Using -OO on a script will remove the __doc__ attributes but the docstrings will still be in the process memory. In the attachments is an example script which demonstrates this with a docstring of ~10 MiB (opening the file in an editor can need some time). Calling "python3 -OO test.py" will result in a memory usage of ~16 MiB on my system (Linux 64 Bit) while test.__doc__ is None.

Do realize this is a one-time memory cost, though, because next execution will load from the .pyo and thus will never load the docstring into memory. If you pre-compile all bytecode with -OO this will never even occur.

> Do realize this is a one-time memory cost, though, because next execution will load from the .pyo and thus will never load the docstring into memory.

Except in 2 cases:

- The bytecode was previously generated with -O.
- The bytecode couldn't be written (for example permission issues or Python was invoked with -B).

So the question is, if there is no longer a reference to the docstring, why isn't it garbage collected?  (I tested adding a gc.collect(), and it didn't make any difference.)

R. David Murray <report@bugs.python.org> wrote:
> So the question is, if there is no longer a reference to the docstring, why isn't it garbage collected?  (I tested adding a gc.collect(), and it didn't make any difference.)

I think it probably is garbage collected but the freed memory is not returned to the OS
by the memory allocator.

Hmm.  If I turn on gc debugging before the def, I don't see anything get collected.  If I allocate a series of new 10K strings, the memory keeps growing.  Of course, that could still be down to the vagaries of OS memory management.  Time to break out Victor's tracemalloc, but I probably don't have that much ambition today :)

It looks like the memory management is based directly on Py_Arenas:

def f():
    """squeamish ossifrage"""
    pass

Breakpoint 1, PyArena_Free (arena=0x9a5120) at Python/pyarena.c:159
159         assert(arena);
(gdb) p arena->a_objects
$1 = ['f', 'squeamish ossifrage']
(gdb) bt
#0  PyArena_Free (arena=0x9a5120) at Python/pyarena.c:159
#1  0x0000000000425af5 in PyRun_FileExFlags (fp=0xa1b780, filename_str=0x7ffff7f37eb0 "docstr.py", start=257, globals=
    {'f': <function at remote 0x7ffff7f04058>, '__builtins__': <module at remote 0x7ffff7f6a358>, '__name__': '__main__', '__file__': 'docstr.py', '__package__': None, '__loader__': <SourceFileLoader(name='__main__', path='docstr.py') at remote 0x7ffff7ede608>, '__cached__': None, '__doc__': None}, locals=
    {'f': <function at remote 0x7ffff7f04058>, '__builtins__': <module at remote 0x7ffff7f6a358>, '__name__': '__main__', '__file__': 'docstr.py', '__package__': None, '__loader__': <SourceFileLoader(name='__main__', path='docstr.py') at remote 0x7ffff7ede608>, '__cached__': None, '__doc__': None}, closeit=1, flags=0x7fffffffe490) at Python/pythonrun.c:2114
#2  0x0000000000423a0c in PyRun_SimpleFileExFlags (fp=0xa1b780, filename=0x7ffff7f37eb0 "docstr.py", closeit=1, flags=
    0x7fffffffe490) at Python/pythonrun.c:1589
#3  0x000000000042289c in PyRun_AnyFileExFlags (fp=0xa1b780, filename=0x7ffff7f37eb0 "docstr.py", closeit=1, flags=0x7fffffffe490)
    at Python/pythonrun.c:1276
#4  0x000000000043bc83 in run_file (fp=0xa1b780, filename=0x9669b0 L"docstr.py", p_cf=0x7fffffffe490) at Modules/main.c:336
#5  0x000000000043c8c5 in Py_Main (argc=3, argv=0x964020) at Modules/main.c:780
#6  0x000000000041cdb5 in main (argc=3, argv=0x7fffffffe688) at ./Modules/python.c:69

So the string 'squeamish ossifrage' is still in arena->a_objects right until
end of PyRun_FileExFlags(), even with -OO.

Do note that .pyc files now encode their optimization levels, so the only thing to potentially do here is change the compiler to toss docstrings out and make sure they are freed when they are parsed to avoid holding on to them.