite:///mlflow.db★
mlflow experiments list
mlflow runs list --experiment-id 1
mlflow artifacts download -u runs:/<id>/model
mlflow gc
mlflow.set_tracking_uri("file:./mlruns")★
mlflow.set_tracking_uri("http://host:5000")★
MLFLOW_TRACKING_URI=http://…
MLFLOW_EXPERIMENT_NAME=my-exp
--backend-store-uri postgresql://…
--default-artifact-root s3://bucket/path
runs:/<run_id>/model
models:/<name>/<version>
models:/<name>/Production
models:/<name>@<alias>★
models:/<model_id>v3
Same project, four jobs. Tracking and Models are used on every run; Projects kick in for reproducibility; the Registry manages production promotion.
Logs every run's params, metrics, tags, and artifact files so experiments stay comparable and reproducible.
An MLproject file + env spec packages your code so anyone — or any machine — can reproduce a run exactly.
A standard packaging format: one saved model, loadable either as its native flavor or as a generic pyfunc.
Centrally versions a named model and tracks which version carries each alias/stage on the road to production.
Manual logging gives full control over every captured value; autologging is one line and handles the common case for most popular frameworks with no extra code.
import mlflow mlflow.set_experiment("text-clf") with mlflow.start_run(): # 1. Log hyperparameters once mlflow.log_params({"lr": 0.01, "epochs": 10}) # 2. Log metrics per step for epoch in range(10): loss = train(epoch) mlflow.log_metric("loss", loss, step=epoch) # 3. Save the trained model mlflow.sklearn.log_model(model, "model") # 4. Optional: tag for filtering mlflow.set_tag("framework", "sklearn")
import mlflow.sklearn from sklearn.ensemble import RandomForestClassifier # Patches sklearn BEFORE training starts mlflow.sklearn.autolog() with mlflow.start_run(): model = RandomForestClassifier(n_estimators=100) model.fit(X_train, y_train) # ↳ params, metrics, model all captured automatically # ── Nested runs for sweeps ────────────────────── with mlflow.start_run(run_name="sweep"): for lr in [0.01, 0.001, 0.0001]: with mlflow.start_run(nested=True): mlflow.log_param("lr", lr) # … train child trial …
The handful of errors that account for most "MLflow isn't working" moments.
| Symptom | Fix |
|---|---|
No such file or directory: mlruns | Run mlflow ui from the directory containing mlruns/, or set MLFLOW_TRACKING_URI explicitly. |
| Authentication failures (Azure / Databricks) | Re-check tokens, service-principal permissions, and that the tracking URI matches the workspace endpoint exactly. |
| Logging is very slow in training loops | Batch metrics with log_metrics({...}) and log every N steps rather than every step to reduce API round-trips. |
Large files silently fail in log_param | Use log_artifact() for files — log_param is for scalar/string values only. |
RestException: Run not found | Double-check experiment ID and run ID — runs are scoped to one experiment and one tracking server. |
Permission errors on mlruns/ | Check filesystem ownership and permissions on the backend store directory or S3 bucket policy. |
autolog captures nothing | Call mlflow.autolog() or the flavor-specific version before the training call, not inside the loop. |