Promouvoir automatiquement uniquement en cas de dépassement de la référence

A freshly retrained model can be worse than the one in production. Never promote a new model blindly just because it is newer.

The live model is the champion; the new candidate is the challenger. Promotion is a contest the challenger must win on the metrics.

Your baseline is the current champion score on a fixed evaluation set. The challenger has to clearly beat that number to earn promotion.

Score both models on the identical, held-out test set. A fair comparison only counts when nothing else changed between them.

Wrap the decision in a check that returns true only if the challenger wins. This gate compares the two scores directly.

Tiny gains can be noise, so require a real margin. Demanding at least a 1% lift avoids churning the production model for nothing.

Compare on the metric that matters for your task, like F1 or AUC, not just raw accuracy. The wrong metric promotes the wrong model.

Put the gate right after evaluation in your DAG. If the challenger loses, the pipeline stops and the champion stays live.

Winning means moving the model stage to Production in the registry. This transition is the one action that changes what users get served.

Record both scores and the outcome every run. That audit trail explains later why a model was or was not promoted.

A challenger can look amazing because the test set leaked or shrank. Sanity-check the evaluation before trusting any promotion.

When should the new challenger model be auto-promoted?

You gated promotion on results: compare challenger to champion on the same data, require a margin, then transition the winner in the registry. ✅