Awesome
xorfilter: Go library implementing xor and binary fuse filters
Bloom filters are used to quickly check whether an element is part of a set. Xor and binary fuse filters are a faster and more concise alternative to Bloom filters. Furthermore, unlike Bloom filters, xor and binary fuse filters are naturally compressible using standard techniques (gzip, zstd, etc.). They are also smaller than cuckoo filters. They are used in production systems.
- Thomas Mueller Graf, Daniel Lemire, Binary Fuse Filters: Fast and Smaller Than Xor Filters, Journal of Experimental Algorithmics (to appear). DOI: 10.1145/3510449
- Thomas Mueller Graf, Daniel Lemire, Xor Filters: Faster and Smaller Than Bloom and Cuckoo Filters, Journal of Experimental Algorithmics 25 (1), 2020. DOI: 10.1145/3376122
This Go library is used by
- coherence-go-client: the Oracle Coherence client
- Matrixone: a Hyperconverged cloud-edge native database
We are assuming that your set is made of 64-bit integers. If you have strings or other data structures, you need to hash them first to a 64-bit integer. It is not important to have a good hash function, but collision should be unlikely (~1/2^64). A few collisions are acceptable, but we expect that your initial set should have no duplicated entry.
The current implementation has a false positive rate of about 0.3% and a memory usage of less than 9 bits per entry for sizeable sets.
You construct the filter as follows starting from a slice of 64-bit integers:
filter,_ := xorfilter.PopulateBinaryFuse8(keys) // keys is of type []uint64
It returns an object of type BinaryFuse8
. The 64-bit integers would typically be hash values of your objects.
You can then query it as follows:
filter.Contains(v) // v is of type uint64
It will always return true if v was part of the initial construction (Populate
) and almost always return false otherwise.
An xor filter is immutable, it is concurrent. The expectation is that you build it once and use it many times.
Though the filter itself does not use much memory, the construction of the filter needs many bytes of memory per set entry.
For persistence, you only need to serialize the following data structure:
type BinaryFuse8 struct {
Seed uint64
SegmentLength uint32
SegmentLengthMask uint32
SegmentCount uint32
SegmentCountLength uint32
Fingerprints []uint8
}
When constructing the filter, you should ensure that there are not too many duplicate keys for best results.