Booking.com's 100M-embedding test shows purpose-built vector DBs outperform general-purpose search engines
CShorten30 · x · 2026-10-11
- Booking.com ran its own benchmark with 100M embeddings, filtered queries, and concurrent writes, concluding purpose-built vector databases behave very differently from general-purpose search engines with vector features bolted on.
- In Weaviate, index and storage engine are designed together: HNSW is an in-memory graph persisted via an append-only log; objects and filters live in a custom LSM store, with filters surfaced as bitmaps that steer HNSW traversal (ACORN), keeping writes cheap appends.
- HFresh goes further by placing vector clusters directly in the LSM store so the index rebalances itself in the background without full rebuilds.
- Other systems inherit storage built for other workloads: page-based databases turn every update into new row versions plus WAL and VACUUM overhead.
More from Infra
- Sentry CEO: startups gluing OTel onto ClickHouse will be the next o11y failures — zeeg · 2026-10-12
- AI networking DSP moats questioned: real differentiation lies with TSMC microrings and CPO — jwt0625 · 2026-10-12
- B200 benchmark: NVFP4 decodes up to ~8% faster than MXFP4 in low-batch LLM inference — StasBekman · 2026-10-11
- Fractile CEO explains why frontier labs don't all build their own chips — No Priors · 2026-10-11
- Defending Musk's 'no custom equipment': how standardization inertia slows chip fab innovation — jwt0625 · 2026-10-11
- JPMorgan, Morgan Stanley struggle 6 months to offload $38B Oracle data center debt — SumitGup · 2026-10-11