ribbonGo

March 14, 2026 · View on GitHub

A high-performance Ribbon filter implementation in Go — the space-efficient probabilistic data structure that is practically smaller than Bloom and Xor filters.

Based on the paper:

"Ribbon filter: practically smaller than Bloom and Xor" Peter C. Dillinger & Stefan Walzer, 2021 (arXiv:2103.02515)

What is a Ribbon Filter?

A Ribbon filter is a static, space-efficient probabilistic data structure for approximate set membership queries — the same problem Bloom filters solve, but using significantly less space.

Given a set of keys, a Ribbon filter can answer "is this key in the set?" with:

No false negatives — if a key is in the set, the filter always says yes.
Configurable false positive rate — non-member keys have a tunable probability of being incorrectly reported as members (FPR ≈ 2^−r for r result bits).
Near-optimal space — approaches the information-theoretic lower bound, using only ~1–5% more space than the minimum possible.

Ribbon vs Bloom vs Xor

Property	Bloom	Xor	Ribbon
Space overhead	~44% over optimal	~23% over optimal	~1–5% over optimal
Construction	Fast, incremental	Fast, static	Moderate, static
Supports deletion	No (without counting)	No	No
FPR at 1% target (bits/key)	9.6	9.8	≈ 8.0

Ribbon filters are ideal for read-heavy, write-once workloads — LSM-tree engines (RocksDB, LevelDB), static lookup tables, networking data planes, and any scenario where a set is built once and queried millions of times.

Quick Start

Installation

go get github.com/RibbonGo/ribbonGo

Requires Go 1.25+.

Usage

package main

import (
    "fmt"
    "log"

    "github.com/RibbonGo/ribbonGo"
)

func main() {
    // Create a filter with default settings (w=128, r=7, FPR ≈ 0.78%)
    f := ribbon.New()

    // Build the filter from a set of keys
    keys := []string{"apple", "banana", "cherry", "date", "elderberry"}
    if err := f.Build(keys); err != nil {
        log.Fatal(err)
    }

    // Query membership
    fmt.Println(f.Contains("banana"))    // true  (always correct for members)
    fmt.Println(f.Contains("fig"))       // false (probably — FPR ≈ 0.78%)
}

Custom Configuration

// Trade space for faster construction with a narrower ribbon width
f := ribbon.NewWithConfig(ribbon.Config{
    CoeffBits:           64,    // w=64: balanced speed/space
    ResultBits:          8,     // r=8: FPR ≈ 0.39%
    FirstCoeffAlwaysOne: true,  // deterministic pivot (faster)
    MaxSeeds:            256,   // hash seed retries before failure
})

if err := f.Build(keys); err != nil {
    // err is ribbon.ErrConstructionFailed if banding fails
    log.Fatal(err)
}

API Reference

The public API is intentionally minimal — a single type with four functions.

Types

Type	Description
`Ribbon`	The main filter type. Create → Build → Contains.
`Config`	Construction parameters (ribbon width, result bits, etc.).
`ErrConstructionFailed`	Sentinel error returned when banding fails for all seed retries.

Functions

Function	Signature	Description
`New`	`func New() *Ribbon`	Create a filter with defaults: w=128, r=7, fcao=true, maxSeeds=256
`NewWithConfig`	`func NewWithConfig(cfg Config) *Ribbon`	Create a filter with custom parameters. Panics on invalid config.
`Build`	`func (r *Ribbon) Build(keys []string) error`	Construct the filter from a set of unique keys. May be called multiple times.
`Contains`	`func (r *Ribbon) Contains(key string) bool`	Test membership. Zero allocations, safe for concurrent use after Build.

Config Fields

Field	Type	Default	Description
`CoeffBits`	`uint32`	128	Ribbon width w ∈ {32, 64, 128}. Larger → more compact, slower to build.
`ResultBits`	`uint`	7	Fingerprint bits r ∈ [1, 8]. FPR ≈ 2^−r.
`FirstCoeffAlwaysOne`	`bool`	true	Force bit 0 of coefficient rows to 1 for deterministic pivoting.
`MaxSeeds`	`uint32`	256	Maximum hash seed retries before returning `ErrConstructionFailed`.

Features

Paper-faithful implementation — every design decision traces to a specific section of Dillinger & Walzer (2021), with §N citations throughout the code
Configurable ribbon width — w ∈ {32, 64, 128} to trade construction speed for space efficiency
Configurable result bits — r bits per key for FPR ≈ 2^−r
Dynamic slot computation — overhead ratio grows logarithmically with n, ported from RocksDB's empirical tables (ribbon_config.cc)
firstCoeffAlwaysOne optimisation — deterministic pivot from paper §4, skipping leading-zero scan
SoA memory layout — Struct-of-Arrays for maximum cache utilisation (8 coefficients per cache line at w=64)
Width-specialised inner loops — pure uint64 path for w≤64, separate lo/hi ops for w=128
Software-pipelined prefetching — AddRange hides L2/L3 latency by prefetching the next key's slot
Zero allocations on hot paths — no heap escapes during construction or query
Two-phase hash pipeline — hash keys once, remix per seed attempt; uses XXH3 (64-bit)
Research-friendly — all paper parameters exposed; reference (slow*) implementations included for cross-validation

Benchmarks

Apple M3 Pro · Go 1.25 · ARM64 · r = 7 result bits · firstCoeffAlwaysOne = true

All benchmarks follow the methodology of Dillinger & Walzer (2021), testing at both n = 10⁶ and n = 10⁸ keys.

Build Performance

Construction throughput and space efficiency at scale.

n	Width	ns/key	bits/key	Overhead
10⁶	w=32	56.89	10.54	31.81%
10⁶	w=64	64.56	8.959	11.99%
10⁶	w=128	106.3	8.380	4.749%
10⁸	w=32	355.3	11.62	45.30%
10⁸	w=64	266.2	9.406	17.58%
10⁸	w=128	384.7	8.585	7.314%

At n = 10⁶, w=128 achieves 8.38 bits/key — only 4.7% above the information-theoretic minimum (7 bits for r=7). At n = 10⁸, it remains under 8.6 bits/key with just 7.3% overhead.

Query Performance

Lookup latency per key (n = 10⁶).

Width	Positive (ns/op)	Negative (ns/op)
w=32	37.25	36.85
w=64	53.70	52.49
w=128	88.66	84.73

Query time scales linearly with ribbon width, as expected — the inner loop performs a dot product over w bits. Positive and negative queries have nearly identical cost.

Space Efficiency

Bits per key at both scales, with packed (information-theoretic) comparison.

n	Width	bits/key	packed bits/key	Overhead
10⁶	w=32	10.54	9.227	31.81%
10⁶	w=64	8.959	7.839	11.99%
10⁶	w=128	8.380	7.332	4.749%
10⁸	w=32	11.62	10.17	45.30%
10⁸	w=64	9.406	8.231	17.58%
10⁸	w=128	8.585	7.512	7.314%

w=128 at n = 10⁶ uses only 8.38 bits/key — compare with Bloom filters at 9.6 bits/key for the same FPR. That's a 12.7% space saving over Bloom.

Run Benchmarks Yourself

# Full benchmark suite
go test -bench=. -benchmem -count=3 ./...

# Paper-aligned benchmarks at n=10⁶ and n=10⁸
go test -run=^$ -bench='BenchmarkRibbon' -benchtime=3s -count=1

# Specific benchmark
go test -run=^$ -bench='BenchmarkRibbonBuild/w=128/n=1000000' -benchtime=3s

Architecture

The implementation follows the paper's full algorithmic pipeline:

Key → [Hash] → [Bander] → [Solver] → [Filter/Query]
       §2         §2,§4       §2           §2

Pipeline Layers

Layer	File	Description
uint128	`uint128.go`	128-bit integer type for coefficient rows when w=128.
Hash	`hash.go`	Two-phase pipeline: hash each key once with XXH3, then cheaply remix per seed to derive `(start, coeffRow, result)` triples.
Bander	`bander.go`	On-the-fly Gaussian elimination over GF(2). Converts hashed equations into an upper-triangular banded matrix. Hottest code path during construction.
Solver	`solver.go`	Back-substitution: solves the upper-triangular system to produce the compact solution vector encoding the filter.
Filter	`filter.go`	Query evaluation: one hash, one dot product over GF(2). Also provides false-positive rate estimation.
Builder	`builder.go`	Orchestrates the full pipeline: hashing → banding → solving → filter construction. Includes RocksDB-style dynamic slot computation.
Public API	`ribbon.go`	Sole public surface: `Ribbon`, `Config`, `New()`, `NewWithConfig()`, `Build()`, `Contains()`.

Key Optimisations

SoA layout — Coefficient data is stored in parallel arrays (coeffLo []uint64, coeffHi []uint64, result []uint8) instead of an array of structs. For w≤64, coeffHi is nil — zero memory, zero operations. This doubles the number of coefficients per cache line compared to AoS layout.

Width specialisation — add() dispatches to addW64() (pure uint64, ~10 ARM64 instructions per elimination step) or addW128() (separate lo/hi uint64 ops). The generic uint128.rsh() branch dispatch is avoided entirely.

Software-pipelined prefetching — addRange() touches coeffLo[nextKey.start] while processing the current key, pulling the next cache line into L1. This mirrors RocksDB's BandingAddRange approach and yields 20–36% throughput improvement at scale.

Dynamic overhead ratio — The number of slots m is computed using RocksDB-style empirical tables where overhead grows logarithmically with n. This ensures reliable banding at all scales, unlike a fixed overhead formula that breaks down at large n.

Two-phase hashing — Keys are hashed once with XXH3. On seed retry, only a cheap remix is applied to derive new (start, coeff, result) triples — no re-hashing of the original key data.

How It Works

At a high level, a Ribbon filter encodes set membership as a system of linear equations over GF(2):

Hash each key to a triple: starting position s, a w-bit coefficient row c, and an r-bit result r.
Band the equations into an upper-triangular matrix via Gaussian elimination (the "banding" step).
Solve the triangular system by back-substitution, producing a compact solution vector Z.
Query a candidate key by computing its triple, then checking if c · Z[s..s+w] == r.

The "ribbon" name refers to the banded structure of the coefficient matrix — each equation touches only w consecutive columns starting at position s, forming a ribbon-like diagonal band.

Why is it smaller than Bloom?

A Bloom filter wastes space because its bit-setting positions are independent — you can't pack information as tightly. A Ribbon filter encodes key membership as equations, allowing the solver to pack nearly r bits of information per key into the solution vector. The information-theoretic minimum is r bits/key, and Ribbon achieves within 1–5% of this.

Testing

23 test functions across the public API, plus extensive internal tests — all passing.

Tests cover:

Correctness — single insertion, no false negatives, false-positive rate validation, collision chains, 128-bit boundary crossing, redundant/contradictory equations
All configurations — w ∈ {32, 64, 128} × firstCoeffAlwaysOne ∈ {true, false} × r ∈ {1, 4, 7, 8}
Scale — builds at n = 100,000 verified against expected FPR
Edge cases — empty input, single key, nil/empty filter, rebuild semantics, invalid config panics
Cross-validation — optimised add() vs reference slowadd(), addRange() vs add()-loop — all verified slot-by-slot

go test -v -count=1 ./...

Project Structure

ribbon/
├── ribbon.go              # Public API (Ribbon, Config, New, Build, Contains)
├── builder.go             # Pipeline orchestration, slot computation
├── bander.go              # Gaussian elimination (banding)
├── hash.go                # XXH3-based two-phase hash pipeline
├── solver.go              # Back-substitution solver
├── filter.go              # Query evaluation and FPR estimation
├── uint128.go             # 128-bit integer type
├── ribbon_test.go         # Public API tests (23 functions)
├── ribbon_bench_test.go   # Paper-aligned benchmarks
├── filter_test.go         # Internal filter/pipeline tests
├── filter_bench_test.go   # Internal benchmarks
├── bander_test.go         # Bander unit tests
├── bander_bench_test.go   # Bander benchmarks
├── hash_test.go           # Hash layer tests
├── hash_bench_test.go     # Hash benchmarks
├── solver_test.go         # Solver tests
├── solver_bench_test.go   # Solver benchmarks
├── uint128_test.go        # uint128 tests
├── go.mod
├── go.sum
├── LICENSE
└── README.md

Contributing

Contributions are welcome! This project follows the paper's design closely — please read the relevant paper section before modifying any algorithm.

Getting Started

git clone https://github.com/RibbonGo/ribbonGo.git
cd ribbonGo
go test ./...

Guidelines

Paper-first — every design decision should cite a specific section (§N) of Dillinger & Walzer (2021)
RocksDB cross-references — use [RocksDB: FunctionName in file.h] format for implementation parallels
All 6 configs — tests must cover w ∈ {32, 64, 128} × firstCoeffAlwaysOne ∈ {true, false}
Reference implementations — include slow* variants for cross-validation of optimised code
Zero allocations — hot paths must not escape to heap; verify with go test -bench=X -benchmem
Naming — unexported everything (standardBander, not StandardBander); constants use kCamelCase

References

Dillinger, P. C. & Walzer, S. (2021). Ribbon filter: practically smaller than Bloom and Xor. arXiv:2103.02515
RocksDB Ribbon filter: ribbon_impl.h, ribbon_alg.h, ribbon_config.cc
XXH3 hash function (Go): github.com/zeebo/xxh3