pgvector 0.4.0 performance

Summary

The post benchmarks pgvector 0.4.0 for approximate nearest-neighbor search and assesses whether PostgreSQL can support production-scale vector workloads. Using the ANN Benchmarks methodology, pre-warmed RAM, the dbpedia-entities-openai-1M dataset of 1 million 1,536-dimensional OpenAI embeddings, and Supabase ARM instances, it varies probes, lists, concurrency, and compute size. On an 8-core, 32 GB instance, pgvector reached 0.91 accuracy@10 at 380 QPS with 10 probes and 0.98 accuracy at 140 QPS with 40 probes; a 64-core, 256 GB server reached about 1,800 QPS at 0.91 accuracy. The results indicate predictable scaling, while higher lists can improve query speed at the cost of index-build time, and the roadmap identifies HNSW, product quantization, and parallel index builds as planned improvements.

Takeaways

1. 1

   On a 2XL instance, using 10 probes produced 0.91 accuracy@10 at 380 QPS, while 40 probes produced 0.98 accuracy at 140 QPS.

2. 2

   Increasing the lists constant can improve query speed for similar accuracy, but it increases index-build time and may require more maintenance_work_mem.

3. 3

   The benchmarks report about 1,800 QPS at 0.91 accuracy on a 64-core, 256 GB server, while the roadmap lists HNSW, product quantization, and parallel index builds as planned improvements.