Oct. 16th, 2014

I was reading Allison's blog post on how to start exploring Python internals, and one of the suggestions was: try implementing a Python library function without looking at it! I thought this sounded like splendid fun; also, one of the suggestions was namedtuple and I actually REALLY LIKE namedtuple but don't have occasion to use it often enough. So I dove in! Stuff I learned so far doing this:
  • Metaclasses! I already knew about these in a vague "it's like a thing that creates classes or something" sort of way, and since it seems like namedtuple creates class-like objects, I thought it'd be a good place to start. Probably the most interesting thing I discovered: the plain old type method, which I've always used just to check the types of objects, can also be used to dynamically create new classes! This seems like a super-odd and unintuitive dual functionality, and I found a throwaway comment that claimed this was due to historic/backwards compatibility reasons, but I wasn't able to determine what these reasons were. (Let me know if you know!)

  • With type() alone, you can create a pretty decent named tuple, which I coded up like so. Granted, it's (a) not a tuple at all, and (b) does some slightly frownyface manhandling of class properties, and (c) doesn't implement all the functionality of namedtuple... BUT, it does handle my most common use case for namedtuple, which tends to be: "Hey, I want a kind-of-throwaway class that'll be used only in a small section of the code—but that throwaway class will make what I'm doing SO MUCH MORE READABLE." Thus, tada! Instant objects with sensible properties!

  • But for some reason I got to wondering: could you make a function that, say, knows to simply create a Foo when you call namedtuple('Foo', 'my properties'), rather than having to do Foo = namedtuple('Foo', 'my properties')? It turns out the answer is YES, but you have to do evil things to make it happen. Essentially, Python maintains dictionaries of variables for you—try typing globals() or locals() into your Python interpreter to see!

    In order to auto-generate our Foo class, then, we want to add Foo to the local variable dictionary of the caller. (Meaning: if we're calling namedtuple('Foo', 'my properties') within our main method, we want Foo to be created in that main method, not just within the namedtuple call.) Turns out there's a _getframe function you can use to get, say, the current frame, or the parent frame... and then just tack Foo onto the parent frame and you're good to go!

    But that's all a terrible idea and you shouldn't do it. It's not good for you. It's not good for the planet. Don't be like me.
I've got an actual, good-for-the-planet implementation of namedtuple underway, so hopefully I can share a real gist of that with you all soon!

Edit: Ned pointed out that the super(self.__class__, self).__init__() call I had in my init functions for the janky and trolly tuples wasn't quite right—calling super on our hand-rolled class gets us a NoneType, so it doesn't really make sense to call it. I updated the code to be more correct now. Thanks, Ned!