Should delayed objects modify instances of classes?

from dask import delayed
from nose.tools import assert_true

class MyClass(dict):
    pass

@delayed(pure=True)
def foo_dict(arg):
    assert_true(isinstance(arg, dict))

@delayed(pure=True)
def foo_myclass(arg):
    assert_true(isinstance(arg, MyClass))


# works fine
res = foo_dict(MyClass())
res.compute()
# raises an error
res = foo_myclass(MyClass())
res.compute()

This returns an error because when passing through delayed, the instance goes from a MyClass instance to dict.

Is this intended? For my use cases, it doesn’t bother me much, but I want to better understand the inner workings of the delayed object, and what assumptions I can make. If I replaced delayed with a curry instance, this works (of course, now don’t do the last compute step). I’m not sure if this helps narrow things down.

Thanks for the great library!

[martindurant]

I have checked, and within the function foo, when running in a worker process, arg does have the type MyClass, but id(type(arg)) and id(MyClass) do not match: i.e., presumably deserialization has created two identical class definitions. If you pass the class, it works fine:

    @delayed(pure=True)
    def foo(arg, cls):
        assert_true(isinstance(arg, cls))

res = foo(MyClass(), MyClass)

[jrmlhermitte]

ah yes thanks, I should have seen that. I have noticed that issue in the past. A class defined in the same file code is run is not the same as a class imported (I don’t understand the internals why, but it makes intuitive sense for a namespace safe system).

thanks again