MLOps: Running Machine Learning in Production

What is MLOps

MLOps (machine learning operations) is the set of practices that keep a machine learning model working in production: how it is deployed, how it is monitored, how it is retrained and how you know which version made each decision. It borrows ideas from DevOps but adds something of its own: in a model, not only the code changes, the data it works with changes too.

Why ML models degrade in production

A model learns from the data that existed when it was trained. Over time, reality changes: new products come in, customers change, a camera moves, the lighting changes or a process is modified. The software stays on and keeps answering, but its predictions become less and less reliable. This is called data drift. Without monitoring, nobody notices until the errors reach the business.

How MLOps works

An MLOps cycle usually includes:

• Versioning: of data, labels, code and model, so you can reproduce a result and explain a past decision.
• Controlled deployment: first in a test environment or in parallel, without acting, and then gradually.
• Monitoring: availability, latency, distribution of predictions, uncertain cases and comparison with a human reference.
• Retraining: with new data when metrics drop, and validation before replacing the previous model.
• Rollback: being able to restore the previous version if the new one fails.

An example

A demand forecasting model works well for months. A new campaign changes buying patterns and the error starts to rise. The monitoring dashboard catches it, the team retrains with recent data, compares the new version with the current one over the same period and releases it only if it improves. If something goes wrong, it goes back to the previous one.

MLOps best practices

• Measure the model with business metrics, not only technical ones.
• Store which model and data version produced each result.
• Test in shadow mode before letting the model act on its own.
• Keep human review on a sample and on uncertain cases.
• Monitor the model like any other piece of production equipment.

How we apply it at DPI

MLOps is part of our machine learning service, where we train prediction, classification and optimization models on each company's data and validate them with business metrics. In industrial quality control we follow the same principles: version data, labels and model, run in shadow mode before automating, and monitor the system, because if conditions change the model can stop being reliable even though the software is still running.