How does the MLOps lifecycle differ from an ordinary software delivery lifecycle?

Traditional software delivery mainly manages application code and releases. MLOps builds on those practices, but it also manages training data, experiments, model versions, and prediction quality over time.

A simple example

Consider a shopping app with two features.

Shopping cart: A developer changes the discount calculation, tests the code, and releases the update.

Product recommendations: The app uses a trained model to suggest products. During a festival season, shopping habits change. The code may still run perfectly, but the recommendations may become less helpful. The team might need new data, a newly evaluated model, and a safe release.

What makes the MLOps lifecycle different?

• Data: Teams must check and manage the data used to train and run the model.
• Experiments: Training may involve testing different approaches and comparing results.
• Testing: In addition to software tests, teams check data quality and model performance.
• Deployment: A release may include the prediction service and an automated training pipeline.
• Monitoring: Teams watch both system health and the usefulness of predictions.
• Updates: A model may need retraining even when the application code hasn't changed.

Two simplified lifecycles

Software: Write code → Test → Build → Deploy → Monitor → Update.

MLOps: Prepare and validate data → Train → Evaluate → Version → Deploy → Monitor → Retrain when needed.

The MLOps workflow still includes software engineering activities; these steps highlight the additional responsibilities.

Does MLOps replace DevOps?

No. MLOps extends DevOps practices such as Continuous Integration (CI) and Continuous Delivery (CD). It may also use Continuous Training (CT) to train and evaluate candidate models when new data arrives or performance declines. A newly trained model should pass quality checks before it replaces the deployed version.

In short: Software delivery keeps applications reliable as code changes. MLOps also helps keep model predictions useful as the world and its data change.