Calling st.register_type_strategy imports numpy

[rsokl]

This is the "offending" block:

hypothesis/hypothesis-python/src/hypothesis/strategies/_internal/types.py

Lines 493 to 501 in 65b4b21

	try: # pragma: no cover
	import numpy as np
	from hypothesis.extra.numpy import array_dtypes, array_shapes, arrays, scalar_dtypes
	_global_type_lookup[np.dtype] = array_dtypes()
	_global_type_lookup[np.ndarray] = arrays(scalar_dtypes(), array_shapes(max_dims=2))
	except ImportError:
	pass

This test

hypothesis/hypothesis-python/tests/numpy/test_import.py

Lines 16 to 29 in 65b4b21

	SHOULD_NOT_IMPORT_NUMPY = """
	import sys
	from hypothesis import given, strategies as st
	@given(st.integers() | st.floats() | st.sampled_from(["a", "b"]))
	def test_no_numpy_import(x):
	assert "numpy" not in sys.modules
	"""
	def test_hypothesis_is_not_the_first_to_import_numpy(testdir):
	# We only import numpy if the user did so first.
	result = testdir.runpytest(testdir.makepyfile(SHOULD_NOT_IMPORT_NUMPY))
	result.assert_outcomes(passed=1, failed=0)

can be made to fail by modifying the src-string to be:

SHOULD_NOT_IMPORT_NUMPY = """
import sys
from hypothesis import given, strategies as st

@given(st.integers() | st.floats() | st.sampled_from(["a", "b"]) | st.just(st.register_type_strategy(int, st.integers())))
def test_no_numpy_import(x):
    assert "numpy" not in sys.modules
"""

This was causing a headache for me when figuring out why numpy was being imported the instant I ran pytest... this was because I happened to have installed numpy + hypothesis + pydantic (which uses st.register_type_strategy via our entrypoint).

This could be a big headache for testers over at, say, PyTorch, who want to test that they don't cause numpy to be imported.

[Zac-HD]

If I expected this to require a nasty patch - i.e. worse than our numpy special case in register_random - I wouldn't consider it worth fixing.

That's because once we make our pytest plugin lazily load Hypothesis (#3140), this will only trigger if you deliberately import hypothesis and somehow call st.register_type_strategy() in the test that Numpy is not imported, which usually happens in a fresh process.

Other libraries will never get this special treatment, but Numpy is already exceptional in our code and the ecosystem.