Wrong data generation using dataclasses, Sequence and from_type

[mvaled]

Hypothesis from_type sometimes generates bytes when the dataclass has an attribute of type Sequence[X]:

>>> from dataclasses import dataclass
>>> from typing import Sequence
>>> @dataclass
... class Item:
...     name: str
...     

>>> @dataclass
... class Row:
...     name: str
...     items: Sequence[Item]
...     

>>> strategies.from_type(Row).example()   # <--- Wrong data type
Row(name='', items=b'\x00')

>>> strategies.from_type(Row).example()
Row(name='\x13', items=[])

>>> strategies.from_type(Row).example()
Row(name='', items=[Item(name='')])

[mvaled]

I'm using hypothesis 4.50.6. But the issue is reproducible in master, commit 89d494d

[mvaled]

The issue seems to be in the implementation of from_typing_thing:

>>> from_typing_type(Sequence[Item])                                                                                                                                                          
one_of(binary(), lists(elements=builds(Item)))

which includes binary().

[mvaled]

@Zac-HD

I found the underlying cause is that:

>>> try_issubclass(typing.ByteString, Sequence[Item])                                                                                                                                         
True

[mvaled]

I think typing.ByteString should not be a subclass of Sequence. There are other things to consider:

>>> from_typing_type(Sequence)                                                                                                                                                                
one_of(binary(), lists(elements=text()))

One may argue that if binary() is valid, so is text(), and the result should be one_of(binary(), text(), lists(elements=text())).

---

In Python we don't have a type for a single char and or byte; so it's difficult to express Sequence[Byte] or Sequence[Char]; and make ByteString a subclass of the first and str and subclass of the last.

[mvaled]

@Zac-HD

I've created a PR #2258 but I feel it deserves a tests. However, I'm not sure if it's ok to depends of dataclasses for tests.