|
| 1 | +--- |
| 2 | +layout: post |
| 3 | +author: Tony Schneider |
| 4 | +title : Classic Rails Autoloading and Singleton Classes |
| 5 | +date : 2020-01-23 |
| 6 | +tags : software |
| 7 | +--- |
| 8 | + |
| 9 | +Rails constant autoloading is a really nifty feature. |
| 10 | + |
| 11 | +It lets us reference constants without explicitly requiring them. |
| 12 | +It is also what allows Rails to pick up changes to files without having to constantly restart your server. |
| 13 | + |
| 14 | +It's one of those features that goes under appreciated. |
| 15 | +If it's working you probably don't even know it's there -- until something goes wrong :laughing: |
| 16 | + |
| 17 | +One gotcha I've run into a number of times is how it interacts with Ruby's `class << self` syntax. |
| 18 | + |
| 19 | +## 10,000 Foot View |
| 20 | + |
| 21 | +When you write Ruby code, you are likely defining lots of constants. |
| 22 | +Module names are constants, class names are constants, and of course constants are constants! |
| 23 | + |
| 24 | +When you run the code, Ruby expects you to have defined the constants you use. |
| 25 | +Typically this is done by requiring the ruby file(s) that you plan to use. |
| 26 | + |
| 27 | +```ruby |
| 28 | +# in foo.rb |
| 29 | +class Foo |
| 30 | +end |
| 31 | + |
| 32 | +# in bar.rb |
| 33 | +require "foo" |
| 34 | +class Bar |
| 35 | + def initialize |
| 36 | + @foo = Foo.new |
| 37 | + end |
| 38 | +end |
| 39 | +``` |
| 40 | + |
| 41 | +If we were to reference `Foo` in `bar.rb` without the `require "foo"`, because `const_get` fails, ruby calls `const_missing`. |
| 42 | +Without any further intervention, it would raise an... uninitialized constant error a la `NameError (uninitialized constant Foo)` |
| 43 | + |
| 44 | +The `const_missing` method serves as a hook for dynamically resolving the missing constant. |
| 45 | + |
| 46 | +Classic Rails autoloading works by doing exactly this! |
| 47 | +It implements `const_missing` and relies on file location conventions to figure out where to find the missing constant. |
| 48 | + |
| 49 | +What happens when you break convention you ask? |
| 50 | +Welp, let’s just say it might be a bit of a learning experience. |
| 51 | + |
| 52 | +## First some Ruby |
| 53 | + |
| 54 | +```ruby |
| 55 | +class YourClass |
| 56 | + def self.hello |
| 57 | + "Why hello there" |
| 58 | + end |
| 59 | + |
| 60 | + class << self |
| 61 | + def hello2 |
| 62 | + "Why hello there" |
| 63 | + end |
| 64 | + end |
| 65 | +end |
| 66 | + |
| 67 | +YourClass.hello # => Why hello there |
| 68 | +YourClass.hello2 # => Why hello there |
| 69 | +``` |
| 70 | + |
| 71 | +You may have come across this syntax in your ruby usage. |
| 72 | + |
| 73 | +The `class << self` syntax isn’t simply an alternate syntax for defining class methods. |
| 74 | +While ultimately it results in a class method, you’re doing so by opening the class’s "eigenclass". |
| 75 | +Sounds intimidating at first, but _eigen_ just means "self" in German. |
| 76 | + |
| 77 | +Remember that in the ruby object model you have classes and instances of those classes. |
| 78 | +A user defined class is an instance of the class _Class_ (whew :sweat_smile:). |
| 79 | + |
| 80 | +When inside an instance method, `self` is a way to refer to the instance _itself_. |
| 81 | +This usage of `self` is different because we’re operating in the scope of the class instead of the instance. |
| 82 | +As a result, in this example, `self` is actually `YourClass`. |
| 83 | + |
| 84 | +Walking around saying the word “Eigenclass” sure does make you sound smart, but I find it's easier to reason about when referred to as a "singleton class". |
| 85 | + |
| 86 | +### What's a Singleton Class? |
| 87 | + |
| 88 | +If every class you define is an instance of `Class`, where do your class's class methods actually live? |
| 89 | + |
| 90 | +One option would be to define them as instance methods on the class `Class`. |
| 91 | +I think you'd be forgiven for thinking that given the "class" vs "instance of class" distinction discussed above. |
| 92 | + |
| 93 | +While in a way poetic, we wouldn't want Ruby to define our class methods as instance methods on the class `Class` because it would mean all instances of `Class` would have our class method. |
| 94 | + |
| 95 | +```ruby |
| 96 | +Class.new.hello #=> Why hello there |
| 97 | +Class.new.hello2 #=> Why hello there |
| 98 | +``` |
| 99 | + |
| 100 | +That would be... insanity. |
| 101 | + |
| 102 | +To alleviate this, each ruby class has an anonymous singleton class that it uses to store class methods. |
| 103 | +In other words, your class methods are actually defining instance methods on this singleton class. |
| 104 | +Similarly, when you call your class method, your class calls an instance method on the class’s singleton class. |
| 105 | + |
| 106 | +You can actually see your class’s singleton class by doing `YourClass.singleton_class`! |
| 107 | + |
| 108 | +```ruby |
| 109 | +# should be the same list |
| 110 | +YourClass.singleton_class.instance_methods(false) |
| 111 | +YourClass.singleton_methods |
| 112 | +``` |
| 113 | + |
| 114 | +As the name implies, you cannot (thank goodness) create instances of the singleton class: |
| 115 | + |
| 116 | +```ruby |
| 117 | +YourClass.singleton_class.new #=> NOPE |
| 118 | +``` |
| 119 | + |
| 120 | +## So What’s the Difference? |
| 121 | + |
| 122 | +When you use the `def self.` approach, ruby takes care of defining an instance method on the singleton class for you. |
| 123 | + |
| 124 | +You’ll notice the `class << self` variant doesn’t use `def self.` at all. |
| 125 | + |
| 126 | +Hopefully now you see why — because the `class << self` syntax opens up the singleton class allowing you to define instance methods on it directly. |
| 127 | +As a result of doing this, your class now has access to those methods as class methods! |
| 128 | + |
| 129 | +Probably the most common reason I see folks reaching for `class << self` is to take advantage of this instance method definition as a way to define private class methods without resorting to the admittedly awkward `private_class_method` method. |
| 130 | + |
| 131 | +The other difference that’s seems less talked about is the impact to `Module.nesting` which is crucial to any autoloading implementation. |
| 132 | +Because you’re defining methods in different scopes (`YourClass` vs `YourClass.singleton_class`), you’re going to get different answers when `Module.nesting` is called. |
| 133 | + |
| 134 | +```ruby |
| 135 | +class A |
| 136 | + class << self |
| 137 | + def foo |
| 138 | + Module.nesting |
| 139 | + end |
| 140 | + end |
| 141 | + |
| 142 | + def self.bar |
| 143 | + Module.nesting |
| 144 | + end |
| 145 | + |
| 146 | + def baz |
| 147 | + Module.nesting |
| 148 | + end |
| 149 | +end |
| 150 | + |
| 151 | +A.foo # => [#<Class:A>, A] |
| 152 | +A.bar # => [A] |
| 153 | +A.new.baz # => [A] |
| 154 | +``` |
| 155 | + |
| 156 | +## Back to Rails |
| 157 | + |
| 158 | +In Rails 5, here is roughly how the "classic" autoloading algorithm works (taken from the autoloading guides): |
| 159 | + |
| 160 | +``` |
| 161 | +if the class or module in which C is missing is Object |
| 162 | + let ns = '' |
| 163 | +else |
| 164 | + let M = the class or module in which C is missing |
| 165 | + |
| 166 | + if M is anonymous |
| 167 | + let ns = '' |
| 168 | + else |
| 169 | + let ns = M.name |
| 170 | + end |
| 171 | +end |
| 172 | + |
| 173 | +loop do |
| 174 | + # Look for a regular file. |
| 175 | + for dir in autoload_paths |
| 176 | + if the file "#{dir}/#{ns.underscore}/c.rb" exists |
| 177 | + load/require "#{dir}/#{ns.underscore}/c.rb" |
| 178 | + |
| 179 | + if C is now defined |
| 180 | + return |
| 181 | + else |
| 182 | + raise LoadError |
| 183 | + end |
| 184 | + end |
| 185 | + end |
| 186 | + |
| 187 | + # Look for an automatic module. |
| 188 | + for dir in autoload_paths |
| 189 | + if the directory "#{dir}/#{ns.underscore}/c" exists |
| 190 | + if ns is an empty string |
| 191 | + let C = Module.new in Object and return |
| 192 | + else |
| 193 | + let C = Module.new in ns.constantize and return |
| 194 | + end |
| 195 | + end |
| 196 | + end |
| 197 | + |
| 198 | + if ns is empty |
| 199 | + # We reached the top-level without finding the constant. |
| 200 | + raise NameError |
| 201 | + else |
| 202 | + if C exists in any of the parent namespaces |
| 203 | + # Qualified constants heuristic. |
| 204 | + raise NameError |
| 205 | + else |
| 206 | + # Try again in the parent namespace. |
| 207 | + let ns = the parent namespace of ns and retry |
| 208 | + end |
| 209 | + end |
| 210 | +end |
| 211 | +``` |
| 212 | + |
| 213 | +The part we're going to focus on is the condition that says: |
| 214 | + |
| 215 | +``` |
| 216 | + if M is anonymous |
| 217 | + let ns = '' |
| 218 | +``` |
| 219 | + |
| 220 | +From the section above we discovered the class's singleton class is an anonymous class. |
| 221 | +Because of this, this condition is going to expect unloaded constants explicitly defined in our singleton class to be located in the top level namespace. |
| 222 | + |
| 223 | +So let's go to an example you might see in the wild: |
| 224 | + |
| 225 | +```ruby |
| 226 | + module SomeNamespace |
| 227 | + class PolicyService |
| 228 | + |
| 229 | + def self.create_policy |
| 230 | + RatingService.create |
| 231 | + end |
| 232 | + |
| 233 | + class << self |
| 234 | + def create_policy2 |
| 235 | + RatingService.create |
| 236 | + end |
| 237 | + end |
| 238 | + end |
| 239 | + end |
| 240 | +``` |
| 241 | + |
| 242 | +### Variant 1: `def self.` (class scope) |
| 243 | + |
| 244 | +As defined, the nesting inside `PolicyService.create_policy` is: |
| 245 | + |
| 246 | +```ruby |
| 247 | +# [ |
| 248 | +# SomeNamespace::PolicyService |
| 249 | +# SomeNamespace |
| 250 | +# ] |
| 251 | +``` |
| 252 | + |
| 253 | +As a result, `PolicyService.create_policy` works as expected, first checking for `RatingService` in `SomeNamespace::PolicyService`, then `SomeNamespace` and finally at the top level via `::RatingService`. |
| 254 | + |
| 255 | +## Variant 2: `class << self` (singleton class scope) |
| 256 | + |
| 257 | +Subtly different, the nesting for `PolicyService.create_policy2` is: |
| 258 | + |
| 259 | +```ruby |
| 260 | +# [ |
| 261 | +# #<Class:SomeNamespace::PolicyService>, |
| 262 | +# SomeNamespace::PolicyService, |
| 263 | +# SomeNamespace |
| 264 | +# ] |
| 265 | +``` |
| 266 | + |
| 267 | +If the constant is already loaded by something else, great, no autoloading required. |
| 268 | + |
| 269 | +However, if the constant is missing, rails is going to look at the top level namespace. |
| 270 | +If the constant isn't defined at the top level namespace, you will get a `NameError`. |
| 271 | + |
| 272 | +Worse yet, if there is a constant defined at the top level namespace, it might not be the desired constant! :scream: |
| 273 | + |
| 274 | +This can be a real pain in the neck to track down since you are unlikely to always load classes in the same order when running tests. |
| 275 | +Similarly, in development, your classes will get reloaded to reflect the changes you've made, potentially causing them to be reloaded in a new order. |
| 276 | + |
| 277 | +Sound familiar? :smiling_imp: |
| 278 | + |
| 279 | +## Conclusion |
| 280 | + |
| 281 | +I'd recommend **not** using `class << self` when you're working in a Rails autoloaded directory (e.g `app/**/*`). |
| 282 | + |
| 283 | +If you're using private class methods so much that you feel the need to crack open the singleton class, perhaps there's an instance hiding in your code waiting to be discovered. |
| 284 | + |
| 285 | +Hopefully next time you see a missing constant error, you’ll be able to track it down faster with this knowledge in your toolbox. |
| 286 | + |
| 287 | +—-- |
| 288 | + |
| 289 | +Consider this post a farewell letter to our dear friend (and occasional mortal enemy), Classic Rails Autoloading. |
| 290 | + |
| 291 | +Rails 6 reworks (thankfully :sweat_smile:) how autoloading is done using a library called [Zeitwerk](https://github.com/fxn/zeitwerk) and I'm really excited for it! |
0 commit comments