Why do I get unexpected behavior in Python isinstance after pickling?

Question

Putting aside whether the use of isinstance is harmful, I have run into the following conundrum when trying to evaluate isinstance after serializing/deserializing an object via Pickle:

from __future__ import with_statement
import pickle

# Simple class definition
class myclass(object):
    def __init__(self, data):
        self.data = data

# Create an instance of the class
x = myclass(100)

# Pickle the instance to a file
with open("c:\\pickletest.dat", "wb") as f:
    pickle.dump(x, f)

# Replace class with exact same definition
class myclass(object):
    def __init__(self, data):
        self.data = data

# Read an object from the pickled file
with open("c:\\pickletest.dat", "rb") as f:
    x2 = pickle.load(f)

# The class names appear to match
print x.__class__
print x2.__class__

# Uh oh, this fails...(why?)
assert isinstance(x2, x.__class__)

Can anyone shed some light on why isinstance would fail in this situation? In other words, why does Python think these objects are of two different classes? When I remove the second class definition, isinstance works fine.

Answer 1

This is how the unpickler works (site-packages/pickle.py):

def find_class(self, module, name):
    # Subclasses may override this
    __import__(module)
    mod = sys.modules[module]
    klass = getattr(mod, name)
    return klass

To find and instantiate a class.

So of course if you replace a class with an identically named class, the klass = getattr(mod, name) will return the new class, and the instance will be of the new class, and so isinstance will fail.