Simplest Hash Functions

(purplesyringa.moe)

43 points | by ibobev 4 days ago

8 comments

  • jandrewrogers 2 hours ago
    A “simplest” hash function is completely dependent on what you are using the hash function for and the guarantees you want a hash function to make. An optimal permutation of an integer is different from a small string hash is different from a block checksum. Literally, you are optimizing the algorithm for entirely different properties. No algorithm can satisfy all of them even approximately.

    The full scope of things hash functions are commonly used for requires at least four algorithms if you care about performance and optimality. It is disconcertingly common to see developers using hash algorithms in contexts where they are not fit for purpose. Gotta pick the right tool for the job.

    • eru 1 hour ago
      You are right!

      For example, when you know that your input is uniformly randomly distributed, then truncation is a perfectly good hash function. (And a special case of truncation is the identity function.)

      The above condition might sound too strong to be practical, but when you are eg dealing with UUIDs it is satisfied.

      Another interesting hash function: length. See https://news.ycombinator.com/item?id=6919216 for a bad example. For a good example: consider rmlint and other file system deduplicators.

      These deduplicators scan your filesystem for duplicates (amongst other things). You don't want to compare every file against every other file. So as a first optimisation, you compare files only by some hash. But conventional hashes like sha256 or crc take O(n) to compute. So you compute cheaper hashes first, even if they are weaker. Truncation, ie only looking at the first few bytes is very cheap. Determining the length of a file is even cheaper.

    • andai 35 minutes ago
      This comment feels about half as long as it ought to be. Can you say more?
  • benob 1 hour ago
    I just realized that a hash function is nothing less than the output of a deterministic random number generator xored with some data
    • derriz 30 minutes ago
      Could you explain what you mean here?

      Hashes are _functions_ so provide the same output given the same input.

      If you don't reseed the RNG after every hash computation, then you break this vital property of hashes.

      And if you do reseed, then your claim boils down to "every hash function is just an XOR against a contstant" which certainly is not true either.

    • eru 1 hour ago
      Sorry, what?

      That might we one very particular way to write a hash function, but it's far from the only one.

      Believe it or not, for some purposes taking the first few bytes of a string or even just the length of the string are good hash functions.

      • andai 32 minutes ago
        Well, that's technically also a deterministic random number generator! (I want to say it's not a great one, but... that's apparently context-dependent!)

        What are those purposes?

  • teo_zero 2 hours ago
    While I generally like to reinvent the wheel, for hash functions I strongly recommend to use a proved good one. Djb2 by the venerable Daniel Bernstein satisfies all the requirements of TFA.

      h = 5381
      while still has data:
        h = h * 33 + next_byte()
      return h
    
    PS of course if you think the multiplication is overkill, consider that it is nothing more than a shift and an addition.
    • purplesyringa 3 minutes ago
      Djb2 is hardly a proven good hash :) It's really easy to find collisions for it, and it's not seeded, so you're kind of screwed regardless. It's the odd middle ground between "safely usable in practice" and "fast in practice", which turns out to be "neither safe nor fast" in this case.
  • kuzivaai 1 hour ago
    The point about hash tables using top bits instead of bottom bits is the kind of thing that feels obvious once someone says it and yet here we are. Genuine question: have you seen any real-world hash table implementations that actually do this, or is it purely "this is what we should have done 40 years ago"?
    • adrian_b 4 minutes ago
      Which bits are best is completely dependent on the particular hash function.

      For instance, if the last operation during computing a hash function was the kind of integer multiplication where the result has the same length as the operands, then the top bits are the best (because they depend on more of the input bits than the bottom result bits).

      On the other hand, if the last operation during computing a hash function was the kind of integer multiplication where the result has a length that is the sum of the lengths of the operands (i.e. where the high bits of the result are not discarded), then the best bits of the result are neither the top bits nor the bottom bits, but the middle bits, for the same reason as before, they are the bits that depend on more of the input bits than either the bottom bits or the top bits of the result.

      When the last operation is an addition, because the carry bits propagate from the bottom bits towards the top bits, the top bits are the best, because the higher a bit is in the result there are more input bits on which it depends.

    • SkiFire13 1 hour ago
      I think older processors used to have a slower implementation for shifts, which made this slower.

      Nowadays swisstable and other similar hashtables use the top bits and simd/swar techniques to quickly filter out collisions after determining the starting bucket.

    • teo_zero 1 hour ago
      Sorry, I've read and reread TFA but the concept still evades me. Is it that, since it's easier for a hash function to have higher entropy in the higher bits than in the lower ones, it would be more logical for hash tables to discard the lower bits and keep the higher ones?
      • purplesyringa 1 minute ago
        Yes, that's my point. It's not true that all hash functions have this characteristic, but most fast ones do. (And if you're using a slow-and-high-quality hash function, the distinction doesn't matter, so might as well use top bits.)
  • Charon77 3 hours ago
    > Like addition

    I'm perplexed to the claim that addition is cheaper than XOR, especially since addition is built upon XOR, am I missing anything? Is it javascript specific?

    • pjscott 2 hours ago
      The wording was a bit unclear. The previous paragraph mentions wanting something cheaper than "those pesky XORs and multiplications". The multiplication is the expensive part; the (very cheap) XORs are just mildly annoying because you have to think about what they're doing.
    • 4k0hz 2 hours ago
      At least on x86, multiple additions and multiplications can be done with a single `lea` instruction so it's preferable to XOR. Though I have no idea about other architectures, compiler implementations, any interpreters...
      • ygra 2 hours ago
        That only helps with multiplications by statically known word sizes (4x, 8x, etc.) and not arbitrary x·y. It can help with many smaller constant multipliers if the complete is clever, but it has to be known at compile time.
  • johanvts 2 hours ago
    This is a excellent article for anyone looking for some more in-depth analysis of tabulation based hashing methods: https://arxiv.org/abs/1505.01523
  • Sharlin 1 hour ago

      def hash(str):
        len(str)
    
    O(1), baby!
    • nkrisc 7 minutes ago
      If you use the identity function as your hashing function then is it O(0) because you are done before you start?
    • eru 1 hour ago
      You'd probably want 'return len(str)', if this is Python?

      In any case, for some applications this is indeed a great hash function. Programs like rmlint use it as part of their checks for duplicate files: if files have different lengths, they can't have the same content after all.

      • andai 30 minutes ago
        Re: missing return keyword

        Actually, in Python, None is a valid key...

        (I'm so sorry. JavaScript has ruined me.)

        • tetha 22 minutes ago
          Any `return c` for some constant is a valid and correct hash function. It just has a lot of collisions and degenerates hash-maps to terrible performance. That was in fact my first thought when I read "simplest hash functions".
    • zimpenfish 1 hour ago
      O(1) only if you're working in a language with length-stored strings (like Pascal[0]), right?

      In something like C with its generic strings[1], it would surely have to be O(n) since you have to scan the entire string to calculate its length?

      (I have always been terrible at big-O, mind.)

      [0] There's probably more of them by now.

      [1] ie. not a specific length-stored string type.

  • bryanrasmussen 3 hours ago
    a hash function that produce hashes that are already in the hash table should, IMO, not be called a hash function.

    I get why technically it is a hash function, but still, no.

    • baruch 2 hours ago
      It is mathematically impossible for a proper hash function (one with an output range smaller than its input range) to not have collisions. The proof uses the Pigeon Hole Principle https://en.wikipedia.org/wiki/Pigeonhole_principle
      • bryanrasmussen 2 hours ago
        ok damn, I did not know this, obviously. Thanks.

        I guess I've never actually had this problem because was always hashing things that were static, or specialty cases like password hashes where the salt obviously guarantees uniqueness.

        • kro 1 hour ago
          It's very very unlikely to get collisions there, but still not impossible. Whenever you map data of arbitrary length (infinite possibilities) to a limited length collisions are possible.
          • eru 1 hour ago
            Doesn't even have to be arbitrary length.

            Whenever you map into a smaller space, you get collisions. The bigger space doesn't have to be infinite.

    • ahazred8ta 3 hours ago
      A perfect hash function https://en.wikipedia.org/wiki/Perfect_hash_function has to be specially constructed for each desired set of inputs. Generic hash functions cannot be 'perfect'.
    • phinnaeus 3 hours ago
      Here is a hash function that does not have hash collisions:

        fn hash(data):
          return data
      • tux3 42 minutes ago
        That is a function, but not a hash function!
      • Charon77 3 hours ago
        Well it no longer constrains the data in a fixed output length.
        • dbdr 2 hours ago
          Sure, but if you constrain to fixed output length, you will definitely have collisions (Pigeon Hole Principle). There's no way around that.
          • hsbauauvhabzb 2 hours ago
            padding with zeroes to a fixed length and prepending the original length would suffice, but you’d have to have a fixed length of double infinity to account for both the length information and the hash information, and the hash is less efficient than the original information.
      • oliver236 2 hours ago
        what programming language is this?