GENREL05-BP02 Replicate embedding data across all regions of availability
Inference to a foundation model may be available over a local availability region, or could be a large region of availability. Make sure your data is available across all regions of availability to adequately service inference requests.
Desired outcome: When implemented, this best practice improves the reliability of your generative AI workload by validating that models have access to the appropriate data to service inference requests across an entire Region of availability.
Benefits of establishing this best practice: Scale horizontally to increase aggregate workload availability - Data replication across a region of availability enables horizontal scaling of the data access infrastructure and supports consistent serving of inference requests.
Level of risk exposed if this best practice is not established: Medium
Implementation guidance
Validate that data sources for your generative AI workloads are replicated and made available across all of the designated regions of availability. Configure data replication pipelines which replicate data creation, updates, and deletions across data sources, providing a consistent experience for users. Customers should leverage a durable storage layer which is available across all desired regions.
Amazon S3 is a common choice since it is a durable, scalable, and reliable storage layer which integrates simply with several data analytics and vector storage solutions. A modern data architecture backed by Amazon S3 is a recommended choice for multi-region data availability at scale. Consider using Amazon S3 or a similar storage layer. Develop data pipelines to distribute data across regions. Amazon S3's bucket replication capability is a managed version of a data pipeline which replicates data across regions.
Alternatively, data pipelines can be developed and orchestrated manually using Amazon Glue. These data pipelines should run frequently enough to satisfy data availability requirements across regions. Once replicated, verify that the data is processed by a replicated vector storage layer.
Implementation steps
-
Create two OpenSearch clusters across two regions, where one is a leader and one is a follower.
-
Create a request for an outbound connection from the follower to the leader.
-
Accept the inbound request from the leader.
-
Modify the leader security configuration to facilitated cluster replication.
-
Create an index for replication on the leader cluster.
-
Run cluster replication from the follower cluster.
-
Index documents on the leader cluster.
-
Test document replication on the follower cluster.
Resources
Related practices:
Related guides, videos, and documentation:
Related examples:
