Feature sharing across teams in MLOps - Commands & Configuration

Introduction

When multiple teams work on machine learning projects, they often need to reuse the same data features. Feature sharing helps teams avoid repeating work and keeps features consistent across projects.

When your data science team wants to reuse customer age and location features in different ML models.

When a new team joins and needs access to existing features without rebuilding them.

When you want to keep feature definitions consistent to avoid errors in model training.

When you want to track and update features centrally so all teams get the latest version.

When you want to speed up model development by sharing tested and validated features.

Commands

Start the MLflow tracking server to store and share feature metadata and artifacts centrally.

Terminal

mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./mlruns --host 0.0.0.0 --port 5000

Expected Output

2024/06/01 12:00:00 INFO mlflow.server: Starting MLflow server...

2024/06/01 12:00:00 INFO mlflow.server: Listening at http://0.0.0.0:5000

→

--backend-store-uri - Specifies where to store metadata

→

--default-artifact-root - Specifies where to store feature files

→

--host - Makes server accessible on all network interfaces

Log a feature named 'customer_age' with a sample value to the MLflow tracking server for sharing.

Terminal

mlflow run . -P feature_name=customer_age -P feature_value=35

Expected Output

2024/06/01 12:01:00 INFO mlflow.projects: Running command 'python feature_log.py --feature_name customer_age --feature_value 35'

Feature 'customer_age' logged with value 35

Run ID: 1234567890abcdef

Download the logged feature artifacts from MLflow to use in another team or project.

Terminal

mlflow artifacts download -r 1234567890abcdef -d ./downloaded_features

Expected Output

Successfully downloaded artifacts to ./downloaded_features

→

-r - Specifies the run ID to download artifacts from

→

-d - Specifies the local directory to save artifacts

Key Concept

If you remember nothing else from this pattern, remember: centralizing feature storage lets all teams reuse and update features easily without duplication.

Code Example

MLOps

import mlflow
import argparse

def log_feature(feature_name: str, feature_value: int):
    with mlflow.start_run() as run:
        mlflow.log_param("feature_name", feature_name)
        mlflow.log_metric("feature_value", feature_value)
        print(f"Feature '{feature_name}' logged with value {feature_value}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--feature_name', type=str, required=True)
    parser.add_argument('--feature_value', type=int, required=True)
    args = parser.parse_args()
    log_feature(args.feature_name, args.feature_value)

Common Mistakes

Not running the MLflow server before logging features

Features cannot be stored or shared without the server running, causing errors.

Always start the MLflow tracking server before logging or retrieving features.

Logging features with inconsistent names or formats

This causes confusion and errors when teams try to reuse features.

Agree on a naming convention and data format for features before sharing.

Downloading artifacts without specifying the correct run ID

You may get wrong or no feature data, breaking your model pipeline.

Always use the exact run ID from the feature logging step to download artifacts.

Summary

Start the MLflow server to store and share features centrally.

Log features with clear names and values using MLflow commands.

Download shared features by specifying the correct run ID for reuse.