Model maintenance
Periodical model retraining is necessary to prevent model staleness.
As the input data distribution changes, the performance of the model will weaken over time. Depending on the frequency of the data variations, you can update the model manually or schedule an automatic model update.
While the existing endpoint is being redeployed, the endpoint will not have a downtime if it is still active. After the updated endpoint is created, it replaces the existing endpoint. You still get predictions based on the old endpoint before it is replaced.