Qdrant builds Universal I/O to run vector search mmap code on object storage
qdrant_engine · x · 2026-10-27
Qdrant engineers explain how they moved their vector database onto object storage without rewriting its data structures:
- Motivation: a cold mmap read takes 100µs on local SSD but 10–100ms over the network; vectors, HNSW graphs, and payload indexes all relied on mmap.
- Approach: instead of maintaining two versions of every data structure, they built a shared abstraction called Universal I/O — callers queue reads and collect results, and each backend decides how to execute them.
- Remote caching: since object-storage files are immutable or append-only, fetched blocks never go stale; a local mmap mirror with a bitmap tracking 16KiB blocks handles HIT/MISS/COVERED paths.
- Diagrams show an async read pipeline on Tokio workers where reads are scheduled immediately and results collected in finish order.
Related event: Qdrant's Universal I/O Lets Vector DBs Run Directly on Object Storage(3 posts)→
More from coding & agent
- huggingface_hub 2.3.0 cuts 1M-file listings to 20s, adds live Bucket following — lhoestq · 2026-10-27
- Nous Research ships Advanced Delegation plugin letting Hermes Agent route subagents to different models — NousResearch · 2026-10-27
- A dentist with a 3D printer discovered OpenAI Codex — and never looked back — gabrielchua · 2026-10-27
- Stanford's PaperFormalizer turns research papers into Lean-verified versions with author in the loop — StanfordUniversity · 2026-10-27
- Hugging Face TRL makes score centering + masked importance sampling default in async GRPO — _lewtun · 2026-10-27
- a16z: agents follow tech's oldest law — chaos first, then rein it in — a16z Newsletter · 2026-10-27