ByTech.BloomFilter

A high-performance, benchmark-driven Bloom filter library for .NET 10. Zero-allocation hot path, span-first API, versioned binary serialization, thread-safe variant, counting filter with deletion, batch operations, named filter factory, DI integration, and strong statistical validation.

---

What is a Bloom filter?

A Bloom filter is a space-efficient probabilistic data structure for membership testing. It can tell you "definitely not in the set" or "possibly in the set" — never the reverse.

• No false negatives — if an element was added, MayContain always returns true
• Possible false positives — MayContain may return true for elements never added
• Space-efficient — uses far less memory than a hash set for large cardinalities
• O(k) operations — both Add and MayContain are constant-time

Common use cases: cache filtering, duplicate detection, network routing, database query optimization, and pre-screening large datasets.

---

Quick start

using ByTech.BloomFilter;

// Standard filter
var filter = BloomFilterBuilder
    .ForExpectedInsertions(1_000_000)
    .WithFalsePositiveRate(0.01)
    .Build();

filter.Add("hello"u8);
filter.Add("world");
filter.MayContain("hello"u8);  // true
filter.MayContain("other");    // false

// Thread-safe filter (lock-free adds)
using var safeFilter = BloomFilterBuilder
    .ForExpectedInsertions(1_000_000)
    .WithFalsePositiveRate(0.01)
    .BuildThreadSafe();

// Counting filter (supports deletion)
var counting = CountingBloomFilterBuilder
    .ForExpectedInsertions(100_000)
    .WithFalsePositiveRate(0.01)
    .Build();

counting.Add("item"u8);
counting.MayContain("item"u8);  // true
counting.Remove("item"u8);      // true (decremented)
counting.MayContain("item"u8);  // false

// Batch operations (works on any IBloomFilter)
IBloomFilter bf = filter;
bf.AddRange(new[] { "a"u8.ToArray(), "b"u8.ToArray() }
    .Select(x => (ReadOnlyMemory<byte>)x).ToArray());
bf.ContainsAll(...);  // all present?
bf.ContainsAny(...);  // any present?

// DI integration (separate package: ByTech.BloomFilter.DependencyInjection)
services.AddBloomFilter(bf =>
{
    bf.AddFilter("users", b => b.WithExpectedInsertions(1_000_000).WithFalsePositiveRate(0.01));
    bf.AddThreadSafeFilter("sessions", b => b.WithExpectedInsertions(500_000).WithFalsePositiveRate(0.001));
});

// Resolve via factory
var factory = serviceProvider.GetRequiredService<IBloomFilterFactory>();
var usersFilter = factory.Get("users");

---

Features

• Zero-allocation hot path — Add and MayContain allocate nothing on the heap
• Span-first API — primary path is ReadOnlySpan<byte>; string and byte[] overloads delegate to it
• Thread-safe variant — ThreadSafeBloomFilter with lock-free Interlocked.Or for concurrent adds
• Counting Bloom filter — CountingBloomFilter with 4-bit counters supporting Add/Remove/MayContain
• IBloomFilter interface — common surface for all filter types; polymorphic usage via factory or DI
• Batch operations — AddRange, ContainsAll, ContainsAny on all filter types + string/generic extensions
• Named filter factory — IBloomFilterFactory / BloomFilterFactory for managing multiple named filters
• DI integration — ByTech.BloomFilter.DependencyInjection package with AddBloomFilter() for IServiceCollection
• Generic T support — IBloomFilterKeySerializer<T> + extension methods for typed keys
• Double-hashing position derivation — XxHash128 split into h1/h2, generating k positions in O(1)
• Packed ulong[] bit storage — 64-bit word operations for high throughput
• Configurable parameters — specify expected insertions and target false positive rate; optimal m and k computed automatically
• Versioned binary serialization — persist and restore with round-trip correctness and CRC32 corruption detection
• ETW telemetry — opt-in EventSource counters (bloom.items_added, bloom.queries) with zero cost when no listener
• Diagnostics — saturation metrics, popcount, estimated current false positive rate

---

API

Standard Bloom filter

var filter = BloomFilterBuilder
    .ForExpectedInsertions(10_000_000)
    .WithFalsePositiveRate(0.001)
    .Build();

filter.Add("key"u8);
filter.MayContain("key"u8);  // true
filter.Clear();
filter.Snapshot();            // diagnostics: bits set, fill ratio, estimated FPR

Thread-safe Bloom filter

using var filter = BloomFilterBuilder
    .ForExpectedInsertions(1_000_000)
    .WithFalsePositiveRate(0.01)
    .BuildThreadSafe();

// Safe for concurrent Add + MayContain from multiple threads
// Add uses lock-free Interlocked.Or — no contention between writers
// Clear acquires exclusive lock

Counting Bloom filter

var filter = CountingBloomFilterBuilder
    .ForExpectedInsertions(100_000)
    .WithFalsePositiveRate(0.01)
    .Build();

filter.Add("item"u8);
filter.MayContain("item"u8);  // true
filter.Remove("item"u8);      // true — counters decremented
filter.MayContain("item"u8);  // false

// 4-bit counters, saturate at 15 (sticky — prevents overflow-induced false negatives)
// Uses 4x memory of standard filter

Batch operations

// Available on all filter types via IBloomFilter
IBloomFilter filter = BloomFilterBuilder
    .ForExpectedInsertions(100_000)
    .WithFalsePositiveRate(0.01)
    .Build();

// Byte-based batch
var items = new ReadOnlyMemory<byte>[] { "a"u8.ToArray(), "b"u8.ToArray(), "c"u8.ToArray() };
filter.AddRange(items);
filter.ContainsAll(items);  // true — all present
filter.ContainsAny(items);  // true — at least one present

// String batch (extension methods)
filter.AddRange(new[] { "hello", "world" });
filter.ContainsAll(new[] { "hello", "world" });  // true

Named filter factory

var factory = new BloomFilterFactory();
factory.Register("users", BloomFilterBuilder.ForExpectedInsertions(1_000_000).WithFalsePositiveRate(0.01).Build());
factory.Register("sessions", BloomFilterBuilder.ForExpectedInsertions(500_000).WithFalsePositiveRate(0.001).BuildThreadSafe());

var usersFilter = factory.Get("users");
factory.TryGet("missing", out var f);  // false

DI integration

// Package: ByTech.BloomFilter.DependencyInjection
services.AddBloomFilter(bf =>
{
    bf.AddFilter("users", b => b.WithExpectedInsertions(1_000_000).WithFalsePositiveRate(0.01));
    bf.AddThreadSafeFilter("sessions", b => b.WithExpectedInsertions(500_000).WithFalsePositiveRate(0.001));
    bf.AddCountingFilter("temp-keys", b => b.WithExpectedInsertions(10_000).WithFalsePositiveRate(0.05));
});

// Resolve anywhere via IBloomFilterFactory
var factory = serviceProvider.GetRequiredService<IBloomFilterFactory>();
var filter = factory.Get("users");

Generic T via serializer

public class GuidSerializer : IBloomFilterKeySerializer<Guid>
{
    public int GetMaxByteCount(Guid value) => 16;
    public int Serialize(Guid value, Span<byte> destination)
    {
        value.TryWriteBytes(destination);
        return 16;
    }
}

var filter = BloomFilterBuilder.ForExpectedInsertions(10_000).WithFalsePositiveRate(0.01).Build();
var serializer = new GuidSerializer();

filter.Add(myGuid, serializer);
filter.MayContain(myGuid, serializer);

Serialization

// Save
using var stream = File.Create("filter.bin");
BloomFilterSerializer.WriteTo(filter, stream);

// Load
using var input = File.OpenRead("filter.bin");
var restored = BloomFilterSerializer.ReadFrom(input);

---

Benchmarks

Environment: .NET 10.0.5, x64 RyuJIT, Release, Windows 11

Add throughput

Method	Key size	Mean	Allocated
Add(byte[])	16 B	32 ns	0 B
Add(byte[])	128 B	42 ns	0 B
Add(string)	16 B	55 ns	0 B
Add(string)	128 B	68 ns	0 B

Query throughput

Method	Mean	Allocated
MayContain(present)	32 ns	0 B
MayContain(absent)	22 ns	0 B

Zero heap allocations confirmed across all hot-path operations via [MemoryDiagnoser].

dotnet run --project benchmarks/ByTech.BloomFilter.Benchmarks -c Release

See benchmark methodology for environment requirements and interpretation rules.

---

How it works

Parameter calculation

Given expected insertions n and target false positive rate p:

m = -(n × ln(p)) / (ln(2)²)     — optimal bit count
k = (m / n) × ln(2)              — optimal hash function count

Double hashing

XxHash128 produces 128 bits → split into h1 (low 64) and h2 (high 64). Positions derived via:

position(i) = (h1 + i × h2) mod m    for i = 0, 1, ..., k-1

Bit storage

Bits are packed into a ulong[] array. Set and test operations use word-level indexing and bit masking — no per-operation allocations. The thread-safe variant uses Interlocked.Or for atomic bit-setting.

---

Thread safety

Type	Concurrent Add	Concurrent Query	Concurrent Add+Query	Clear
BloomFilter	No	No	No	No
ThreadSafeBloomFilter	Yes (lock-free)	Yes	Yes	Exclusive lock

The standard BloomFilter requires external synchronization. ThreadSafeBloomFilter uses Interlocked.Or for lock-free atomic bit-setting — bits only transition 0→1, making this safe without locks.

---

Telemetry

Opt-in ETW counters via System.Diagnostics.Tracing.EventSource:

Counter	Type	Description
bloom.items_added	Rate	Items added per second
bloom.queries	Rate	Queries per second
bloom.false_positive_estimate	Gauge	Estimated current FPR

dotnet counters monitor --process-id <pid> --counters ByTech.BloomFilter

Zero overhead when no listener is attached.

---

Project structure

ByTech.BloomFilter.slnx
├── src/
│   ├── ByTech.BloomFilter/                          # Core library
│   │   ├── Configuration/                           # Parameter planning, validation
│   │   ├── Hashing/                                 # XxHash128 double-hashing
│   │   ├── Storage/                                 # BitStore, ConcurrentBitStore, CountingBitStore
│   │   ├── Serialization/                           # Versioned binary format with CRC32
│   │   └── Diagnostics/                             # Snapshot, EventSource
│   └── ByTech.BloomFilter.DependencyInjection/      # DI integration package
├── tests/
│   ├── ByTech.BloomFilter.Tests/                    # Unit, integration, statistical, DI
│   └── ByTech.BloomFilter.FuzzTests/                # Fuzz testing
├── benchmarks/
│   └── ByTech.BloomFilter.Benchmarks/               # BenchmarkDotNet suite (5 classes)
└── .docs/                                           # Project documentation

---

Limitations

• Counting filter uses 4x memory — 4-bit counters vs 1-bit positions
• Counter saturation — counters at 15 are sticky and cannot be decremented
• Little-endian serialization — binary format assumes little-endian host (standard for all .NET 10 platforms)
• Maximum capacity — 2^37 bits (~16 GB); exceeding this throws during construction
• Counting filter serialization — not yet implemented (standard filter serialization only)

---

Version

Current: v1.2.0

---

License

Copyright 2026 ByTech. Licensed under the Apache License, Version 2.0.