A common design pattern in clojure is to have a single massive data-structure that represents all application state, but to take it apart recursively whilst updating so that each function is only passed the pieces of state that are relevant. That way you can look at a function and immediately know what it can and can't read/write out of the global state.
A nice way of doing this in an imperative language might be to pass bits of the global data structure by reference (ala lenses) so that a given function can only read/write state contained in its arguments. The data is still global in the sense that it is all accessible from the root object and is not encapsulated, but access can be restricted on a per-function basis.
That's true. The equivalent effect in the "all local" paradigm is local objects the state of which can't meaningfully be understood or manipulated by neither the programmer nor other objects, leading again to unexpected behavior that is painful to track down.
A lot of the precautions that are common sense when working with global data are already instinctively followed (or at least understood) by most programmers I think. In effect the lenses are already working when there's an understanding when and where manipulation occurs. This mostly coincides with the notion that complex data should be manipulated by a well-defined model and wherever sensible there should be only one mechanism for doing it.
def foo(x):
x_plus_one = add_one(x)
x_plus_one_times_two = times_two(x_plus_one)
return x_plus_one_times_two
Here is a function with some local state. Which part do you feel "can't meaningfully be understood or manipulated by [...] the programmer"? I'm not trying to bait you, but I'm struggling to grasp the argument you are making.[1] https://en.wikipedia.org/wiki/Referential_transparency_%28co...
def foo():
a = []
append_foo(a)
return a
def append_foo(a):
a.append('foo')
But then I wouldn't call this "local mutability", in the sense that you modify your list in a different part of the code. Not a problem in a trivial example like this, definitively an issue when the actual action performed is ten levels deep.Abstractly speaking, when I say "A often has the property B", you can't meaningfully counter that with "you're wrong, here's an example where !B".
Fortunately the confusion is cleared up now.
You can't return locally created variables [beyond primitives].
Beyond simple primitive reading/returning functions, you often create some arrays or structs, then pass a pointer to the necessary arrays and structs to your functions, so it can read and write them.
This way the functions can stay pure.
This is just free vs. bound variables. You're absolutely right that it would be a pain to write things such that there are no free variables from parent environments. However, those functions aren't necessarily from the top-level (global) environment.
>However, in striving to ban explicit global state we have developed an astonishing array of cruft and complexity which counteracts the effects we wanted to achieve in the first place.
I think it's important to avoid global, mutable state in most cases. Relying on functions and constants from the top-level environment is fine, but storing a program's state in the top-level environment is just asking for a whole lot of painful debugging.
One of the reasons global state gets a bad rap is because we're always trying to minimize side effects, and that's a worthy goal. The idea not only makes sure the program's components test well individually, it also enables better component re-use (although I'm still convinced that re-use without refactoring is mostly a myth in practice). However, in striving to ban explicit global state we have developed an astonishing array of cruft and complexity which counteracts the effects we wanted to achieve in the first place.
I agree with the article that well-defined and clean global data makes a lot of code easier to handle and it also eliminates unnecessary work, both on the human and the machine side. Of course, this idea breaks down again when the global substrate becomes muddled and structurally broken. At which point we've come full circle.
In my opinion a mixed approach is probably advisable for most projects and in fact, that's the way we already do things in many cases, even if we're forbidden to use the actual phrase "global variables". Judicious use of both paradigms yields the best results in my opinion. Maybe it's time to actually start calling global state by its real name, without expecting to be stigmatized for it.