Model deployment

How this question breaks down

1. 1RegistryCore18 minRecord every trained model version with what made it, how it scored and who approved it, then let serving load whatever a named alias points to, so promoting and rolling back become a one-line metadata change.
2. 2CanaryCore20 minRun the new model on copied live traffic while the old one still answers, compare their outputs, then let it decide for a small slice of real users, and judge it on signals that arrive in hours because the true labels take weeks.
3. 3RollbackCore17 minMoving the alias back to the last good version takes one command, but it only restores behaviour if the features, threshold and indexes that version expects come back with it, and it never undoes the decisions and training data the bad version already produced.

Work through the parts in order and you’ve covered the full design. Each part is one topic.