Benchmarks

Overview

Performance benchmarks overview

Latency and throughput of KiteDB's single-file engine, vector index, and language bindings, measured on one machine on October 3, 2026 with MVCC on (the default). Every table names the raw log in docs/benchmarks/results/ that its numbers come from, along with the dataset and durability settings of that run. docs/BENCHMARKS.md has the full notes and the earlier results.

Benchmark pages

Graph benchmarks: single-file engine latency from Rust and Python, MVCC on and off, sync modes, write scaling, open time and paging
Vector benchmarks: vector index insert, build, lookup, and search (Rust), and IVF against IVF-PQ
Cross-language benchmarks: Rust and Python side by side, plus the cost of the TypeScript fluent API

Test environment

Apple M5 Pro, 15 cores (5 Super + 10 Performance), 48 GB RAM
macOS 26.6 (Darwin 25.6.0), AC Power, energy mode high power
Rust 1.88.0
Node 24.21.0
Python 3.12.8
SQLite 3.47.1 (baseline)
Commit 5460a2c, release builds

Each configuration ran five times, interleaved with the others, with a fixed seed; each figure is the run with the median value, copied from the log.

Highlights

These are the numbers on the homepage, rounded. The graph and vector pages list them at full log precision.

Graph operations

Operation	Workload	p50	p95
Edge exists	random pair, 50k edges	42 ns	83 ns
Key lookup	random existing key	42 ns	208 ns
1-hop traversal	outgoing edges, random node	42 ns	84 ns
Commit 100 nodes	one transaction, WAL append	17.9 µs	46.8 µs
Commit 100 edges	one transaction, WAL append	16.1 µs	25.9 µs

Source: 2026-10-03-r2-single-file-raw-rust-mvcc-normal.txt (10k nodes, 50k edges, 3 edge types, 10 edge props, 1k vectors of 128 dims, sync=normal, MVCC on)

Vector index

Operation	p50	p95
Insert one vector (10k inserts)	833 ns	1.8 µs
build_index() (single run)	87.1 ms	n/a
Get vector (random)	166 ns	334 ns
Search (top-10 of 10k × 768d, IVF)	192 µs	270 µs

Source: 2026-10-03-r2-vector-bench-rust.txt (10k vectors, 768 dims, 1k iterations, k=10, nProbe=10, cosine, IVF with 100 clusters)

These runs use MVCC, the default since 0.3.0, and every commit is group-committed. The graph benchmarks compare the same run with MVCC off. The February 2026 results in docs/BENCHMARKS.md ran without MVCC on another machine (an Apple M4), so a difference from them mixes the code and the hardware. macOS's clock ticks every 41.67 ns, so 42 ns, the smallest nonzero time these benchmarks report, means one tick or less.

Bindings snapshot

p50 latency for the same graph workload from Rust and through the Python bindings.

Operation	Rust	Python
Key lookup (random existing key)	42 ns	125 ns
1-hop traversal (out, random node)	42 ns	250 ns
Edge exists (random pair)	42 ns	125 ns
Batch write (100 nodes)	17.9 µs	24.7 µs

Source: 2026-10-03-r2-single-file-raw-rust-mvcc-normal.txt, 2026-10-03-r2-single-file-raw-python-mvcc-normal.txt (10k nodes, 50k edges, 3 edge types, 10 edge props, 1k vectors of 128 dims, sync=normal, MVCC on)

Cross-language benchmarks has every operation, plus the TypeScript API.

Parallel write scaling

With MVCC, writer threads build their transactions in parallel, and commits that arrive together are written as one group. With syncMode=Normal, eight writers reach 12.69K/s with transactions of 200 nodes and 200 edges, 2.6x the rate of one writer, and 462.48K/s with one-node transactions, against 386.48K/s for one writer. The graph benchmarks have the 1, 4 and 8-writer measurements in both sync modes.

Running benchmarks

One script reruns every benchmark on this site and writes dated logs; run it on a quiet machine, from the repository root:

Shell

1ray-rs/scripts/bench-refresh.sh --list   # the matrix2ray-rs/scripts/bench-refresh.sh          # every run on this site3cd ray-docs && bun run bench:data        # this site's data from the new logs

The individual commands, from docs/BENCHMARKS.md. Rust core, graph operations:

Shell

1cd ray-rs2cargo run --release --example single_file_raw_bench --no-default-features -- \3  --nodes 10000 --edges 50000 --iterations 10000 \4  --wal-size 268435456 --no-auto-checkpoint --seed 42 --sync-mode normal --mvcc

Python bindings, graph operations:

Shell

1cd ray-rs/python/benchmarks2python3 benchmark_single_file_raw.py \3  --nodes 10000 --edges 50000 --iterations 10000 \4  --wal-size 268435456 --no-auto-checkpoint --seed 42 --sync-mode normal --mvcc

TypeScript, fluent vs low-level API:

Shell

1cd ray-rs2node --import @oxc-node/core/register benchmark/bench-fluent-vs-lowlevel.ts --mvcc --seed 42

Rust vector index:

Shell

1cd ray-rs2cargo run --release --example vector_bench --no-default-features -- \3  --vectors 10000 --dimensions 768 --iterations 1000 --k 10 --n-probe 10 --seed 42 --no-output