Probablistic filters are high-speed, space-efficient data structures that support set-membership tests with a one-sided error. These filters can claim that a given entry is definitely not represented in a set of entries, or might be represented in the set. That is, negative responses are conclusive, whereas positive responses incur a small false positive probability (FPP).
The trade-off for this one-sided error is space-efficiency. Cuckoo Filters and Bloom Filters require approximately 7 bits per entry at 3% FPP, regardless of the size of the entries. This makes them useful for applictations where the volume of original data makes traditional storage impractical.
Bloom filters have been in use since the 1970s and are well understood. Implementations are widely available. Variants exist that support deletion and counting, though with expanded storage requirements.
Cuckoo filters were described in Cuckoo Filter: Practically Better Than Bloom, a paper by researchers at CMU in 2014. Cuckoo filters improve on Bloom filters by supporting deletion, limited counting, and bounded FPP with similar storage efficiency as a standard Bloom filter.
Below is side-by-side simulation of the inner workings of Cuckoo and Bloom filters.