MLflow Skinny

3.10.1 · active · verified Sun Mar 29

MLflow Skinny is a lightweight Python package that provides core MLflow functionalities for experiment tracking and model management, omitting heavier dependencies like SQL storage, the MLflow UI, server, and extensive data science libraries. It serves as a foundation for users who need only the tracking and logging capabilities. MLflow is an open-source platform designed to streamline the entire machine learning lifecycle, supporting experiment tracking, reproducible code packaging, and model deployment. The current version is 3.10.1 and it requires Python >=3.10. The library maintains an active development status with frequent patch and minor releases, often on a monthly cadence.

Warnings

breaking MLflow 3.x changed the default tracking URI from file-based (./mlruns) to SQLite (sqlite:///mlflow.db). `mlflow-skinny` does not include `sqlalchemy`, `alembic`, or `sqlparse` by default, leading to `UnsupportedModelRegistryStoreURIException` if a tracking URI is not explicitly set or these dependencies are not manually installed.
Fix: Explicitly set `mlflow.set_tracking_uri('file:///path/to/mlruns')` to use a file-based store without extra dependencies, or `pip install mlflow-skinny sqlalchemy alembic sqlparse` to enable SQLite support.
gotcha `mlflow-skinny` explicitly excludes the MLflow UI and server components. Running `mlflow ui` after installing only `mlflow-skinny` will result in an error like 'Unable to display MLflow UI - landing page (index.html) not found'.
Fix: To use the MLflow UI and server, you must install the full `mlflow` package (`pip install mlflow`) or manually build the UI assets if running from source (not recommended for general users).
gotcha Many MLflow features, particularly model flavors (e.g., `mlflow.sklearn`, `mlflow.tensorflow`), model serving (`mlflow models serve`), or advanced artifact storage, require additional dependencies not bundled with `mlflow-skinny`. Trying to use these features without the necessary extra packages will lead to `ImportError` or `ModuleNotFoundError`.
Fix: Install `mlflow-skinny` along with the specific extra dependencies needed for your use case (e.g., `pip install mlflow-skinny scikit-learn pandas` for `mlflow.sklearn` functionality, or `pip install mlflow-skinny flask` for serving).
breaking MLflow 3.x introduced several breaking changes. For example, the `run_uuid` attribute on `RunInfo` objects was removed and replaced by `run_id`. Some model flavors (e.g., `fastai`, `mleap`, `diviner`, `promptflow`) were deprecated or removed. The `log_model` API can now be called directly without `mlflow.start_run()` context.
Fix: Consult the MLflow 3.x Migration Guide for a comprehensive list of changes. Update code to use `run_id` instead of `run_uuid` and adjust model logging patterns as necessary.

Install

pip install mlflow-skinny Base installation
pip install mlflow-skinny scikit-learn numpy pandas With scikit-learn autologging support
pip install mlflow-skinny sqlalchemy alembic sqlparse With SQLite tracking backend support

Imports

mlflow
```
import mlflow
```
mlflow.sklearn
```
import mlflow.sklearn
```
MlflowClient
```
from mlflow import MlflowClient
```

Quickstart

This quickstart demonstrates how to use `mlflow-skinny` for experiment tracking with scikit-learn's autologging feature. It logs a `RandomForestRegressor` model, its parameters, and metrics automatically. Note that for this example to run, `scikit-learn`, `numpy`, and `pandas` must be installed alongside `mlflow-skinny`. It explicitly sets a file-based tracking URI to avoid issues with MLflow 3.x's default SQLite backend, which `mlflow-skinny` doesn't support out-of-the-box.

import os
import mlflow
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_diabetes
from sklearn.ensemble import RandomForestRegressor

# Ensure required dependencies are installed for this example (sklearn, numpy, pandas)
# pip install mlflow-skinny scikit-learn numpy pandas

# Set a tracking URI. With mlflow-skinny, a local file-based store (mlruns/) is often preferred
# or ensure `sqlalchemy` is installed for 'sqlite:///mlflow.db' default in MLflow 3.x.
# We use a local directory explicitly to avoid the default SQLite dependency issue.
mlflow.set_tracking_uri("file:///tmp/mlruns_quickstart")

# Enable MLflow's automatic experiment tracking for scikit-learn
# This will log parameters, metrics, and the model automatically
mlflow.sklearn.autolog()

# Load the training dataset
db = load_diabetes()
X_train, X_test, y_train, y_test = train_test_split(db.data, db.target)

# Train a RandomForestRegressor model
# MLflow triggers logging automatically upon model fitting due to autologging
with mlflow.start_run():
    rf = RandomForestRegressor(n_estimators=100, max_depth=6, max_features=3, random_state=42)
    rf.fit(X_train, y_train)

    # You can also manually log additional metrics or parameters if needed
    # mlflow.log_metric("example_custom_metric", 0.95)

print(f"MLflow Run completed. View runs at: {mlflow.get_tracking_uri()}")

view raw JSON →