A growing company has a business-critical key performance indicator (KPI) for the uptime of a machine learning (ML) recommendation system. The company is using Amazon SageMaker hosting services to dev

Deploying across multiple AWS Regions provides disaster recovery, but it is a higher-effort and more complex solution than multi-AZ deployment within a single region, typically for RTOs longer than 5 minutes, and is overkill for the specified RTO.

Using SageMaker auto scaling primarily handles varying workloads by adjusting instance count, but it does not inherently provide high availability against an entire Availability Zone outage if all instances are confined to a single AZ.

Deploying multiple instances for each production endpoint within a VPC that spans at least two subnets in different Availability Zones (AZs) is the standard and most efficient way to achieve high availability for SageMaker hosting services. This configuration automatically distributes traffic and handles failover if an AZ becomes unavailable, meeting the 5-minute Recovery Time Objective (RTO) with minimal effort.

Frequently generating backups and deploying them in a different AZ or Region implies a manual or semi-automated recovery process, which is unlikely to achieve a 5-minute RTO for an active endpoint and represents a disaster recovery strategy rather than continuous high availability.