Exam-style
A compliance team holds 400 million embeddings covering a decade of archived case files. They run about thirty searches a month, and a wait of a second or two is fine. What is the cheapest fit?
Reveal the answer
A. S3 Vectors, storing the embeddings in a vector bucket and querying them through the s3vectors API
S3 Vectors is built for this shape: a large corpus, infrequent queries, and no need for single-digit-millisecond responses. A single vector index holds up to 2 billion vectors, so 400 million sits in one. There is no compute to provision. AWS charges USD$0.06 per GB-month of vector storage in us-east-1, plus per-request query fees, and infrequent queries return in under a second. Every alternative runs compute. A NextGen OpenSearch Serverless collection group does scale to zero after ten minutes idle, but the first search then waits 10 to 30 seconds while workers are restored, and bills OCU-hours once they are. A provisioned domain bills warm nodes and dedicated master nodes around the clock, whether or not anyone searches. An HNSW index over 400 million rows has to be built and held on a provisioned cluster that bills by the hour all month. MemoryDB vector search runs on a single shard and does not scale horizontally, so the whole index has to fit the memory of one hourly-billed node.
Q. You have a huge, rarely-queried vector archive and a wait of a second or two is acceptable. Cheapest fit?
A. S3 Vectors. Embeddings go into a vector bucket, a distinct bucket type queried through the s3vectors API, with no infrastructure to provision. One index holds up to 2 billion vectors. Infrequent queries return in under a second; AWS quotes latency as low as 100ms once queries become frequent. That floor rules the service out for an interactive assistant chasing 50ms.
Why? Match the store to the latency requirement. S3 Vectors bills storage and requests, not a cluster sitting idle between searches.