Instances' __setitem__ methods

[Spencer Pearson]

I was recently trying to implement a dict-like object which would do
some fancy stuff when it was modified, and found that overriding the
__setitem__ method of an instance did not act the way I expected. The
help documentation (from help(dict.__setitem__)) claims that
"d.__setitem__(k,v)" is equivalent to "d[k]=v", but I've produced this
code that, on Python 2.6, acts differently in the two cases.

def print_args( key, value ):
    print "print_args called: key = %s, value = %s" %(key,value)

class MyDict( dict ):
    def __init__( self ):
        dict.__init__( self )
        self.__setitem__ = print_args

    def __setitem__( self, key, value ):
        print "ModelDict.__setitem__ called"
        dict.__setitem__( self, key, value )

d = MyDict()

print "d.__setitem__(0,1):",
d.__setitem__(0,1)

print "d[0]=1:",
d[0]=1


I would expect the two setitems to both call print_args, but that's
not what happens. In the first case, it calls print_args, but in the
second case, the __setitem__ declared in MyDict is called instead.

The documentation at http://docs.python.org/reference/datamodel.html#specialnames
says that for new-style classes, "x[i]" is equivalent to
"type(x).__getitem__(x, i)". I assume that "x[i]=y" has similarly been
changed to be equivalent to "type(x).__setitem__(x, i, y)", since that
would produce the results that I'm getting. Is the help documentation
for dict.__setitem__ just outdated, or am I missing some subtlety
here?

Also: when I say "d.f(*args)", am I correct in thinking that d checks
to see if it has an instance attribute called "f", and if it does,
calls f(*args); and if it doesn't, checks whether its parent class
(and then its grandparent, and so on) has a class attribute called
"f", and if it does, calls f(x, *args)?