python中的pickling方法描述符对象

Question

I am trying to pickle a method_descriptor .

Pickling with pickle or cloudpickle fails:

Python 2.7.10 |Continuum Analytics, Inc.| (default, Oct 19 2015, 18:04:42) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
Anaconda is brought to you by Continuum Analytics.
Please check out: http://continuum.io/thanks and https://anaconda.org

>>> import pickle, cloudpickle

>>> pickle.dumps(set.union)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 1374, in dumps
    Pickler(file, protocol).dump(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 224, in dump
    self.save(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 306, in save
    rv = reduce(self.proto)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/copy_reg.py", line 70, in _reduce_ex
    raise TypeError, "can't pickle %s objects" % base.__name__
TypeError: cannot pickle method_descriptor objects

>>> cloudpickle.dumps(set.union)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/site-packages/cloudpickle/cloudpickle.py", line 602, in dumps
    cp.dump(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/site-packages/cloudpickle/cloudpickle.py", line 111, in dump
    raise pickle.PicklingError(msg)
pickle.PicklingError: Could not pickle object as excessively deep recursion required.

Importing dill somehow makes pickle work, as shown below:

>>> import dill
>>> pickle.dumps(set.union)
'cdill.dill\n_getattr\np0\n(c__builtin__\nset\np1\nS\'union\'\np2\nS"<method \'union\' of \'set\' objects>"\np3\ntp4\nRp5\n.'

>>> f = pickle.loads(pickle.dumps(set.union))
>>> set.union(set([1,2]), set([3]))
set([1, 2, 3])
>>> f(set([1,2]), set([3]))
set([1, 2, 3])

The issue in cloudpickle remains even after the dill import:

>>> cloudpickle.dumps(set.union)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/site-packages/cloudpickle/cloudpickle.py", line 602, in dumps
    cp.dump(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/site-packages/cloudpickle/cloudpickle.py", line 111, in dump
    raise pickle.PicklingError(msg)
pickle.PicklingError: Could not pickle object as excessively deep recursion required.

In my application I rely on cloudpickle to handle functions with globals. So my question is, how can I get cloudpickle to work for method_descriptor objects in Python 2.7?

EDIT: I noticed that the same issue occurs in Python 3.3, but is not present in Python 3.5.

Answer 1

I'm the dill author. When you do import dill , it injects the serialization registry from dill into pickle (basically, puts all the copy_reg -type knowledge from dill into the pickle registry).

>>> import pickle
>>> pickle.Pickler.dispatch
{<type 'function'>: <function save_global at 0x105d0c7d0>, <type 'dict'>: <function save_dict at 0x105d0c668>, <type 'int'>: <function save_int at 0x105d0c230>, <type 'long'>: <function save_long at 0x105d0c2a8>, <type 'list'>: <function save_list at 0x105d0c578>, <type 'str'>: <function save_string at 0x105d0c398>, <type 'unicode'>: <function save_unicode at 0x105d0c410>, <type 'instance'>: <function save_inst at 0x105d0c758>, <type 'type'>: <function save_global at 0x105d0c7d0>, <type 'NoneType'>: <function save_none at 0x105d0c140>, <type 'bool'>: <function save_bool at 0x105d0c1b8>, <type 'tuple'>: <function save_tuple at 0x105d0c488>, <type 'float'>: <function save_float at 0x105d0c320>, <type 'classobj'>: <function save_global at 0x105d0c7d0>, <type 'builtin_function_or_method'>: <function save_global at 0x105d0c7d0>}
>>> import dill
>>> pickle.Pickler.dispatch
{<class '_pyio.BufferedReader'>: <function save_file at 0x106c8b848>, <class '_pyio.TextIOWrapper'>: <function save_file at 0x106c8b848>, <type 'operator.itemgetter'>: <function save_itemgetter at 0x106c8b578>, <type 'weakproxy'>: <function save_weakproxy at 0x106c8c050>, <type 'NoneType'>: <function save_none at 0x105d0c140>, <type 'str'>: <function save_string at 0x105d0c398>, <type 'file'>: <function save_file at 0x106c8b8c0>, <type 'classmethod'>: <function save_classmethod at 0x106c8c230>, <type 'float'>: <function save_float at 0x105d0c320>, <type 'instancemethod'>: <function save_instancemethod0 at 0x106c8ba28>, <type 'cell'>: <function save_cell at 0x106c8bb18>, <type 'member_descriptor'>: <function save_wrapper_descriptor at 0x106c8bc08>, <type 'slice'>: <function save_slice at 0x106c8bc80>, <type 'dict'>: <function save_module_dict at 0x106c8b410>, <type 'long'>: <function save_long at 0x105d0c2a8>, <type 'code'>: <function save_code at 0x106c8b320>, <type 'type'>: <function save_type at 0x106c8c0c8>, <type 'xrange'>: <function save_singleton at 0x106c8bde8>, <type 'builtin_function_or_method'>: <function save_builtin_method at 0x106c8b9b0>, <type 'classobj'>: <function save_classobj at 0x106c8b488>, <type 'weakref'>: <function save_weakref at 0x106c8bed8>, <type 'getset_descriptor'>: <function save_wrapper_descriptor at 0x106c8bc08>, <type 'weakcallableproxy'>: <function save_weakproxy at 0x106c8c050>, <class '_pyio.BufferedRandom'>: <function save_file at 0x106c8b848>, <type 'int'>: <function save_int at 0x105d0c230>, <type 'list'>: <function save_list at 0x105d0c578>, <type 'functools.partial'>: <function save_functor at 0x106c8b7d0>, <type 'bool'>: <function save_bool at 0x105d0c1b8>, <type 'function'>: <function save_function at 0x106c8b398>, <type 'thread.lock'>: <function save_lock at 0x106c8b500>, <type 'super'>: <function save_functor at 0x106c8b938>, <type 'staticmethod'>: <function save_classmethod at 0x106c8c230>, <type 'module'>: <function save_module at 0x106c8bf50>, <type 'method_descriptor'>: <function save_wrapper_descriptor at 0x106c8bc08>, <type 'operator.attrgetter'>: <function save_attrgetter at 0x106c8b5f0>, <type 'wrapper_descriptor'>: <function save_wrapper_descriptor at 0x106c8bc08>, <type 'numpy.ufunc'>: <function save_numpy_ufunc at 0x106c8bcf8>, <type 'method-wrapper'>: <function save_instancemethod at 0x106c8baa0>, <type 'instance'>: <function save_inst at 0x105d0c758>, <type 'cStringIO.StringI'>: <function save_stringi at 0x106c8b6e0>, <type 'unicode'>: <function save_unicode at 0x105d0c410>, <class '_pyio.BufferedWriter'>: <function save_file at 0x106c8b848>, <type 'property'>: <function save_property at 0x106c8c140>, <type 'ellipsis'>: <function save_singleton at 0x106c8bde8>, <type 'tuple'>: <function save_tuple at 0x105d0c488>, <type 'cStringIO.StringO'>: <function save_stringo at 0x106c8b758>, <type 'NotImplementedType'>: <function save_singleton at 0x106c8bde8>, <type 'dictproxy'>: <function save_dictproxy at 0x106c8bb90>}

cloudpickle has (slightly) different pickling functions than dill , and if you are using cloudpickle , it push it's own serialization functions into the pickle registry. If you want to get cloudpickle to work for you, you might be able to monkeypatch a solution… essentially install a module within your application that does import dill as cloudpickle (Nice reference: http://blog.dscpl.com.au/2015/03/safely-applying-monkey-patches-in-python.html )… but that would replace the entire use of cloudpickle with dill in your application context. You could also try a monkeypatch along these lines:

>>> #first import dill, which populates itself into pickle's dispatch
>>> import dill
>>> import pickle
>>> # save the MethodDescriptorType from dill
>>> MethodDescriptorType = type(type.__dict__['mro'])
>>> MethodDescriptorWrapper = pickle.Pickler.dispatch[MethodDescriptorType]
>>> # cloudpickle does the same, so let it update the dispatch table
>>> import cloudpickle
>>> # now, put the saved MethodDescriptorType back in
>>> pickle.Pickler.dispatch[MethodDescriptorWrapperType] = MethodDescriptorWrapper

Note that if you are going to use cloudpickle.dumps directly, you'd have to overload the registry in cloudpickle directly by doing the above monkeypatch on cloudpickle.CloudPickler.dispatch .

I don't guarantee that it will work, nor do I guarantee that it won't screw up other objects from cloudpickle (essentially, I haven't tried it), but it's a potential route to replacing the offending cloudpickle wrapper with the one from dill .

If you want the short answer, I'd say (at least for this case) use dill . ;)

---

EDIT with regard to copyreg :

Here's what's in dill :

def _getattr(objclass, name, repr_str):
    # hack to grab the reference directly
    try:
        attr = repr_str.split("'")[3]
        return eval(attr+'.__dict__["'+name+'"]')
    except:
        attr = getattr(objclass,name)
        if name == '__dict__':
            attr = attr[name]
        return attr

Which is used to register a function with a lower-level reduce function (directly on the pickler instance). obj is the object to pickle.

pickler.save_reduce(_getattr, (obj.__objclass__, obj.__name__, obj.__repr__()), obj=obj)

I believe this translates to a reduce method (used directly in copyreg.pickle ) like this:

def _reduce_method_descriptor(obj):
    return _getattr, (obj.__objclass__, obj.__name__, obj.__repr__())

Answer 2

After much messing around I think I have found a clean answer to this question that works with Python 2.7 and 3.3. Note that Python 3.5 has no issues to begin with.

Before I show the results of my findings I want to credit the multiprocessing.forking module, which was where I got the jist of the code that makes this work.

In what follows, I will use set.union as an example of <class 'method_descriptor'> .

Python 3.5: method descriptors work out of the box

Python 3.5.0 |Continuum Analytics, Inc.| (default, Oct 19 2015, 21:57:25) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import pickle
>>> pickle.dumps(set.union)
b'\x80\x03cbuiltins\ngetattr\nq\x00cbuiltins\nset\nq\x01X\x05\x00\x00\x00unionq\x02\x86q\x03Rq\x04.'
>>> f = pickle.loads(pickle.dumps(set.union))
>>> f({1, 2, 3}, {5})
{1, 2, 3, 5}
>>> 

Python 3.3: use copyreg to provide a way for pickle to work with method_descriptor

Python 3.3.5 |Continuum Analytics, Inc.| (default, Jun  4 2015, 15:22:11) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import pickle
>>> pickle.dumps(set.union)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
_pickle.PicklingError: Can't pickle <class 'method_descriptor'>: attribute lookup builtins.method_descriptor failed

The type of set.union is method_descriptor :

>>> type(set.union)
<class 'method_descriptor'>

We define the reduce function for the method_descriptor and register it with copyreg :

>>> def _reduce_method_descriptor(m):
...     return getattr, (m.__objclass__, m.__name__)
... 
>>> import copyreg
>>> copyreg.pickle(type(set.union), _reduce_method_descriptor)

Success:

>>> pickle.dumps(set.union)
b'\x80\x03cbuiltins\ngetattr\nq\x00cbuiltins\nset\nq\x01X\x05\x00\x00\x00unionq\x02\x86q\x03Rq\x04.'
>>> f = pickle.loads(pickle.dumps(set.union))
>>> f({1, 2, 3}, {5})
{1, 2, 3, 5}

And if we now import cloudpickle the registered pickling function still works:

>>> import cloudpickle
>>> cloudpickle.dumps(set.union)
b'\x80\x02c__builtin__\ngetattr\nq\x00c__builtin__\nset\nq\x01X\x05\x00\x00\x00unionq\x02\x86q\x03Rq\x04.'
>>> f = pickle.loads(pickle.dumps(set.union))
>>> f({1, 2, 3}, {5})
{1, 2, 3, 5}
>>> 

Python 2.7: use copy_reg to provide a way for pickle to work with method_descriptor

In Python 2.7, the module for registering pickle support functions is called copy_reg .

Python 2.7.10 |Continuum Analytics, Inc.| (default, Oct 19 2015, 18:04:42) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
Anaconda is brought to you by Continuum Analytics.
Please check out: http://continuum.io/thanks and https://anaconda.org
>>> import pickle
>>> pickle.dumps(set.union)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 1374, in dumps
    Pickler(file, protocol).dump(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 224, in dump
    self.save(obj)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/pickle.py", line 306, in save
    rv = reduce(self.proto)
  File "/home/pmd/anaconda3/envs/python2/lib/python2.7/copy_reg.py", line 70, in _reduce_ex
    raise TypeError, "can't pickle %s objects" % base.__name__
TypeError: can't pickle method_descriptor objects

The type of set.union is method_descriptor :

>>> type(set.union)
<type 'method_descriptor'>

We define the reduce function for the method_descriptor and register it with copyreg :

>>> def _reduce_method_descriptor(m):
...     return getattr, (m.__objclass__, m.__name__)
... 
>>> import copy_reg
>>> copy_reg.pickle(type(set.union), _reduce_method_descriptor)
>>> pickle.dumps(set.union)
"c__builtin__\ngetattr\np0\n(c__builtin__\nset\np1\nS'union'\np2\ntp3\nRp4\n."

Success:

>>> f = pickle.loads(pickle.dumps(set.union))
>>> f(set([1, 2, 3]), set([5]))

And works with cloudpickle too:

set([1, 2, 3, 5])
>>> import cloudpickle
>>> cloudpickle.dumps(set.union)
'\x80\x02c__builtin__\ngetattr\nq\x00c__builtin__\nset\nq\x01U\x05unionq\x02\x86q\x03Rq\x04.'
>>> f = pickle.loads(cloudpickle.dumps(set.union))
>>> f(set([1, 2, 3]), set([5]))
set([1, 2, 3, 5])
>>>