Feature request – support pgvectorscale
It would be awesome if you could look into adding pgvectorscale as a supported extension. Its StreamingDiskANN index is designed to stay mostly on disk and uses binary compression, so in theory it should need far less RAM than pgvector’s HNSW and make scale-to-zero cold starts and smaller compute tiers a bit friendlier for sizeable vector workloads.
I totally get that this is a young extension and that the local-NVMe benchmark numbers out there likely won’t map 1-to-1 onto Neon’s remote-storage setup, but seems like it could improve pgvector workloads in both performance and RAM-related cost, which matters a lot on Neon.
I'm sure there could be complexities/factors that I am missing that y'all have considered though, so lmk if so.
I totally get that this is a young extension and that the local-NVMe benchmark numbers out there likely won’t map 1-to-1 onto Neon’s remote-storage setup, but seems like it could improve pgvector workloads in both performance and RAM-related cost, which matters a lot on Neon.
I'm sure there could be complexities/factors that I am missing that y'all have considered though, so lmk if so.
