MLflow

Warnings

• breaking MLflow 3.x introduced significant breaking changes, including the complete removal of MLflow Recipes. Many model flavors (fastai, mleap, diviner) are no longer supported directly. The 'routes' and 'route_type' config keys for AI Gateway were removed. The deployment server and `start-server` CLI command have been removed, replaced by `mlflow models serve` or containerized deployments. [1, 3]
  Fix: Review the MLflow 3 Migration Guide. For Recipes, migrate to standard MLflow tracking/model registry or MLflow Projects. For unsupported flavors, use `mlflow.pyfunc` with a custom wrapper or `mlflow.onnx`/`mlflow.pytorch`. Use the new AI Gateway configuration format and `mlflow models serve` for deployments. [1]
• breaking In MLflow 3.x, the `run_uuid` attribute on `RunInfo` objects has been removed and replaced by `run_id`. Additionally, several Git-related run tags (`mlflow.gitBranchName`, `mlflow.gitRepoURL`) were removed. [1]
  Fix: Update code to use `run_id` instead of `run_uuid`. Remove reliance on the deprecated Git tags or implement custom tagging for similar information. [1]
• breaking The Artifacts tab in the MLflow UI for runs no longer displays model artifacts in MLflow 3.x. Model artifacts are now accessed through a dedicated 'Logged Models' page. [1]
  Fix: Navigate to the 'Logged Models' page in the MLflow UI to view model-specific information and artifacts. Update any automation or user guides that rely on the old UI structure. [1]
• deprecated Parameters like `example_no_conversion` and `code_path` have been removed from model logging/saving APIs. `requirements_file` for PyTorch flavor is removed, and `inference_config` from Transformers flavor is also removed. [1]
  Fix: For `code_path`, use the default code directory structure. For PyTorch `requirements_file`, use `pip_requirements` or `extra_pip_requirements`. For Transformers `inference_config`, set the configuration before logging the model. [1]
• gotcha When upgrading a self-hosted MLflow server, it is crucial to stop the server, upgrade the package, run database migrations using `mlflow db upgrade <backend-store-url>`, and then restart the server. MLflow does not natively support live upgrades, and schema migrations can be slow and non-transactional. Always back up your database before migration. [16]
  Fix: Follow the official upgrade procedure carefully. Plan for downtime or implement a rolling upgrade strategy with a load balancer for high availability. Back up your database before any migration. [16]
• gotcha MLflow clients and servers work best when they are on the same version. While a newer server is generally backward compatible with older clients for basic logging, using newer client features (e.g., MLflow Tracing) with an older server might lead to missing endpoints or unexpected behavior. [16]
  Fix: Strive to keep your MLflow client SDK and server versions aligned. If using new client features, ensure your server is also updated to a compatible version. [16]

Install

• pip install mlflow Install MLflow

Imports

• mlflow

  import mlflow

  Main MLflow module for tracking, models, projects.
• MlflowClient

  from mlflow.tracking import MlflowClient

  For lower-level interaction with the MLflow Tracking Server. [21]
• infer_signature

  from mlflow.models import infer_signature

  Utility to infer model signatures automatically. [2]

Quickstart

This quickstart demonstrates how to log parameters, metrics, and a scikit-learn model using the MLflow fluent API. It sets up an experiment, trains a logistic regression model on the Iris dataset, logs its hyperparameters and accuracy, infers the model signature, and registers the model in the MLflow Model Registry. To view the results, start the MLflow UI by running `mlflow ui` in your terminal and navigating to `http://localhost:5000` (or `http://127.0.0.1:5000`). [2, 5, 22]

import mlflow
import pandas as pd
from sklearn import datasets
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from mlflow.models import infer_signature

# Set MLflow tracking URI (optional, defaults to local ./mlruns)
# For a local server, run 'mlflow ui' in your terminal and point to http://127.0.0.1:5000
# os.environ['MLFLOW_TRACKING_URI'] = os.environ.get('MLFLOW_TRACKING_URI', 'http://127.0.0.1:5000')

mlflow.set_experiment("MLflow_Quickstart_Experiment")

# Load the Iris dataset
X, y = datasets.load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Define model hyperparameters
params = {"solver": "lbfgs", "max_iter": 1000, "multi_class": "auto", "random_state": 8888}

with mlflow.start_run():
    # Log hyperparameters
    mlflow.log_params(params)

    # Train the model
    lr = LogisticRegression(**params)
    lr.fit(X_train, y_train)

    # Make predictions and calculate metrics
    y_pred = lr.predict(X_test)
    accuracy = accuracy_score(y_test, y_pred)
    mlflow.log_metric("accuracy", accuracy)

    # Infer model signature
    predictions = lr.predict(X_train) # Use training data for signature inference
    signature = infer_signature(X_train, predictions)

    # Log the model
    mlflow.sklearn.log_model(
        sk_model=lr,
        artifact_path="logistic_regression_model",
        signature=signature,
        registered_model_name="IrisLogisticRegression"
    )

    print(f"Logged model with accuracy: {accuracy}")
    print(f"View runs in MLflow UI: run 'mlflow ui' in your terminal and navigate to http://127.0.0.1:5000")