CoordiNode

Appearance

Benchmarks ​

CoordiNode benchmarks use industry-standard suites (ann-benchmarks, LDBC SNB, YCSB, TSBS, Search Benchmark Game); every result is JSON-recorded with hardware fingerprint + Git SHA so the timeline is reproducible end-to-end.

Live data

The charts below are generated from JSON files at bench-results/ on every commit. CoordiNode results are produced automatically by CI on a dedicated bench host on every push to main, accumulating a full timeline. Competitor baselines (hnswlib, Faiss, MongoDB, etc.) are pinned to a specific version, run manually on the bench host, and replaced when we re-bench against a newer build: no competitor history is kept in-repo, only the current head-to-head. Each result names the host it ran on (see below); compare numbers only between results from the same host.

How to read these charts ​

  • Pareto frontier — higher and to the right is better. Recall (or correctness proxy) on the x-axis, throughput on the y-axis (log scale). Each curve is one engine.
  • Recall / accuracy timeline — peak score across CoordiNode commits. The CoordiNode line moves on every push to main; competitor lines stay flat between manual re-baselines.
  • Throughput @ target quality — the canonical "useful operating point" cell (e.g. QPS @ recall ≥ 0.95 for ANN). The taller the bar, the better the engine handles that point.

Hardware fingerprint ​

Every bench JSON records the host it ran on: CPU model, cores and threads, RAM, OS and architecture. Results up to October 2026 were measured on an Intel i9-9900K desktop (8 cores / 16 threads, 64 GB RAM); later ones on the Linux bench host their JSON names. The hosts are intentionally modest hardware: when CoordiNode reports a number on them, the equivalent server-class run is going to be faster, not slower. The point is comparability: a head-to-head compares engines measured on the same host, and Linux and Windows results are never compared with each other.

Why every modality is on one page ​

CoordiNode is one engine, one transaction and one query language, and a real workload mixes modalities: a graph traversal feeding a vector search, filtered by a predicate over indexed document properties. Benchmarking each modality in isolation is the first step, and the tabs below say plainly which ones have published numbers today. Cross-modality workloads land as the per-modality harnesses mature. The bench harness (crates/coordinode-bench) is modality-agnostic: adding a new dataset is a JSON and chart-spec addition.