st.integers() prefers values, closer to min_value for bounded integers on big enough interval

[skirpichev]

Simple example, here is unbounded case:

>>> import hypothesis.strategies as st
>>> n    = 5000
... ints = st.integers()
... cnt  = 0
... for _ in range(n):
...     if ints.example() < 0:
...         cnt += 1
... print(f"{cnt/n:.0%}")
... 
46%

Here is a small interval (bound are symmetrical):

>>> bits = 16
... mval = (1<<bits) - 1
... ints = st.integers(min_value=-mval, max_value=mval)
... cnt  = 0
... for _ in range(n):
...     if ints.example() < 0:
...         cnt += 1
... print(f"{cnt/n:.0%}")
... 
46%

But what happens with bigger bounds?

Distribution looks asymmetrical:

>>> bits = 32
... mval = (1<<bits) - 1
... ints = st.integers(min_value=-mval, max_value=mval)
... cnt  = 0
... for _ in range(n):
...     if ints.example() < 0:
...         cnt += 1
... print(f"{cnt/n:.0%}")
... 
79%
>>> bits = 64
... mval = (1<<bits) - 1
... ints = st.integers(min_value=-mval, max_value=mval)
... cnt  = 0
... for _ in range(n):
...     if ints.example() < 0:
...         cnt += 1
... print(f"{cnt/n:.0%}")
... 
81%

I'm giving examples with, huh, examples(), but same happens for regular tests with pytest&hypothesis.

[Liam-DeVoe]

Thanks Sergey.

I suspect the problem is here; specifically upper = min(upper, lower + 2**cap_bits - 1).

We should work on some visual-only distribution tests as part of our test suite!

[skirpichev]

We should work on some visual-only distribution tests as part of our test suite!

I noted that same happens with regular tests. That was the problem for me. But examples with .examples() show same picture.