Creating your own Tinyurl style uid

问题:

I'm writing a small article on humanly readable alternatives to Guids/UIDs, for example those used on TinyURL for the url hashes (which are often printed in magazines, so need to be short).

The simple uid I'm generating is - 6 characters: either a lowercase letter (a-z) or 0-9.

"According to my calculations captain", that's 6 mutually exclusive events, although calculating the probability of a clash gets a little harder than P(A or B) = P(A) + P(B), as obviously it includes numbers and from the code below, you can see it works out whether to use a number or letter using 50/50.

I'm interested in the clash rate and if the code below is a realistic simulation of anticipated clash rate you'd get from generating a hash. On average I get 40-50 clashes per million, however bare in mind the uid wouldn't be generated a million times at once, but probably only around 10-1000 times a minute.

What is the probability of a clash each time, and can anyone suggest a better way of doing it?

static Random _random = new Random();

public static void main()
{
    // Size of the key, 6
    HashSet<string> set = new HashSet<string>();
    int clashes = 0;
    for (int n=0;n < 1000000;n++)
    {
        StringBuilder builder = new StringBuilder();

        for (int i =0;i < 7;i++)
        {
            if (_random.NextDouble() > 0.5)
            {
                builder.Append((char)_random.Next(97,123));
            }
            else
            {
                builder.Append(_random.Next(0,9).ToString());
            }
        }

        if (set.Contains(builder.ToString()))
        {
            clashes++;
            Console.WriteLine("clash: (" +n+ ")" +builder.ToString());
        }

        set.Add(builder.ToString());
        _random.Next();
        //Console.Write(builder.ToString());
    }

    Console.WriteLine("Clashes: " +clashes);
    Console.ReadLine();
}

UPDATE: Here's the resulting article from this question

I really asked two questions here so I was cheating. The answer I was after was rcar's, however Sklivvz's is also the answer to the 2nd part (an alternative). Is it possible to make a custom unique id generator in a database, or would it be client side (which would be 2 possible reads first)?

The general idea I was after was using Ids in databases or other stores that can be used by phone or printed material, not a giant 16 byte guid.

UPDATE 2: I put the formula for two mutually exclusive events above instead of 2 independent ones (as getting an 'a' the first time doesn't mean you can't get an 'a' the second time). Should've been P(A and B) = P(A) x P(B)

回答1:

The probability of a collision against one specific ID is:

p = ( 0.5 * ( (0.5*1/10) + (0.5*1/26) ) )^6

which is around 1.7×10^-9.

The probability of a collision after generating n IDs is 1-p^n, so you'll have roughly a 0.17% chance of a collision for each new insertion after 1 million IDs have been inserted, around 1.7% after 10 million IDs, and around 16% after 100 million.

1000 IDs/minute works out to about 43 million/month, so as Sklivvz pointed out, using some incrementing ID is probably going to be a better way to go in this case.

EDIT:

To explain the math, he's essentially flipping a coin and then picking a number or letter 6 times. There's a 0.5 probability that the coin flip matches, and then 50% of the time there's a 1/10 chance of matching and a 50% chance of a 1/26 chance of matching. That happens 6 times independently, so you multiply those probabilities together.

回答2:

Why do you want to use a random function? I always assumed that tinyurl used a base 62 (0-9A-Za-z) representation of a sequential Id. No clashes and the urls are always as short as possible.

You would have a DB table like

Id  URL
 1  http://google.com
 2  ...
... ...
156 ...
... ...

and the corresponding URLs would be:

http://example.com/1
http://example.com/2
...
http://example.com/2W
...

回答3:

Look up the Birthday Paradox, it's the exact problem that you're running into.

The question is: How many people do you need to get together in a room, so that you have a 50% chance of any two people having the same birthdate? The answer may surprise you.

回答4:

Some time ago I did exactly this, and I followed the way Sklivvz mentioned. The whole logic was developed with a SQL server stored procedure and a couple of UDF (user defined functions). The steps were:

• say that you want to shorten this url: Creating your own Tinyurl style uid
• Insert the URL in a table
• Obtain the @@identity value of the last insert (a numeric id)
• Transform the id in a corresponding alphanumeric value, based on a "domain" of letters and numbers (I actually used this set: "0123456789abcdefghijklmnopqrstuvwxyz")
• Return that value back, something like 'cc0'

The conversion was realized thru a couple of very short UDF.

Two conversion called one after the other would return "sequential" values like these:

select dbo.FX_CONV (123456) -- returns "1f5n"

select dbo.FX_CONV (123457) -- returns "1f5o"

If you are interested I can share the UDF's code.

回答5:

Why not just use a hashing algorithm? and use a hash of the url?

if you are using random numbers chances are you will get clashes because they are indeterminate.

hashes arent proovably unique but there is a fairly good chance that the hash of a string will be unique.

Correction

Actually wait you want them to be humanly readable... if you put them in hex they are technically humanly readable.

or you could use an algorithm that converted a hash into a humanly readable string. if the humanly readable string is a different representation of the hash it should also be as "unique" as the hash, ie base 36 of the original hash.

回答6:

I would generate a random value representative of the data that you are going to hash, and then hash that and check clahses rather than trying to simulate with random manually made hashes. This will give you a better indicator. And you will have more randomness because you will have more to randomize (Assuming your data to be hashed is larger :) ).

回答7:

If you're using 6 characters, a-z and 0-9, thats a total of 36 characters. The number of permutations is thus 36^6 which is 2176782336.. so it should only clash 1/2176782336 times.

回答8:

from wikipedia:

When printing fewer characters is desired, GUIDs are sometimes encoded into a base64 or Ascii85 string. Base64-encoded GUID consists of 22 to 24 characters (depending on padding), for instance:

7QDBkvCA1+B9K/U0vrQx1A
7QDBkvCA1+B9K/U0vrQx1A==

and Ascii85 encoding gives only 20 characters, e. g.:

5:$Hj:Pf\4RLB9%kU\Lj 

So if you are concerned with uniqueness, a base64 encoded GUID gets you somewhat closer to what you want, though its not 6 characters.

Its best to work in bytes first, then translate those bytes into hexadecimal for display, rather than working with characters directly.