← Back to the chapter · Previous detail

Technical detail · Production ML Monitoring

Gate feedback and retraining

Operator corrections, confirmed events, user reports, maintenance outcomes, and delayed labels help the next model only when they are traceable, privacy-reviewed, representative enough for the decision, and separated from unverified guesses. Avoid training on the model’s own unverified predictions, which can amplify previous mistakes.

Retraining changes the release boundary. Carry the training data window, label source, privacy review, feature code, threshold, held-out segments, canary results, and rollback rule into the new release record. Time-align delayed maintenance labels, operator overrides, and incident reviews with the original predictions. Separate evidence by site, device type, operating mode, and user impact so the owner can assess the tradeoff.

Record the next retest trigger after acting. A future model, feature, firmware, population, threshold, privacy, or device change should reopen the monitoring record. The evidence then supports a choice among rollback, threshold repair, data-quality repair, and retraining.

← Back to the chapter · Previous detail