The Case Against pgvector | Alex Jacobs
What happens when you try to run pgvector in production and discover all the things the blog posts conveniently forgot to mention
Vector Database
Summary
VectorChord achieves 100M vector indexing in 20 minutes with 12 GB RAM on PostgreSQL, a 120x speedup and 7x memory reduction over pgvector, through hierarchical K-means, reduced contention, and parallelized compaction.
Highlights
- Indexes 100M 768-dim vectors in 20 min with 12 GB RAM vs pgvector's 40 hours and 200 GB on same hardware
- Hierarchical K-means with dimensionality reduction cuts initialization from 30 min (GPU) to 8 min (CPU) and 135 GB to 23 GB
- Reducing lock contention during insertion phase drops build time from 420 minutes to 9 minutes
- Parallelized compaction reduces final phase from 8 minutes to 1 minute
- Enables deployment on cheaper instances like Amazon i7i.4xlarge at $1087/mo without requiring a GPU
auto-generated
Context
Audience
Database engineers and ML infrastructure teams deploying large-scale vector search on PostgreSQL
DomainVector Databases and PostgreSQL Performance
FormatTechnical blog post
AccessFree online
Source Overview
VectorChord indexes 100 million 768-dimensional vectors in 20 minutes using 12 GB memory on a 16 vCPU machine, compared to pgvector's 40 hours and 200 GB.
Optimizations target three phases: hierarchical K-means plus dimensionality reduction cut initialization from 30 to 8 minutes and memory from 135 GB to 23 GB; reduced contention speeds insertion from 420 to 9 minutes; parallelized compaction drops from 8 to 1 minute. These enable indexing on cheaper instances like Amazon i7i.4xlarge without GPUs.
Topics
Related
pgvectorvchordrq indexHierarchical K-means clusteringInverted file indexAmazon EC2 i7i instances
