Avoid sub-optimal models in production.
Similar to deployment, rolling back models can be a tedious process. Instead of manually performing this task, it is recommended to define an automatic process for it.
Automatic rollbacks should be implemented in any production-level ML application.
If, due to changes in the input data or undetected skew, a deployed model performs sub-optimal, it should be rolled back to an earlier, better performing version.
Designing a process for automatic roll-back minimizes the time a deployed model with sub-optimal performance is kept in production.