Data structures
Bloom filter
"A Bloom filter is a space-efficient probabilistic data structure [...] used to test whether an element is a member of a set"[1], with occasional false positives (and quantifiable amount of them), but no false negatives.
A little more mechanically put, a bloom filter is
- a fixed-size structure (a balance between 'small meaning not wasting space' and 'larger to improve effectiveness')
- with constant-time adds and checks,
- where we add() all things we see,
- where a get() of the same key, asking "have we added (seen) X before", is answered either with
- "quite possibly" or
- "definitely not"
Why is that useful at all?
Consider you have huge yet slow-to-access storage.
It would be great if we only access that when we need to.
If we had a structure in RAM that could answer "I definitely don't have this" or "I definitely have this", that would be great, but that would grow alongside the size of the store, so over time you won't be able to guarantee you can keep it in RAM.
It turns out that if you can accept the answer "definitely not" and "possibly yes",
you can make a much smaller, fixed-size data structure so you can ensure it sits in RAM.
As the data in the backing store grows, it merely has to say "possibly" more often where before it could say "definitely not".
How?
While it's not precisely how they work, you can intuit the properties of how Bloom filters work, by modifying a hash map.
Consider
- deciding how many buckets you have up front
- for each bucket, store just true or false (instead a list of all values, as a real hashmap would do)
Since you're still deciding bucket by hash of the key,
- if you fetch a false from that it means you didn't previously set a key like that.
- if you fetch get a true from that, it ends up meaning "the hash can always have collided, so I can only answer possibly".
This also helps the intuition on why having a smaller-sized bloom filter (in the hashmap analogy: fewer buckets) increases the the possibility of getting a 'possibly' rather than a certain 'nope'.
You probably want to make the bloom filter as large as your RAM allows, to increase the amount of "nope"s and decrease the amount of false possibles.
But for some applications, the bloom filter can be quite small, and still e.g. save a reasonable amount of accesses, which is why this is useful even on microcontrollers.
You can quantify how large of a bloom filter you will want for a given set size, and desired error rate - there are calculators.
Alternatives
Depending on RAM abundance, CPU power, and on the case, while a much larger bloom filter works well enough, you may be better served by
- a bit array doing the same task but giving a fully correct answer
- more compact than a hash map, but with more practical restrictions
- has to fit in memory for your given, fixed-size set -- or has to grow with the set (a bloom filter just becomes less efficient)
- also requires you to have a simple, non-colliding mapping into that array
- more compact than a hash map, but with more practical restrictions
- a hash map doing the same task but giving a fully correct answer
- always falling back to slow storage if misses, but just counting on most things being hits
- may not only checks presence from RAM, but may often gives the answer from RAM
- (and a bloom filter in front of a memcache is often pointless)
See also