See the thread starting at
http://mail.python.org/pipermail/python-dev/2006-February/061261.html

This still needs unit tests and docs.

Logged In: YES 
user_id=1188172

Observations:
* Doesn't the PyMemberDef need a sentinel?
* Is PyObject_CallObject() faster than PyEval_CallFunction()
with no arguments present?
* The current patch gives me a segfault at interpreter exit
because there is a dict object whose ma_default_factory was
not initialized to NULL.

Logged In: YES 
user_id=6380

Are you sure you did a "make clean"? Because the dictobject
struct lay-out is changed you may have to do that.

If the problem persists, try adding setting
ma_default_factory to NULL explicitly in dict_new (like it's
already done in PyDict_New) and see if that makes a difference.

BTW the change to PyDict_GetItem must be removed -- it's not
a good idea.

Logged In: YES 
user_id=1188172

Okay. I configured with "--with-pydebug" all the time, and
then comes the segfault.

Without "--with-pydebug", everything seems fine.

Logged In: YES 
user_id=6380

Aha. I'll have to try that. In the mean time, here's a new
patch:

- PyDict_GetItem is no longer involved
- added {NULL} to PyMemberDef array
- set ma_default_factory to NULL in both constructors

Still no unit tests or docs.

I have some misgivings about the API -- perhaps this should
be a subclass.

Logged In: YES 
user_id=6380

Sorry, forgot the upload. Here it is.

Logged In: YES 
user_id=6380

Here's a version that doesn't crash in debug mode.

Neal Norwitz is standing next to me and pointed out that I
was attempting to decref something after tp_free had already
wiped out the object.  D'oh!

Logged In: YES 
user_id=6380

Here's a completely new version after another round of
python-dev.

- The built-in dict type still defines and calls
on_missing(), but the default on_missing() implementation
just raises KeyError(key).  It no longer has a
default_factory attribute.

- You can subclass dict and override on_missing() to do
whatever you want.

- A useful subclass is collections.defaultdict; it defines
an attribute default_factory and its on_missing()
implementation calls that and inserts the resulting value in
the dict (previous versions of the patch had this semantics
in the built-in dict class, which was frowned upon).

- Now with unit tests.

- No docs yet, though.

Assigning to Raymond Hettinger for review.  Raymond, please
assign it back to me for checkin if you're okay with this
(or for revision if you're not :-).  Because of Google's
lawyers I must check this in myself.

Logged In: YES 
user_id=33168

In test_print() how about os.unlink(tfn)?  Why do you play
games with stdout, can't you do print >>f, ... ?  (If you
don't need stdout, it looks like you don't need sys.)

In defdict_on_missing() why do you check if the factory is
None?  It's the only place None is checked for, everywhere
else, it's just NULL.

defdict_repr should be static.

Type for this line in defdict_init() should be Py_ssize_t,
not int:
int n = PyTuple_GET_SIZE(args);

Logged In: YES 
user_id=80475

Neal, the defdict_on_missing check for Py_None is 
necessary because the user can assign None to the factory 
and expect it to have the same effect as NULL.  The other 
checks for NULL all automatically handle the Py_None case 
the same as if the attribute were filled with a real 
factory.

Code nits:
----------

Pre-parse:
  PyEval_CallMethod((PyObject *)mp, "on_missing", "(O)", 
key);
into:
  PyObject_CallMethodObjArgs((PyObject *)mp, "on_missing", 
key, NULL);


The first four lines in defdict_dealloc() can be 
simplified to:
    Py_Clear(dd->default_factory);

Likewise, the first five active lines of defdict_traverse 
shrink to:
    Py_VISIT(dd->default_factory);

The OFF() macro is used only once.  Better to in-line it.


Other
-----

If you don't get a chance to add the docs, just apply the 
patch and I'll write the docs later.

Likewise, I'll update UserDict and DictMixin to keep the 
APIs in-sync.

Logged In: YES 
user_id=80475

One other thought:  The dict.copy method needs to be 
overriden so that default dicts can duplicate themselves.  
Also consider adding a __reduce__ method to support 
deepcopying and pickling (the latter is less useful 
because functions get pickled by name only).

Logged In: YES 
user_id=6380

Thanks both!  Here's a new patch:

- fixed most of Neal's nits
- fixed most of Raymond's first set of nits
- updated UserDict

Notes:

- I can't use PyObject_CallMethodObjArgs without also
  converting "on_missing" to an object.
  Right now I don't care enough to bother (especially
  since it is only called for proper subclasses).
- I can't find Py_Clear(). Raymond, what did you mean?
- It's not clear how (or even whether) to update DictMixin.
- Haven't gotten to copy() yet -- good point though!
- Still no docs.
- In addition to Raymond's explanation, it's also the case
  that NULL appears as None to the user through the magic
  of the T_OBJECT attribute machinery.

Logged In: YES 
user_id=6380

Here's an update that adds:

- defaultdict.copy() implemented
- docstring tweaks

Logged In: YES 
user_id=80475

Py_CLEAR() is a macro defined in object.h.  See example 
uses in enumobject.c, genobject.c, and itertobject.c.

DictMixin should add the on_missing() method, leaving the 
responsibility for calling it to core object's __getitem__ 
method.  That will likely never happen, but it does allow 
the mapping API to be emulated as fully as possible.  Add 
the method right at the very end (after __len__).

Logged In: YES 
user_id=6380

OK. Py_CLEAR() found; DictMixin.on_missing() added.

Still no docs.

Do UserDict and DictMixin need unit tests additions?

There are grumblings in python-dev on the name; but I still
like defaultdict best.

Logged In: YES 
user_id=80475

I just had one other thought.  If the subclass implemented 
__getitem__ and on_missing directly, then there would be 
no need to make any changes to dictobject.c and we could 
leave the mapping API alone (for a less disruptive 
implementation):

class defaultdict(dict):
    default_factory = None
    def __getitem__(self, key):
        try:
            return dict.__getitem__(self, key)
        except KeyError:
            return self.on_missing(self, key)
    def on_missing(self, key):
        if self.default_factory is None:
            raise KeyError(key)
        value = self[key] = self.default_factory(key)
        return value

Anyone needing to hook the on_missing method can do it 
through a subclass of defaultdict rather than dict.

There are two side benefit to this approach.  One, we keep 
the changes all in one object making it easier to 
understand than putting framework methods in the parent.  
Two, this also allows the on_missing call to be done 
directly instead of through PyEval_CallMethod().  That 
will provide a nice speed-up for the common cases where 
defaultdict hasn't been subclassed.

Logged In: YES 
user_id=6380

Here's a new version.  This adds:

- fix for copying functions in copy.py (see python-dev post)
- more unit tests
- latex docs
- docstring tweaks

I haven't got latex installed on my laptop so I can't test
whether the docs are any good.  Can anyone?

I did this on the plane so your proposal below isn't
incorporated yet.  My comment is that I actually *like*
having the functionality in the dict base class since
someone else implementing on_missing() might not need the
default_factory instance variable; subclassing defaultdict
would be kind of wrong for that purpose.

I'm not sure I understand your comment about optimizing away
PyEval_CallMethod() -- don't we still need to use that so
that a subclass can override it?  Or do you propose to do
another check for the exact defaultdict type and if so,
inline it or at least shortcut the call?  I'm not sure I see
the value in such a micro-optimization at this point --
after all it's only called for missing keys which shouldn't
be all that frequent.

Logged In: YES 
user_id=80475

Maybe after sleeping on the idea, you'll like it better.  
Otherwise, we're adding on_missing() to the dict API where 
it cannot possibly be called, only a subclass can do that.

The essence of the patch is meeting the known regular 
default_factory use cases (with list, int, or set).  No 
one has yet proposed use cases for overriding on_missing() 
with other functionality.

So if you code it all in default_dict, you make the patch 
cleaner, leave the dict API unmolested, make it easier to 
understand the call sequence (no parent to child to parent 
steps), and still leave open the possibility that someone 
could override default_dict.on_missing() if they discover 
a need for defaults computed from a key.

Also, I suspect that from a beginner point-of-view that 
having on_missing() defined in dicts will create confusion 
(don't call this method, just override it in a subclass, 
it is only triggered by getitem, there is no relation to 
get() or to setdefault(), ...)

Logged In: YES 
user_id=80475

Without LaTeX on your system, you can still do a pretty 
good check by running the document through:

  python Tools/scripts/texcheck.py collections.tex

Logged In: YES 
user_id=33168

LaTex is clean.

Logged In: YES 
user_id=764593

Two use cases for overriding on_missing, but not needing a 
factory:

(1)  Stacked configuration objects.  If it isn't in the 
first, check the second, etc.  Override on_missing with the 
__getitem__ of the next configuration object.  (Or perhaps a 
for loop over several fallbacks -- but do it in the 
dictionary, instead of at every lookup site.)

(2)  Caching.  One of the problems with setdefault is that 
it always computes the default, which may be expensive.  

Note that in both of these cases, the conceptual model is 
not "all new entries get the same initial value", but rather 
"A normal dictionary, which just happens to be so expensive 
to create that I don't want to prepopulate."

Logged In: YES 
user_id=6380

Michael Chermside had a good point -- the hook should be
called __missing__ (he suggested __on_missing__ but I think
__missing__ is fine).

Here's a new patch, which also implements and documents and
tests the change that I earlier agreed to on python-dev:
dict doesn't define on_missing (__missing__) but just checks
for it -- if a subclass doesn't define it, the default
behavior applies.

__missing__ must be implemented as a method -- it can't be
an instance variable. 

Logged In: YES 
user_id=764593

v9 updates name to __missing__, but not yet for UserDict.

Logged In: YES 
user_id=6380

OK, here's a new one that fixes UserDict.  With unit tests.

Logged In: YES 
user_id=6380

Committed revision 42573.