Skip to content
AI & analytics

MLOps

The engineering practices for taking machine learning models from training to deployment and monitoring them in production, including on fleets of edge devices.

MLOps applies DevOps ideas to models: versioned datasets and code, reproducible training pipelines, automated testing, controlled rollout and monitoring. For IoT the loop is longer than for a web service because the model ships to thousands of devices with different hardware, connectivity and firmware. A model update is effectively a firmware update and goes through the same OTA update and rollback machinery.

Monitoring is the hard part at the edge: devices may not send raw inputs back, so teams rely on summary statistics and sampled data to catch model drift. A sound MLOps setup records which model version runs on which device, measures accuracy against labels collected in the field, and can retrain and redeploy without a site visit. It bridges model training teams and the fleet management teams who operate the devices.

Part of the IoT Glossary. Spotted an error or a missing term? Tell us.