Scaling a Weaviate vector index 26× with one Amazon EBS gp3 command

This article demonstrates how Weaviate's HFresh disk-backed vector index dramatically improves query performance by leveraging Amazon EBS gp3's provisioning capabilities.

• HFresh keeps a small centroid index in memory and reads vectors from disk, making storage I/O the performance bottleneck instead of RAM
• Raising a single gp3 volume from 3,000 to 80,000 IOPS increased throughput from 9.8 to 260 QPS on a 10-million-vector index with 8 GiB memory cap
• P99 latency improved 28-fold from 5.7 seconds to 204 ms with the same hardware and dataset
• Setting PERSISTENCE_LSM_ACCESS_STRATEGY=pread is critical to reach high IOPS; default mmap strategy caps performance at 39,000 IOPS
• Cost per 100 QPS improved 20-fold with only 34% monthly cost increase for the IOPS provisioning
• Volume modifications are live with no restart required; IOPS changes take minutes to hours depending on volume size

Disk-backed vector indexes replace fixed memory requirements with tunable storage performance, enabling cost-effective scaling through EBS provisioning rather than larger instances.