INDEPENDENT. PURE RUST. BUILT FOR SPEED.

Your IP data.
At native speed.

Read, build and update MaxMind databases with typed, borrowed records. Small allocations. Consistent reads. Full control.

Start building ↗Explore benchmarks →

$cargo add libmaxminddb-rs

LOOKUP ENGINEIPv4 / IPv6

ZERO-COPY STRINGS↳ SAFE LIFETIMES

↳

Borrow the bytes

Decode directly into Rust structs. Borrow strings and byte slices from the database instead of copying them.

⇄

Update without downtime

Stage edits, rebuild, then publish with ArcSwap. Active readers keep a consistent generation until their work is done.

⌁

Build your own MMDB

Write IPv4 and IPv6 networks with longest-prefix matching, deduplication and explicit merge strategies.

MEASURED, NOT ASSUMED

Explore the tradeoffs.

Open benchmark lab →

IPv4 RANDOM / 1M QUERIES

— M ops/s

Single-thread lookup throughput from the committed comparison export.

READERS + WRITERS

6 libraries

Rust, C and Go. Compare throughput, tail latency, scaling and memory.

TRACEABLE RESULTS

19 charts / architecture

Every chart links to its source. Results include their recorded environment.

BENCHMARK LAB / CROSS-LIBRARY COMPARISON

Performance,
measured.

Follow the numbers. Compare the same workload across Rust, C and Go.

↓ Export data

◷ MEASURED SNAPSHOT

Measurement context ↗

ARCHITECTURE / MEASURED MACHINE

Each machine has its own measurements. Compare libraries within a run; hardware and toolchain differences also affect results across architectures.

WORKLOAD

RECORDED ENVIRONMENT

Loading measurements…

View source SVG ↗

By the numbers

EXPORTED PRECISION

Library	Measured version	Throughput

FROM THE REPOSITORY / RUNNABLE RUST

Less glue.
More Rust.

Complete examples with typed records and explicit ownership.

API reference ↗

Source ↗

🦀 Start here

Rust 1.98.1 or newer. Default features enable the reader, writer, derives and SIMD ASCII checks.

cargo add libmaxminddb-rs

🔄 Keep readers scoped

Borrowed records live as long as their guard. Old database memory is released when its last guard, snapshot or editor drops.

REPRODUCIBILITY / CONTEXT MATTERS

Every number
has a source.

Benchmark conditions are part of the result.

01 / RECORDED ENVIRONMENT

One machine. Comparable workloads.

02 / REPRODUCE

Run it on your hardware.

From a repository checkout, run the comparison suite and import its generated charts with updated environment metadata.

make bench-compare
# Or use the pinned Docker toolchains:
make bench-compare-docker

# Review website/benchmark-context.json
make website-build

The suite writes SVG charts, an HTML report and JSON/CSV results. Site builds import measurements; they do not run benchmarks.

03 / INTERPRETATION

Real measurements. Visible limitations.

One recorded execution

Precision follows the source

Values are parsed from SVG tooltip labels produced by the comparison suite. Unit conversions preserve their meaning; missing points are never converted to zero. Download the JSON to inspect labels and source hashes.

Concurrency and memory

Concurrency shows 4, 8 and 16 workers. The exported 1-worker series mixes fallback scenarios, so it is excluded. 8/16-worker results oversubscribe this 4-vCPU runner. RSS measures whole processes; it is not the database buffer size. Reader RSS compares four Rust/C readers, not Go.

Workloads and APIs matter

In lookup charts, 1M means query addresses, not database entries. The default City fixture inserts 4,097 IPv4 and 4,096 IPv6 networks. Random queries can miss. IPv6 absent, size scaling and RSS use separate fixtures.

Lookups include decoding on hits. Our reader uses typed borrowed City records; competitors use their own decoding APIs. This is a harness comparison, not an identical decoder microbenchmark.

Compared implementations

MEASURED VERSIONS

Library	Version	Language	Role

Inspect the benchmark source ↗