[Python-Dev] RE: [Python-checkins] python/nondist/sandbox/datetime obj_delta.c,1.14,1.15 test_both.py,1.15,1.16

[Jeremy Hylton]

>>>>> "TP" == Tim Peters <tim.one@comcast.net> writes:

  TP> I'd rather find a way to avoid __reduce__ entirely, though!

Me, too!

  TP> The Python implementation of these things didn't need it, and in
  TP> the date and datetime cases it's creating bigger pickles than it
  TP> should -- __getstate__ and __setstate__ already did all that was
  TP> necessary, and no more than that.  Supplying an argument tuple
  TP> for __reduce__'s benefit loses either way: I either put the real
  TP> date/datetime arguments there, but then the pickle is of a big
  TP> tuple rather than of a tiny string.  Or I put a dummy argument
  TP> tuple there and also include the tiny string for __setstate__,
  TP> but these constructors require at least 3 arguments so that the
  TP> "dummy argument tuple" consumes substantial space of its own.

  TP> So, as it stands, ignoring the new-style-class administrative
  TP> pickle bloat in the Python implementation, the *guts* of the
  TP> pickles produced by the Python implementation are substantially
  TP> smaller than those produced by the C implementation.

The __reduce__() approach seems to be (either or both) overly complex
and underdocumented.  It's a real shame that something simple like
__getstate__() and __setstate__() can't be made to work for new-style
classes.

If I recall correctly, the problem is that there is no way to
distinguish a user-defined new-style class from a builtin type from a
user-defined subclass of a builtin type.  As a result, there's no way
for pickle to decide if it should be looking for __getstate__() or
invoking the complicated machinery that allows subclasses of builtin
types to be pickleable.  Another victim of unification.

> I figured timedelta *was* "a class", although now I guess that, in
> this context, it's not "a class" but "a type".  That's what you get
> when you unify the concepts <wink>.

Jerem