MLflow Cheat Sheet

Updated 2026-05-21

🧠Study flashcards on this topic96 cards · spaced repetition→

MLflow is an open-source platform for managing the full machine learning lifecycle — from experiment tracking and model packaging to registry management and production deployment. It sits at the center of modern MLOps workflows, giving data scientists a unified interface regardless of which framework (scikit-learn, PyTorch, TensorFlow, XGBoost, or LLMs) they use. The key mental model: every piece of work in MLflow is a run inside an experiment, every run can log params, metrics, and artifacts, and every trained artifact can be promoted to a versioned Model Registry — making any team's experiments reproducible, comparable, and deployable with minimal overhead.

What This Cheat Sheet Covers

This topic spans 20 focused tables and 142 indexed concepts, 96 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Tracking API — Runs and ExperimentsTable 2: Artifact LoggingTable 3: Dataset Tracking (mlflow.data)Table 4: AutologgingTable 5: Model Flavors and LoggingTable 6: Model Signatures and Input ExamplesTable 7: Custom pyfunc ModelsTable 8: Model RegistryTable 9: Search and Query APITable 10: Nested Runs and Hyperparameter TuningTable 11: System Metrics LoggingTable 12: Model Evaluation (mlflow.evaluate)Table 13: GenAI / LLM Tracking and TracingTable 14: Prompt RegistryTable 15: LLM / GenAI EvaluationTable 16: Model Serving and DeploymentTable 17: MLflow ProjectsTable 18: MLflow RecipesTable 19: Tracking Server ArchitectureTable 20: AI Gateway

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Tracking API — Runs and Experiments

The tracking API is the foundation of everything in MLflow. Understanding how to create experiments, start runs, and log data programmatically is the entry point before using autologging, the UI, or any advanced feature.

Function	Example	Description
mlflow.set_experiment()	`mlflow.set_experiment("fraud-detection")`	• Sets the active experiment by name • creates it if it doesn't exist • All subsequent runs go into this experiment
mlflow.start_run()	`with mlflow.start_run(run_name="v1") as run:` `mlflow.log_param("lr", 0.01)`	• Opens a new tracking run as a context manager • automatically calls `end_run()` on exit, even if an exception is raised • Preferred over manual `end_run()`.
mlflow.log_param()	`mlflow.log_param("n_estimators", 100)`	Logs a single key-value hyperparameter (both as strings). Call once per param per run — duplicate keys are overwritten.
mlflow.log_params()	`mlflow.log_params({"lr": 0.01, "epochs": 10})`	• Logs multiple parameters at once from a dict • equivalent to calling `log_param()` in a loop but more efficient
mlflow.log_metric()	`mlflow.log_metric("accuracy", 0.94, step=5)`	• Logs a numeric metric • optional `step` enables time-series charting in the UI • MLflow keeps the full history of values
mlflow.log_metrics()	`mlflow.log_metrics({"mae": 0.12, "rmse": 0.34})`	• Logs multiple metrics simultaneously • accepts optional `step` and `timestamp`.
mlflow.set_tag()	`mlflow.set_tag("team", "nlp-squad")`	• Attaches a string metadata tag to the run • tags appear in the UI and are searchable with `ILIKE`.
mlflow.set_tags()	`mlflow.set_tags({"env": "prod", "version": "2.1"})`	Sets multiple tags at once from a dict.
mlflow.active_run()	`run = mlflow.active_run(); print(run.info.run_id)`	• Returns the currently active run object • returns `None` if no run is active • Useful when run ID is needed inside nested functions
mlflow.last_active_run()	`run = mlflow.last_active_run()`	• Returns the most recently completed run, not just the active one • Useful after a `with` block ends
mlflow.get_experiment_by_name()	`exp = mlflow.get_experiment_by_name("fraud-detection")`	• Retrieves an Experiment object by case-sensitive name • returns `None` if not found
mlflow.create_experiment()	`mlflow.create_experiment("new-exp", artifact_location="s3://my-bucket/")`	• Creates a new experiment with optional custom artifact location and tags • returns the experiment ID string
mlflow.set_tracking_uri()	`mlflow.set_tracking_uri("http://myserver:5000")`	• Points the client at a remote tracking server • equivalent to setting `MLFLOW_TRACKING_URI` env var

Table 2: Artifact Logging

Artifacts are files associated with a run — models, plots, data files, configs. MLflow stores them in the artifact store and links them to the run. Choosing the right artifact logging method avoids unnecessary serialization overhead.

Function	Example	Description
mlflow.log_artifact()	`mlflow.log_artifact("model.pkl")`	• Logs a local file or directory to the run's artifact URI • The file is copied as-is • use `artifact_path` arg to organize into subfolders

The rest of this section is part of the full cheat sheet

Table 3: Dataset Tracking (mlflow.data)

MLflow's dataset tracking module captures metadata about training and evaluation data — schema, digest, source lineage — and attaches it to runs. This is separate from logging raw data files; it records what data was used without duplicating the data itself.

Function	Example	Description
mlflow.data.from_pandas()	`ds = mlflow.data.from_pandas(df, source="data/train.csv", name="train-v2")`	• Wraps a Pandas DataFrame into an MLflow Dataset • computes digest hash and schema automatically

The rest of this section is part of the full cheat sheet

Table 4: Autologging

Autologging eliminates manual log_param / log_metric calls by hooking into framework training functions automatically. It can be enabled globally or per-framework, and offers fine-grained control over what gets captured.

Function	Example	Description
mlflow.autolog()	`mlflow.autolog()` `model.fit(X_train, y_train)`	• Enables autologging for all installed supported libraries at once • intercepts `fit()` calls automatically • The simplest entry point

The rest of this section is part of the full cheat sheet

Table 1: Core Tracking API — Runs and Experiments

Function	Example	Description
mlflow.set_experiment()	`mlflow.set_experiment("fraud-detection")`	• Sets the active experiment by name • creates it if it doesn't exist • All subsequent runs go into this experiment
mlflow.start_run()	`with mlflow.start_run(run_name="v1") as run:` `mlflow.log_param("lr", 0.01)`	• Opens a new tracking run as a context manager • automatically calls `end_run()` on exit, even if an exception is raised • Preferred over manual `end_run()`.
mlflow.log_param()	`mlflow.log_param("n_estimators", 100)`	Logs a single key-value hyperparameter (both as strings). Call once per param per run — duplicate keys are overwritten.
mlflow.log_params()	`mlflow.log_params({"lr": 0.01, "epochs": 10})`	• Logs multiple parameters at once from a dict • equivalent to calling `log_param()` in a loop but more efficient
mlflow.log_metric()	`mlflow.log_metric("accuracy", 0.94, step=5)`	• Logs a numeric metric • optional `step` enables time-series charting in the UI • MLflow keeps the full history of values
mlflow.log_metrics()	`mlflow.log_metrics({"mae": 0.12, "rmse": 0.34})`	• Logs multiple metrics simultaneously • accepts optional `step` and `timestamp`.
mlflow.set_tag()	`mlflow.set_tag("team", "nlp-squad")`	• Attaches a string metadata tag to the run • tags appear in the UI and are searchable with `ILIKE`.
mlflow.set_tags()	`mlflow.set_tags({"env": "prod", "version": "2.1"})`	Sets multiple tags at once from a dict.
mlflow.active_run()	`run = mlflow.active_run(); print(run.info.run_id)`	• Returns the currently active run object • returns `None` if no run is active • Useful when run ID is needed inside nested functions
mlflow.last_active_run()	`run = mlflow.last_active_run()`	• Returns the most recently completed run, not just the active one • Useful after a `with` block ends
mlflow.get_experiment_by_name()	`exp = mlflow.get_experiment_by_name("fraud-detection")`	• Retrieves an Experiment object by case-sensitive name • returns `None` if not found
mlflow.create_experiment()	`mlflow.create_experiment("new-exp", artifact_location="s3://my-bucket/")`	• Creates a new experiment with optional custom artifact location and tags • returns the experiment ID string
mlflow.set_tracking_uri()	`mlflow.set_tracking_uri("http://myserver:5000")`	• Points the client at a remote tracking server • equivalent to setting `MLFLOW_TRACKING_URI` env var

Table 2: Artifact Logging

Function	Example	Description
mlflow.log_artifact()	`mlflow.log_artifact("model.pkl")`	• Logs a local file or directory to the run's artifact URI • The file is copied as-is • use `artifact_path` arg to organize into subfolders

Table 3: Dataset Tracking (mlflow.data)

Function	Example	Description
mlflow.data.from_pandas()	`ds = mlflow.data.from_pandas(df, source="data/train.csv", name="train-v2")`	• Wraps a Pandas DataFrame into an MLflow Dataset • computes digest hash and schema automatically

Table 4: Autologging

Function	Example	Description
mlflow.autolog()	`mlflow.autolog()` `model.fit(X_train, y_train)`	• Enables autologging for all installed supported libraries at once • intercepts `fit()` calls automatically • The simplest entry point