Production ML & MLOps

ML System Lifecycle

A production ML system is not just a trained model. It is a loop around data collection, labeling, feature generation, training, evaluation, deployment, monitoring, retraining, rollback, and human operations.
  • Data is the long-term asset
  • Training is only one stage
  • Evaluation gates deployment
  • Production feedback changes the dataset
  • Retraining is a controlled release process
  • Ownership must cover the whole loop
ML lifecycle loop
The loop is continuous because data, users, and product requirements keep moving.
Lifecycle artifacts
ArtifactWhy it matters
Dataset versionReproduce what the model learned from
Feature definitionsKeep training and serving aligned
Model artifactDeploy and roll back exact versions
Eval setCatch regressions before release
Experiment recordKnow what was tried and why
Monitoring dashboardDetect production behavior changes
RunbookRespond when quality or service degrades
Sources
  • Rules of Machine Learning — ML engineering lifecycle
  • Made With ML — MLOps lifecycle
  • Machine Learning Crash Course — Production ML systems