Support __index__() in getrandbits(), randbytes() and others

[serhiy-storchaka]

Feature or enhancement

Support of the __index__() method in getrandbits() and randbytes() is inconsistent.

• Random.getrandbits() supports __index__(), but SystemRandom.getrandbits() does not.
• SystemRandom.randbytes() supports __index__(), but Random.randbytes() does not.

For consistence, both classes should support it.

sample() fails if its k argument is not integer. k should have __index__(), but it also should support comparison with integers and multiplication by list and integer. It is easy to make it more general, and only require __index__(). It will also produce more meaningful error.

binomialvariate() seems "working" with non-integer n argument, but the result is questionable. Passing non-integer value is most likely an error, and it would be better to detect it.

cc @rhettinger

[rhettinger]

Personally, I don't think this is needed. It is just overhead that doesn't benefit real users.

In pure python code, we commonly do much less error checking on inputs than we do in C code.

[serhiy-storchaka]

Yes, unfortunately it has a cost.

The worst case:

$ ./python -m timeit -s 'from random import randbytes' 'randbytes(1)'
Unpatched: 2000000 loops, best of 5: 145 nsec per loop
Patched:   2000000 loops, best of 5: 161 nsec per loop

It is less for larger argument:

$ ./python -m timeit -s 'from random import randbytes' 'randbytes(128)'
Unpatched: 500000 loops, best of 5: 616 nsec per loop
Patched:   500000 loops, best of 5: 633 nsec per loop

SystemRandom.getrandbits() is so slow, that the difference is insignificant.

$ ./python -m timeit -s 'import random; getrandbits = random.SystemRandom().getrandbits' 'getrandbits(1)'
Unpatched: 200000 loops, best of 5: 1.45 usec per loop
Patched:   200000 loops, best of 5: 1.47 usec per loop

For binomialvariate(), the effect is much weaker for n > 1.

$ ./python -m timeit -s 'from random import binomialvariate' 'binomialvariate()'
Unpatched: 1000000 loops, best of 5: 214 nsec per loop
Patched:   1000000 loops, best of 5: 229 nsec per loop
$ ./python -m timeit -s 'from random import binomialvariate' 'binomialvariate(2)'
Unpatched: 500000 loops, best of 5: 575 nsec per loop
Patched:   500000 loops, best of 5: 588 nsec per loop
$ ./python -m timeit -s 'from random import binomialvariate' 'binomialvariate(10)'
Unpatched: 200000 loops, best of 5: 1.11 usec per loop
Patched:   200000 loops, best of 5: 1.12 usec per loop

On other hand, operator.index() is already called multiple times for randrange() and randint().

[rhettinger]

I recommend closing this. It only solves a personal itch, not a real user issue.

[picnixz]

I'd suggest documenting that if someone wants to support non-integer out of the box (like, just call int(x) everytime). Or simply document that only int objects should be passed, without mentioning that some methods accept arbitrary objects supporting __index__.