v.25.4New Feature
Add an in-memory cache for deserialized vector
Added an in-memory cache for deserialized vector similarity indexes. This should make repeated approximate nearest neighbor (ANN) search queries faster. The size of the new cache is controlled by server settingsvector_similarity_index_cache_sizeandvector_similarity_index_cache_max_entries. This feature supersedes the skipping index cache feature of earlier releases. #77905 (Shankar Iyer).
Why it matters
This feature improves the performance of ANN search queries by caching deserialized vector similarity indexes in memory, reducing the overhead of repeated index deserialization. It replaces the previous skipping index cache mechanism with a more efficient caching solution.How to use it
Users can control the size of the cache using the server settingsvector_similarity_index_cache_size and vector_similarity_index_cache_max_entries. Adjust these settings to optimize memory usage and cache effectiveness according to your workload.