Assigning to an instance's __class__ attribute in Python

Question

Under what circumstances is it possible, and when is it impossible, to assign to an instance's __class__ attribute in Python?

Error messages such as TypeError: __class__ assignment: only for heap types don't really do it for me.

• Eli Bendersky: @Rafe: actually, Django uses this trick in its authentication code
• tacaswell: late to the party, but matplotlib does this in production code too (in some of the 3D plotting).

Answer 1

You can only assign to the __class__ attribute of an instance of a user-defined class (i.e. defined using the class keyword), and the new value must also be a user-defined class. Whether the classes are new-style or old-style does not matter. (You can't mix them, though. You can't turn an old-style class instance into a new-style class instance.) See also this issue in the Python bug tracker, which also complains that the error message is somewhat hard to understand.

Just to add what Rafe said in the above comment: Never do this in production.

• user395760: Also, from my experimenting it seems that the presence/absence of __slots__ also makes a difference: "X deallocator differs from Y" from instance_with_slots.__class__ = ClassWithoutSlots and instance_without_slots.__class__ = ClassWithSlots
• porgarmingduod: Ah, I was confused because did a.__class__ = B.__class__ when I obviously should have been doing a.__class__ = B
• Luke Taylor: Both matplotlib and django do this in production, according to the above comments
• jedwards: @dtheodor what a bold and frankly silly statement. Yes, Python chooses to expose some of its quite advanced and easy-to-misuse machinery, but it is a choice: __class__ could have been made a read-only attribute. Saying that these mature and exceptionally well supported projects are "doing it wrong" because they do something that we advise newcomers against on stackoverflow is like saying that engineers are doing it wrong because they're not using sliderules, or space shuttle welders doing it wrong because they're not using soldering irons.
• Sven Marnach: @jedwards Not everything that language designers allow to do is actually a good idea, even if popular projects do it. I don't have the time to write a detailed analysis, but after looking at the code it is my conviction that both Django and Matplotlib would be better off not using this hack. It's too much magic for virtually no gain.