Getting too low level of randomness with LCGRandom about ceylon-random

[jvasileff]

Sure, I'll take a look. In the mean time, you can also try the platform rng to rule in/out the generator itself - https://modules.ceylon-lang.org/repo/1/com/vasileff/ceylon/random/api/0.0.5/module-doc/api/index.html#platformRandom

[renatoathaydes]

Ah, that's pretty cool... I will try it!

[jvasileff]

@renatoathaydes the statistical tests still pass, and I'm not seeing any significant issues. Where exactly are you seeing the duplicates? Given that with the default values, generateStrings varies the output length uniformly from 0 to 99 characters, I would expect duplicates for short strings to happen fairly often.

Without doing any math, the output for

    for (i in 0:100) {
        print(frequencies(generateStrings()).filter((entry) => entry.item > 1));
    }

looks reasonable to me, with a fairly common occurrence of multiple 0-length strings.

And this also looks reasonable, with several one and two length duplicates:

    for (i in 0:10k) {
        value nonEmptyStrings
            = frequencies(generateStrings())
                .filter((entry) => entry.item > 1)
                .map(Entry.key).filter(not("".equals));

        if (!nonEmptyStrings.empty) {
            print(nonEmptyStrings);
        }
    }

[renatoathaydes]

With the platform-specific Random, the test passes every time.

Because the length is also randomly chosen, it shouldn't happen often that you get two Strings of similar length.

This may be the problem but I find it strange that the other 2 implementations of Random I tried always succeed.

[jvasileff]

Yes, that is strange for the implementation to matter. But you'll have to let me what sort of call to generateStrings() is causing problems for me to take a close look.

As for the length, you will almost always get strings of the same length for 100 samples of 100 random lengths.

[jvasileff]

BTW, if the problem you are seeing is duplicate results for consecutive calls to generateStrings(), especially on JS, the fix is to create an LCGRandom() in the calling function, and use the same one for each call to generateStrings().

The reason for this is that the seed is based on millis + nanos, which isn't always a great source of unique numbers on JS.

[renatoathaydes]

You just found the problem, I believe :)

Anyway, what I had said about length, is that you need both length and contents to be the same to get a duplicate, which is really unlikely, I would think... but of course, there will be Strings with the same length...

[renatoathaydes]

I mentioned your library in the docs of https://github.com/renatoathaydes/specks
as I use it to create random test input.

Thanks!

[jvasileff]

both length and contents to be the same

yes, but for zero-length strings, the contents will always be the same

I mentioned your library

thanks, and congrats on the release!