[Python-Dev] Class decorators can't be pickled, which breaks multiprocessing and concurrent.futures. Any plans of improving this?

[Serhiy Storchaka]

11.08.18 23:08, Santiago Basulto пише:
> Hello folks! I'm using the `concurrent.futures.ProcessPoolExecutor` with 
> a couple of functions that have been decorated with a class decorator. 
> Both `concurrent.futures` and `multiprocessing` breaks because "the 
> object's can't be pickled". There's a really simple fix for this, which 
> is just, instead of "decorating" the function (with the @), instantiate 
> the decorator and use it directly.
> 
> Example. This is my (very simple, for demonstration purposes) decorator:
> 
>      class CheckOnlyIntegers:
>          def __init__(self, fn):
>              self.fn = fn
> 
>          def __call__(self, *args):
>              if not all([type(arg) == int for arg in args]):
>                  raise ValueError("Invalid param is not an integer")
>              return self.fn(*args)
> 
> If I define a simple `add` function and decorate it using the 
> `CheckOnlyIntegers` decorator:
> 
>      @CheckOnlyIntegers
>      def add(x, y):
>          return x + y
> 
> and try using a regular `ProcessPoolExecutor().submit(add, 2, 3)`, it 
> fails with:
> 
> ```
> Can't pickle <function add at 0x10ace3e18>: it's not the same object as 
> __main__.add.
> ```

By default instances of Python classes are pickled by pickling a set of 
their attributes. Functions are pickled by name. But since the name of 
self.fn corresponds to the decorated function, not the function itself, 
it can't be pickled.

You can implement the explicit pickle support for your decorator that 
bypass this limitation.

     def __reduce__(self):
         return self.fn.__qualname__

Now the decorated function will be pickled by name.

It may help to set also self.__module__ = self.fn.__module__ in the 
constructor.