Chap 8. Productionization/Maintenance/Adoption

A model that proved itself offline does not mean that it works well online. It needs to be battle-tested with real data.

In practice, it now switches from offline training to online training.

On the performance side, it needs to keep tracking and monitoring the performance is consistent with the offline results, and worked/showed benefits as it is expected.

In terms of maintainness, issues such as dependency's version updates compatibility, bug fixes / new features, new releases, code refactoring if there are traffic changes, scalability need to be taken into consideration.