S3 Vectors best practices
Amazon S3 Vectors delivers purpose-built, cost-optimized vector storage for use by AI-enabled applications and semantic search of your content stored in Amazon S3. Designed to provide S3 level elasticity and durability for storing vector datasets with sub-second query performance for cold queries and as low as 100 milliseconds for warm queries, S3 Vectors is ideal for applications that need to build and grow vector indexes. With S3 Vectors you can use a dedicated set of API operations to store, access, and perform similarity queries on vector data without provisioning any infrastructure. For more information, see Working with S3 Vectors and vector buckets.
To ensure the maximum benefit from S3 Vectors, we recommend that you perform the following best practices.
- Inserting and deleting vectors
-
Your application can achieve up to one thousand PutVectors or DeleteVectors requests per second per vector index, or can insert or delete up to two thousand five hundred vectors per second per vector index — whichever limit is reached first. If you exceed the request rates, you might receive a
429 TooManyRequestsExceptionerror.To optimize costs, we recommend inserting and deleting vectors in large batches, up to the maximum batch size of 500 vectors per API request. If your workload requires smaller batches, you can send concurrent requests up to the 1,000 requests per second limit. To reach maximum throughput of 2,500 vectors per second, you can send 5 batches per second with 500 vectors each, or 1,000 batches per second with an average of 2.5 vectors each.
- Accessing and querying vectors in an S3 vector index
-
Your application can achieve hundreds of QueryVectors, GetVectors, or ListVectors requests per second per S3 vector index. If you exceed the request rates, you might receive a
429 TooManyRequestsExceptionerror. We recommend you use a retry mechanism and configure your application to send fewer requests. - Scaling across vector indexes
-
To improve query performance per vector index, consider configuring your application to divide vectors across multiple vector indexes when possible. For example, if you have multi-tenant workloads and your application queries each tenant independently, consider storing each tenant's vectors in a separate vector index. For more information, see Vector indexes.
- Implementing multi-tenancy with separate vector indexes
-
You can achieve multi-tenancy by organizing your vector data using a single vector index for each tenant. You can use IAM and bucket policies to restrict each tenant's access to only their designated vector index. This approach helps maintain data isolation and simplifies management by removing the need to create separate buckets for each tenant. For more information, see Identity and Access management in S3 Vectors.
- Configuring non-filterable metadata fields for vector indexes
-
When creating a vector index, configure metadata fields that don't require filtering as non-filterable metadata keys. For example, store text chunks for vector embeddings as non-filterable metadata fields when you need them only for reference. For more information, see Non-filterable metadata.
