Benchmarking Billion-Scale Vector Indexes
Couchbase, Thursday, November 5th, 2026
A Couchbase engineer explains how the team benchmarked vector indexes at billion scale: metrics, datasets, ground truth and partitioning.
Couchbase describes how it benchmarked vector search at the 1-billion-vector scale, focusing on P99 latency, Recall@K, QPS at a fixed recall target, index build time, memory (RSS) and CPU.
It used three datasets: SIFT-1B for scale, LAION-100M for real float32 embeddings, and a self-built 10M-vector, 4096-dimension BookCorpus set embedded with Llama 2.
To get ground truth without a huge machine, it split the data into partitions and computed exact neighbors per partition.
Keying queries to partitions also enabled partition elimination, cutting work per query by 90% on a 10-partition layout and tightening tail latency.