Limiting the scope of mutation

Functional programming teaches us about the dangers of mutability. But some object-oriented code mutates variables freely, in different scopes, along different code paths, and in many places.

To maintain our ability to reason about code, we can learn from functional programming and treat mutation as a risk to be managed. Even if we stop short of mandatory immutability and using constructs like monads, we can still manage mutations and recognize them as a source of bugs and confusion.

I’m always nervous when I see code like this:

class Unwieldy < GrabBag::Base
  include BizarroHelpers
  # ... 500 lines of spaghetti

  def do_something(wibble)
    result = []
    intermediate_result = []

    # ... 20 lines of code

    intermediate_result << weird_helper(wibble)

    # ... 20 lines of code

    intermediate_result << biscuits? ? gravy : grits

    # ... 20 lines of code

    if weirder_helper(wibble)
      intermediate_result.reject!(&:magic)
    else
      result << inexplicable_helper(wibble)
    end

    # ... 20 lines of code

    if confusing_predicate_method?
      result += intermediate_result
      return result
    end

    result << :deus_ex_machina
    result
  end

  # ... 500 lines of spaghetti
end

We might not be able to fix all the spaghetti code and we’re sure to have difficulty grokking what comes from where (the class inherits from GrabBag::Base and mixes in BizzaroHelpers). But if we’re writing do_something, we can use a few rules of thumb to avoid getting too confused and to make our code more maintainable.

To summarize the ideas presented below: Generally, avoid declaring a basic object and repeatedly mutating it in a large scope. Prefer instantiating objects using literals, other objects, and the builder pattern/fluent API. Run the code with mutations in an attached scope, or at least right next to the declaration. And avoid building logic into the interaction of control flow with mutation.

Keep mutation near the declaration.

The examples here use a Rust HashMap. I chose that because the standard library doesn’t give us a way to construct one using a literal.

The following Rust code has a few benefits: it keeps all mutations close together and it immediately makes lunch immutable once the assigning is done.

#![allow(unused_variables)]

use std::collections::HashMap;

fn main() {
    let mut lunch = HashMap::new();
    lunch.insert("salad", "house");
    lunch.insert("entree", "sandwich");
    lunch.insert("dessert", "cookie");
    lunch.insert("drink", "milk");
    let lunch = lunch; // `let` binding without `mut` makes `lunch` immutable.
}

Better yet, put mutation in a scope attached to the declaration.

This is overkill here, but it demonstrates a point. Notice that the scope in which the HashMap is mutable is only in the body of the closure passed to fold:

let lunch = [
    ("salad", "house"),
    ("entree", "sandwich"),
    ("dessert", "cookie"),
    ("drink", "milk"),
]
.iter()
.fold(HashMap::new(), |mut result, (k, v)| {
    result.insert(k, v);
    result
});

Ruby has some nice patterns with Object#tap and Object#yield_self, for example:

immutable = MyClass.new.tap |mutable| do
  mutable.blip = 8
  mutable.blop = 10
end

An instance of MyClass is mutated after being instantiated, but before it is bound to immutable. We do that with a scope in which the instance of MyClass is mutated, but mutable goes out of scope at the end of the tap block; the result is now bound to immutable and—we hope—not mutated any more. This doesn’t actually prevent anyone from mutating immutable (that’s just a name, after all), but it makes our intent clear and helps the next developer follow the same pattern.

Better yet, use a literal or the builder pattern to instantiate the object.

A Rust crate called maplit exposes a macro for creating literal hashmaps:

let lunch = hashmap! {
    "salad" => "house",
    "entree" => "sandwich",
    "dessert" => "cookie",
    "drink" => "milk",
};

If you have control over objects, you can allow them to be initialized with a literal. For example:

use std::collections::HashMap;

// The "Newtype" pattern. Lunch is basically a `HashMap`, but we're making our
// own type as a thin wrapper over it (in Rust it's a struct, but it would be a
// class in many languages) so that we control its construction and which
// methods we expose.
struct Lunch(HashMap<String, String>);

impl Lunch {
    // You can ignore the `&str` and `to_string()` stuff for this example and
    // view them all as strings. It's a Rust quirk.
    fn new(init: Vec<(&str, &str)>) -> Self {
        let mut result = HashMap::with_capacity(init.len());
        for (k, v) in init {
            result.insert(k.to_string(), v.to_string());
        }
        Self(result)
    }
}

fn main() {
    let lunch = Lunch::new(vec![
        ("salad", "house"),
        ("entree", "sandwich"),
        ("dessert", "cookie"),
        ("drink", "milk"),
    ]);
}

Or you can use the builder pattern with an object you define:

#![allow(unused_variables)]

use std::collections::HashMap;

// You'd do this with named struct fields; just making this example more like
// the others.
struct Lunch(HashMap<String, String>);

impl Lunch {
    fn new() -> Self {
        Lunch(HashMap::with_capacity(4))
    }

    fn with_salad(mut self, salad: &str) -> Self {
        self.0.insert("salad".to_string(), salad.to_string());
        self
    }

    fn with_entree(mut self, entree: &str) -> Self {
        self.0.insert("entree".to_string(), entree.to_string());
        self
    }

    fn with_dessert(mut self, dessert: &str) -> Self {
        self.0.insert("dessert".to_string(), dessert.to_string());
        self
    }

    fn with_drink(mut self, drink: &str) -> Self {
        self.0.insert("drink".to_string(), drink.to_string());
        self
    }
}

fn main() {
    let lunch = Lunch::new()
        .with_salad("house")
        .with_entree("sandwich")
        .with_dessert("cookie")
        .with_drink("milk");
}

Assign from rather than in conditional expressions.

In many languages, if/else is an expression rather than a statement. When that is so, make use of this to assign the value of the if/else expression rather than assigning within the if. That fits in with the rest of the recommendations because it entails mutation in one scope that affects a larger scope.

Thus, prefer

my_result = if predicate?
              value_a
            else
              value_b
            end

to

if predicate?
  my_result = value_a
else
  my_result = value_b
end

Among other reasons, it assures that my_result is always defined, and it makes clear that the intent of the if/else is to bind a value to a variable. (And of course, don’t also do some side-effecting thing in the conditional. Remember the command–query separation principle.)

In JavaScript, ternaries are expressions while if/else blocks are not. So you can use ternaries in lieu of if/else expressions in some cases. But ternaries are probably best confined to small amounts of code because the ? and : operators are easy to lose track of.

Do not embed logic in the interaction of flow control and mutation.

This is a general statement of the specific example of if/then statements from the previous section. When your business logic is immanent from the interaction of control flow and mutation, it can be quite difficult to reason about your code, the code is more difficult to maintain, and bugs are more likely.

The tongue-in-cheek example at the beginning is a case of this. A common pattern is for a result variable to be mutated throughout a large function, with some mutation occurring inside conditionals, and result sometimes short-circuit returned. The author of the code may think he or she is helping you out because you don’t wind up deep within nested conditionals. Yet the effect is often worse: the logic is just as confusing, but now the logic is implicit: the state of your result variable is determined by the combination of not having returned at line X, having entered the else block at line Y, and so forth.