Hi all, I would like to share an extension that I am building called pgvectorfast.
We love pgvector for its simplicity and ease of use, but one complaint we often see is that performance at larger scales can lag behind dedicated vector databases. So I wanted to see how much of that gap could be closed without leaving Postgres. I found that PostgreSQL’s shared buffer pool was the main bottleneck in these vector search workloads. So I built pgvectorfast, a drop-in enhancement to pgvector with a fast buffer pool.
I am currently seeing ~6X speedup across in-memory and out-of-memory vector search workloads using HNSW index with the compatibility of pgvector, and it often outperforms dedicated vector databases.
Hi all, I would like to share an extension that I am building called pgvectorfast.
We love pgvector for its simplicity and ease of use, but one complaint we often see is that performance at larger scales can lag behind dedicated vector databases. So I wanted to see how much of that gap could be closed without leaving Postgres. I found that PostgreSQL’s shared buffer pool was the main bottleneck in these vector search workloads. So I built pgvectorfast, a drop-in enhancement to pgvector with a fast buffer pool. I am currently seeing ~6X speedup across in-memory and out-of-memory vector search workloads using HNSW index with the compatibility of pgvector, and it often outperforms dedicated vector databases.
Repo: https://github.com/mitdbg/pgvectorfast PGXN: https://pgxn.org/dist/pgvectorfast
It’s still early and not battle-tested, and I would greatly appreciate feedback and suggestions for improvement!