Performance of accessing class variables in Python

问题:

I wonder if there is any difference in performance when accessing a class variable (a dict) inside a method of the same class using:

self.class_variable_dict.add(some_key, some_value)

and

ClassName.class_variable_dict.add(some_key, some_value)

obviously, both will work as long as there is no instance variable with the same name, but is there any reason/use case for which we should prefer one over the other?

回答1:

Accessing it via ClassName rather than via self will be slightly faster, since if you access it via self it must first check the instance namespace. But I don't expect the difference to be at all significant, unless you have profiling information to suggest that it is.

So I would recommend using whichever one you think is easier to read/understand as a human.

Semantically, they will be different only if the class_variable_dict variable gets shadowed somewhere -- in particular, if (a) self defines a variable of the same name; or (b) self is an instance of a subclass of ClassName, and that subclass (or one of its bases that's still a subclass of ClassName) defines a variable of the same name. If neither of those is true, then they should be semantically identical.

Edit:

delnam has a good point: there are factors that might make either faster. I stand by my assertion that the difference will be trivial unless it's in a very very tight loop. To test it, I created the tightest loop I could think of, and timed it with timeit. Here are the results:

• access via class var: 20.226 seconds
• access via inst var: 23.121 seconds

Based on several runs, it looks like the error bars are about 1sec -- i.e., this is a statistically significant difference, but probably not worth worrying about. Here's my test code:

import timeit

setup='''
class A:
    var = {}
    def f1(self):
        x = A.var
    def f2(self):
        x = self.var

a = A()
'''
print 'access via class var: %.3f' % timeit.timeit('a.f1()', setup=setup, number=100000000)
print 'access via inst var: %.3f' % timeit.timeit('a.f2()', setup=setup, number=100000000)

回答2:

Let's look at what the different options all do.

In [1]: class Foo:
   ...:     bar = {}
   ...:     

In [2]: import dis
In [3]: dis.dis(lambda: Foo.bar.add(1,2))
  1           0 LOAD_GLOBAL              0 (Foo) 
              3 LOAD_ATTR                1 (bar) 
              6 LOAD_ATTR                2 (add) 
              9 LOAD_CONST               1 (1) 
             12 LOAD_CONST               2 (2) 
             15 CALL_FUNCTION            2 
             18 RETURN_VALUE         

In [4]: dis.dis(lambda: Foo().bar.add(1,2))
  1           0 LOAD_GLOBAL              0 (Foo) 
              3 CALL_FUNCTION            0 
              6 LOAD_ATTR                1 (bar) 
              9 LOAD_ATTR                2 (add) 
             12 LOAD_CONST               1 (1) 
             15 LOAD_CONST               2 (2) 
             18 CALL_FUNCTION            2 
             21 RETURN_VALUE 

As you can see from that, both styles generate the same bytecode aside from creating the object in the second case.

---

The other facet to this is that it doesn't matter. Use whatever expresses your objective in the most precise way. Only optimize if the speed matters.

I'd recommend just going with ClassName.dict_value[key] = value

回答3:

Instance variables vs. class variables in Python

See also here same question where I posted a broader test based upon @Edward Loper test including module variables as well which are even faster than class variables.