st.integers(min_value=0, max_value=1e100) almost only returns large numbers

[ichumuh]

hypothesis.__version__
Out[70]: '3.66.0'
a = [st.integers(min_value=0, max_value=1e100).example() for x in range(999)]
len([x for x in a if x >= 1e97])
Out[72]: 997
len([x for x in a if x < 1e97])
Out[73]: 2
len([x for x in a if x < 1e94])
Out[74]: 0

The same happens with floats(). As a workaround I'm currently using filter, which gives a better distribution

a = [st.integers().filter(lambda x: x > 0 and x < 1e100).example() for x in range(999)]
len([x for x in a if x >= 1e97])
Out[75]: 0
len([x for x in a if x < 1e97])
Out[76]: 999

I'm on Ubuntu 16.04 using python 2.7.

[Zac-HD]

First and most generally, .example() gives a highly biased distribution - it's designed to show illustrative rather than representative examples.

That said, considering this alongside #909 and #1212 makes me thing that we may well have some issues with data variety from such strategies. (though note again that it's not specified what will be generated, just that some things could be generated and that we try to find bugs)

[ichumuh]

Thanks for your reply. I was just trying to make the example as simple as possible. I first noticed the issue when a test case with @given(st.floats(min_value=-1e100, max_value=1e100)) didn't find a bug in one of my functions that only occurred with numbers relatively close to 0. Whereas the filter alternative did.

[Zac-HD]

I've just been looking at integers() for #1616, and the root cause is that we delegate from BoundedIntStrategy directly down to cu.integer_range which is basically a uniform distribution. Instead, we probably want to check if the range is large (e.g. >16-bit) and if so use something more like the psudeo-geometric approach from WideRangeIntStrategy.