Capacity Planning for Vector Search
Starting in DataStax Enterprise version 6.9, Vector Search was added as a way to do semantic associations among data as an extension to storage attached indexes (SAI).
From an operational standpoint, Vector Search behaves like any other database index. Writing data requires additional CPU resources to index it. When reading data, Vector Search (and SAI) will require additional work to consult the indexes, gather results, and send them to the application client. Therefore, capacity planning needs to consider this overhead when it comes to CPU usage, available memory, speed of storage, and per-node data density. Specifically, we recommend nodes that use Vector Search have at least 16 CPU cores, at least 64 GB memory, and fast storage (e.g. SSD or NVMe based).
|
Thoroughly test before deploying to production. DataStax highly recommends testing with tools such as NoSQLBench with your desired configuration. Be sure to test common administrative operations, such as bootstrap, repair, and failure, to make certain your hardware selections meet your business needs. See Testing Your Cluster Before Production. |
In addition to the database cluster, DataStax optionally provides a Data API that can abstract Vector Search data and indexes behind an easy to use JSON collection oriented interface. The Data API lives in a separate stateless service that is deployed as containers. The Data API can be scaled separately from the cluster and will depend on request throughput.
