__sizeof__ is not mentioned on page Doc/reference/datamodel.rst together with other special methods. There is no __sizeof__ in index.

Could you possibly write a patch for 2.7, 3.2 and 3.3?

> Could you possibly write a patch for 2.7, 3.2 and 3.3?

No, not for documentation.

is __sizeof__ part of the language or just an implementation detail?

See issue2898.

It is part of runtime services and is mentioned in the documentation for 'sys.getsizeof' [1]:

"""
getsizeof() calls the object’s __sizeof__ method and adds an additional garbage collector overhead if the object is managed by the garbage collector.
"""

[1] http://docs.python.org/library/sys.html?highlight=__sizeof__#sys.getsizeof.

> It is part of runtime services and is mentioned in the documentation for
> 'sys.getsizeof' [1]:

Yes, but it is not mentioned in the index.

issue2898 smells like an implementation detail to me. Also, sys.getsizeof() isn't necessarily authoritative here, since the sys module is full of all sorts of implementation details.

I'm not saying __sizeof__ shouldn't be made an official part of the language, but we should be careful before we effectively etch something in stone for all Python implementations.  It may be worth getting feedback from python-dev, particularly from the major Python implementers.

This may be CPython specific.

Also, sys.getsizeof() hasn't proved to be very useful (I never see anyone use it or see it in production code).  Worse, it doesn't even make sense in the case of shared-key dictionaries and whatnot.  Accordingly, there is little reason to make this more public than it already is.  I would be happy to see it written-off as an implementation detail.

sys.getsizeof() is not for production code, it's for experimenting and getting information at the command prompt. It's quite useful when you know its limitations.

I'd argue that implementing __sizeof__ for user types doesn't make much sense (except C extension types, which are CPython-specific).

rhettinger: users frequently need sys.getsizeof. See, for example,

http://stackoverflow.com/questions/1331471/in-memory-size-of-python-stucture
http://stackoverflow.com/questions/449560/how-do-i-determine-the-size-of-an-object-in-python
http://stackoverflow.com/questions/2117255/python-deep-getsizeof-list-with-contents
http://stackoverflow.com/questions/11301295/measure-object-size-accurately-in-python-sys-getsizeof-not-functioning
[and so on]

In what cases, do you think, it generates meaningless results? I find the results for the shared-keys dictionary quite meaningful and natural.

sys.getsizeof is certainly CPython-specific. However, __sizeof__ is not just an implementation detail of sys.getsizeof, just as __len__ is not an implementation detail of len(). Authors of extension types are supposed to implement it if object.__sizeof__ is incorrect.

> sys.getsizeof is certainly CPython-specific. However, __sizeof__ is not
> just an implementation detail of sys.getsizeof, just as __len__ is not an
> implementation detail of len(). Authors of extension types are supposed to
> implement it if object.__sizeof__ is incorrect.

I think there is one difference between __len__ and __sizeof__. __sizeof__ 
should be overloaded only for C-implemented classes. IMHO, it is a part of C 
API.

> I think there is one difference between __len__ and __sizeof__. __sizeof__ 
> should be overloaded only for C-implemented classes. IMHO, it is a part of C 
> API.

That is a reasonable point. So documenting it along with the type slots
might be best.