cc @pitrou @serhiy-storchaka @rhettinger

I didn't change the documentation. Or do you think that it must be said explicitly that Random subclass implement randbytes() with getrandbits()? My minor documentation issue is that SystemRandom subclasses don't get a randbytes() implementation which uses getrandbits(): SystemRandom subclasses get SystemRandom.randbytes() which calls os.urandom().

I didn't change the documentation. (...) My minor documentation issue is that SystemRandom subclasses don't get a randbytes() implementation which uses getrandbits() (...)

Oh, I found a project in the wild which creates a SystemRandom subclass:
https://pypi.org/project/random-xe/

So I modified my PR to change also SystemRandom subclasses, and I completed the documentation.

I updated the PR to address @serhiy-storchaka's comments and I added more tests on two other kinds of subclasses.

I am not sure that pickling the randbytes method of Random subclass works correctly. If it does not work it may be better to use the second option for the randbytes implementation.

I am not sure that pickling the randbytes method of Random subclass works correctly. If it does not work it may be better to use the second option for the randbytes implementation.

pickle can be used to serialize/deserialize instances. __init_subclass__() is called when a new class is created.

Example:

import random
import pickle

class MyRandom(random.Random):
    def __init__(self):
        super().__init__()
        self.getrandbits_calls = []
        print("init", self)

    def getrandbits(self, n):
        self.getrandbits_calls.append(n)
        return 0

rng = MyRandom()
print(rng.randbytes != random.Random.randbytes)
print()
rng2 = pickle.loads(pickle.dumps(rng))
print(rng2.randbytes != random.Random.randbytes)

Output:

init <__main__.MyRandom object at 0xefaff0>
True

init <__main__.MyRandom object at 0xf2f320>
True

Two MyRandom instances are created, but MyRandom.randbytes is only set once by Random.__init_subclass(MyRandom): when the class is created.

I meant pickling the method.

pickle.dumps(MyRandom.randbytes)

I meant pickling the method.
pickle.dumps(MyRandom.randbytes)

Alright, that's broken because you asked me to ensure that Random._randbytes_getrandbits is called "randbytes". I reverted this change, now it works again. Example:

import random
import pickle

class MyRandom(random.Random):
    def __init__(self):
        super().__init__()

    def getrandbits(self, n):
        return 0

rng = MyRandom()
print(MyRandom.randbytes is random.Random._randbytes_getrandbits)
randbytes2 = pickle.loads(pickle.dumps(MyRandom.randbytes))
print(randbytes2 is random.Random._randbytes_getrandbits)

Output:

True
True

I think we can have this cake and eat it. Define randbytes() in __init_subclass__() and set its __module__ and __qualname__.

@serhiy-storchaka:

If it does not work it may be better to use the second option for the randbytes implementation.

Which second option? Remove the C implementation and always implement randbytes() using getrandbits() in Lib/random.py?

Well, I proposed a solution to keep the optimization (C implementation) and don't require subclasses to have to override a 6th method (randbytes).

I'm not super excited by adding more special cases __init_subclass__() (this PR). I prefer PR #19631 (random.BaseRandom) approach which uses more "regular Python code".

I let @rhettinger and you decide which approach is the best.

When I read https://bugs.python.org/issue40346 it's not even clear if we support subclassing or not, nor if performance matters or not. I'm now very confused :-( Now at least, we have multiple choices, and I let you decide which one is the best ;-)

The second option which I mentioned in my comment #19700 (comment) and repeated in #19700 (comment).

create an internal randbytes function in init_subclass and set its qualname (which depends on the class name) manually.

Well, it seems like there are different ways to implement this change. But I'm not sure that this approach is the best. For now, I will leave the PR as it is.

I'm now waiting for a feedback on https://bugs.python.org/issue40346#msg367202

I abandon my BaseRandom PR #19631 and this PR since there is no clear consensus on these changes: see https://bugs.python.org/issue40346

I wrote PR #19797 instead: "Remove C implementation of Random.randbytes()".