bloomfilter

Bloom Filter

This bloom filter implementation is a Go port of https://github.com/jasondavies/bloomfilter.js

The ability to build a bloom filter on a server in Go and evaluate that filter on a client in Javascript can have immense value for comparing application state in distributed single page applications with offline read/write capabilities and large data sets.

There are a lot of open source bloom filter implementations available on the internet. These implementations are mostly incompatible. Every repo adds its own special sauce to its hashing algorithms and hash derivation methods. After scouring the internet for a multi-language bloom filter implemenation, none appeared that fit the requirements.

So, the most popular actively maintained javascript bloom filter on Github was selected and ported to Go.

The reference implementation uses a non-standard fnv1a hashing algorithm. It is also less than 120 lines of javascript. This project proves that the reference implementation can be ported by a skilled developer to another language in a day.

Go Example

bloomfilter is thread safe with 95% test coverage

package main

import (
	"encoding/base64"
	"fmt"

	"github.com/httpimp/bloomfilter"
)

func main() {
	m, k := bloomfilter.EstimateParameters(10, 1e-6)
	bf := bloomfilter.New(m, k)
	bf.Add([]byte("foo"))
	bf.Add([]byte("bar"))
	encoded := base64.StdEncoding.EncodeToString(bf.ToBytes())
	fmt.Println(m)
	fmt.Println(k)
	fmt.Println(string(encoded))
}

288
20
iCCACAiAACAACIgAAAIIAqCAAIgogCAIAIgACAIAigiAAIqA

Javascript Example

Now we can take that same base64 encoded byte array and evaluate it with bloomfilter.js in the browser

var bits = sjcl.codec.base64.toBits("iCCACAiAACAACIgAAAIIAqCAAIgogCAIAIgACAIAigiAAIqA");
var bloom = new BloomFilter(bits, 20);
console.log(bloom.test("foo"));
console.log(bloom.test("bar"));
console.log(bloom.test("baz"));
bloom.add("baz");
console.log(sjcl.codec.base64.fromBits(bloom.buckets));

true
true
false
iCCACAiAACAACIgAQAIIAqSIEKgowKEKAowIGBIgyomAAIqI

Go Example Again

After deserializing the filter in javascript and altering it, we can send it back to the server again to confirm that it now includes the additional element.

package main

import (
	"encoding/base64"
	"log"

	"github.com/httpimp/bloomfilter"
)

func main() {
	decoded, err := base64.StdEncoding.DecodeString("iCCACAiAACAACIgAQAIIAqSIEKgowKEKAowIGBIgyomAAIqI")
	if err != nil {
		panic(err)
	}
	bf := bloomfilter.NewFromBytes(decoded, 21)
	log.Println(bf.Test([]byte("foo")))
	log.Println(bf.Test([]byte("bar")))
	log.Println(bf.Test([]byte("baz")))
	log.Println(bf.Test([]byte("bork")))
}

true
true
true
false

Performance

This is not the most efficient bloom filter available for Go. There are plenty of good options if you don't need portability.

That said, it still does perform reasonably well.

> go test -bench .
goos: linux
goarch: amd64
pkg: github.com/httpimp/bloomfilter
BenchmarkSeparateTestAndAdd-2            1000000              1224 ns/op
BenchmarkSeparateAdd-2                   2000000               702 ns/op
PASS
ok      github.com/httpimp/bloomfilter  3.320s

Will Fitzgerald's bloom is an excellent bloom filter written in Go

https://github.com/willf/bloom

> go test -bench .
goos: linux
goarch: amd64
pkg: github.com/willf/bloom
BenchmarkEstimated-2                    2000000000               0.10 ns/op
BenchmarkSeparateTestAndAdd-2            2000000               761 ns/op
BenchmarkCombinedTestAndAdd-2            3000000               757 ns/op
BenchmarkAdd-2                           3000000               599 ns/op
PASS
ok      github.com/willf/bloom  9.580s

This project is about 15% slower for Add and 40% slower for Test and Add compared with bloom. This performance penalty may be an acceptable trade-off in exchange for the portability this project is designed to provide. Weigh your application's efficiency requirements for bloom filter use and creation against the benefit of sharing filters with javascript applications before integrating this package.

Standing on the shoulders of giants

Thanks to @jasondavies for creating the reference implementation, a functioning bloom filter in < 120 lines of code. This port took about 6 hours.

Thanks to @willf for estimation.