I'm not sure __sizeof__ is implemented correctly:

>>> from decimal import Decimal
>>> import sys
>>> d = Decimal(123456789123456798123456789123456798123456789123456798)
>>> d
Decimal('123456789123456798123456789123456798123456789123456798')
>>> sys.getsizeof(d)
24

... looks too small.

It isn't implemented at all. The Python version also always returns 96,
irrespective of the coefficient length. Well, arguably the coefficient
is a separate object in the Python version:

96
>>> sys.getsizeof(d._int)
212

For the C version I'll do the same as in longobject.c.

In full:

>>> d = Decimal(100000000000000000000000000000000000000000000000000000000000000000000)                
>>> sys.getsizeof(d)                                                                                  
96                                                                                                    
>>> sys.getsizeof(d._int)                                                                             
212

There are really two options:

a) if an object is a container, and the contained is accessible to reflection (preferably through gc.get_referents), then the container shouldn't account for the size of the contained.
b) if the contained is not accessible (except for sys.get_objects() in a debug build), then the container should provide the total sum.

A memory debugger is supposed to find all objects (e.g. through gc.get_objects, and gc.get_referents), eliminate duplicate references, and then apply sys.getsizeof for each object. This should then not leave out any memory, and not count any memory twice.

In the C version of decimal, do distinct Decimal objects ever share coefficients?  (This would be an obvious optimization for methods like Decimal.copy_negate;  I don't know whether the C version applies such optimizations.)  If there's potential for shared coefficients, that might make the "not count any memory twice" part tricky.

> In the C version of decimal, do distinct Decimal objects ever share  
> coefficients?  (This would be an obvious optimization for methods  
> like Decimal.copy_negate;  I don't know whether the C version  
> applies such optimizations.)  If there's potential for shared  
> coefficients, that might make the "not count any memory twice" part  
> tricky.

I know of three strategies to deal with such a case:
a) expose the inner objects, preferably through tp_traverse, and
    don't account for them in the container,
b) find a "canonical" owner of the contained objects, and only
    account for them along with the canonical container.
c) compute the number N of shared owners, and divide the object
    size by N. Due to rounding, this may be somewhat incorrect.

Mark Dickinson <report@bugs.python.org> wrote:
> In the C version of decimal, do distinct Decimal objects ever share coefficients?

The coefficients are members of the mpd_t struct (libmpdec data type),
and they are not exposed as Python objects or shared.

Cache locality is incredibly important: I have a patch that reserves
a static coefficient of four words inside the PyDecObject. This patch
speeds up _decimal by roughly another 30-40% for regularly sized decimals.

If the decimal grows beyond that, libmpdec automatically switches to
a dynamically allocated coefficient.

I think sharing would probably slow things down a bit.

> and they are not exposed as Python objects or shared.

Okay, thanks.  Sounds like this isn't an issue at the moment then.

+1 for having getsizeof report the total size used.

New changeset 010aa5d955ac by Stefan Krah in branch 'default':
Issue #14520: Add __sizeof__() method to the Decimal object.
http://hg.python.org/cpython/rev/010aa5d955ac

Thanks for the explanations. The new __sizeof__() method should now
report the exact memory usage.