• works around the failing health check discussed in Cubed xarray tests #4 (comment)

The idea is to use the less complicated form of metaprogramming in python (decorators) to delay the call to hypothesis.given. This is done by applying the function decorator delayed_given, which stores strategies and partial strategy functions passed to it in a new property on the test function.

Then when a test class is created, we apply the class decorator initialize_tests that iterates over the test methods, accesses the property we just defined, evaluates any partial strategy functions with a fixed set of parameters, passes the result to given and finally replaces the test function with its decorated version.

As an example, this would amount to something similar to:

class TestNumpyReduction(ReductionTests, NumpyTestMixin):
 ...
 
 given_args, given_kwargs = instantiate_given(
 test_variable.__hypothesis_given__,
 array_strategy_fn=array_strategy_fn,
 )
 test_variable = given(*given_args, **given_kwargs)(test_variable)
 
 ...

Assuming this works properly (I'll still have to write tests), we can define test classes like this:

class ReductionTests(DuckArrayTestMixin):
 @pytest.mark.parametrize("op", ["cumsum", "cumprod"])
 @delayed_given(partial(xrst.variables))
 def test_variable_cumulative(self, op, variable):
 ...

and they'd be used like this:

@initialize_tests
class TestReductionNumpy(ReductionTests, NumpyTestMixin):
 pass

This is obviously a bit more complex than normal code, but the end result does not look too magical to me. What do you think?

cc @TomNicholas, @Zac-HD