Skip to content
This repository has been archived by the owner on Sep 23, 2018. It is now read-only.

Latest commit

 

History

History
123 lines (80 loc) · 3.93 KB

README.md

File metadata and controls

123 lines (80 loc) · 3.93 KB
Raw

Important: Zeroth, consider if a Cuckoo filter could be right for your use-case.

GoDoc travis

Face-meltingly fast, thread-safe, marshalable, unionable, probability- and optimal-size-calculating Bloom filter in go

Copyright © 2014-2016,2018 Barry Allard

MIT license

WTF is a bloom filter

**TL;DR: **Probabilistic, extra lookup table to track a set of elements kept elsewhere to reduce expensive, unnecessary set element retrieval and/or iterator operations when an element is not present in the set. It's a classic time-storage tradeoff algoritm.

Properties

See wikipedia for algorithm details

Impact What Description
Good No false negatives know for certain if a given element is definitely NOT in the set
Bad False positives uncertain if a given element is in the set
Bad Theoretical potential for hash collisions in very large systems and/or badly hash.Hash64-conforming implementations
Bad Add only Cannot remove an element, it would destroy information about other elements
Good Constant storage uses only a fixed amount of memory

Naming conventions

(Similar to algorithm)

Variable/function Description Range
m/M() number of bits in the bloom filter (memory representation is about m/8 bytes in size) >=2
n/N() number of elements present >=0
k/K() number of keys to use (keys are kept private to user code but are de/serialized to Marshal and file I/O) >=0
maxN maximum capacity of intended structure >0
p maximum allowed probability of collision (for computing m and k for optimal sizing) >0..<1
  • Memory representation should be exactly 24 + 8*(k + (m+63)/64) + unsafe.Sizeof(RWMutex) bytes.
  • Serialized (BinaryMarshaler) representation should be exactly 72 + 8*(k + (m+63)/64) bytes. (Disk format is less due to compression.)

Binary serialization format

All values in Little-endian format

Offset Offset (Hex) Length (bytes) Name Type
0 00 8 k uint64
8 08 8 n uint64
16 10 8 m uint64
24 18 k (keys) [k]uint64
24+8*k ... (m+63)/64 (bloom filter) [(m+63)/64]uint64
24+8*k+8*((m+63)/64) ... 48 (SHA384 of all previous fields, hashed in order) [48]byte
  • bloomfilter.Filter conforms to encoding.BinaryMarshaler and `encoding.BinaryUnmarshaler'

Usage

import "github.com/steakknife/bloomfilter"

const (
  maxElements = 100000
  probCollide = 0.0000001
)

bf, err := bloomfilter.NewOptimal(maxElements, probCollide)
if err != nil {
  panic(err)
}

someValue := ... // must conform to hash.Hash64

bf.Add(someValue)
if bf.Contains(someValue) { // probably true, could be false
  // whatever
}

anotherValue := ... // must also conform to hash.Hash64

if bf.Contains(anotherValue) {
  panic("This should never happen")
}

err := bf.WriteFile("1.bf.gz")  // saves this BF to a file
if err != nil {
  panic(err)
}

bf2, err := bloomfilter.ReadFile("1.bf.gz") // read the BF to another var
if err != nil {
  panic(err)
}

Design

Where possible, branch-free operations are used to avoid deep pipeline / execution unit stalls on branch-misses.

Get

go get -u github.com/steakknife/bloomfilter  # master is always stable

Source

Contact

License

MIT license

Copyright © 2014-2016 Barry Allard