The existing @experimental decorator doesn't work particularly well when
trying to decorate a class. Previously, for vanilla class decorations,
things would mostly work:
@experimental class Foo: pass foo = Foo()
ExperimentalWarning: "Foo" is an experimental function.
That error is mostly correct except for the fact that Foo is supposed
to be an experimental class and not an experimental function. But
it's more than just a copy error - because the @experimental decorator
is returning an _inner function, Foo truly is of type function.
The most immediate consequence is that nothing can properly subclass
@experimental class Bar(Foo): pass
E TypeError: function() argument 'code' must be code, not str
It's a bit of a cryptic error but it's throwing because you can't
subclass a function. https://bugs.python.org/issue6829
An interesting Dagster-specific consequence is that you can't mark
anything that subclasses ConfigurableClass as @experimental:
@experimental class Configurable(ConfigurableClass): ...
E TypeError: issubclass() arg 1 must be a class
It's throwing because issubclass expects its first argument to
be a class and we're giving it a function:
One option is to take the approach that we took when introducing
@experimental_decorator - we could add a new @experimental_class and
leave it up to the author to correctly decide when to decorate with
@experimental and when to decorate with @experimental_class.
Instead, I've decided to conditionally switch on whether the callable
being decorated is a function or a class. If it's a function, we
do the same wrapping behavior we've always done. If it's a class, we
augment __init__ to also include an experimental_class_warning but
we still return a class.