rails--rails/guides/source/autoloading_and_reloading_constants.md

DO NOT READ THIS FILE ON GITHUB, GUIDES ARE PUBLISHED ON https://guides.rubyonrails.org.

Autoloading and Reloading Constants

This guide documents how constant autoloading and reloading works.

After reading this guide, you will know:

• Key aspects of Ruby constants
• What are the autoload_paths and how does eager loading work in production?
• How constant autoloading works
• What is require_dependency
• How constant reloading works
• Solutions to common autoloading gotchas

---

Introduction

Ruby on Rails allows applications to be written as if their code was preloaded.

In a normal Ruby program classes need to load their dependencies:

require 'application_controller'
require 'post'

class PostsController < ApplicationController
  def index
    @posts = Post.all
  end
end

Our Rubyist instinct quickly sees some redundancy in there: If classes were defined in files matching their name, couldn't their loading be automated somehow? We could save scanning the file for dependencies, which is brittle.

Moreover, Kernel#require loads files once, but development is much more smooth if code gets refreshed when it changes without restarting the server. It would be nice to be able to use Kernel#load in development, and Kernel#require in production.

Indeed, those features are provided by Ruby on Rails, where we just write

class PostsController < ApplicationController
  def index
    @posts = Post.all
  end
end

This guide documents how that works.

Constants Refresher

While constants are trivial in most programming languages, they are a rich topic in Ruby.

It is beyond the scope of this guide to document Ruby constants, but we are nevertheless going to highlight a few key topics. Truly grasping the following sections is instrumental to understanding constant autoloading and reloading.

Nesting

Class and module definitions can be nested to create namespaces:

module XML
  class SAXParser
    # (1)
  end
end

The nesting at any given place is the collection of enclosing nested class and module objects outwards. The nesting at any given place can be inspected with Module.nesting. For example, in the previous example, the nesting at (1) is

[XML::SAXParser, XML]

It is important to understand that the nesting is composed of class and module objects, it has nothing to do with the constants used to access them, and is also unrelated to their names.

For instance, while this definition is similar to the previous one:

class XML::SAXParser
  # (2)
end

the nesting in (2) is different:

[XML::SAXParser]

XML does not belong to it.

We can see in this example that the name of a class or module that belongs to a certain nesting does not necessarily correlate with the namespaces at the spot.

Even more, they are totally independent, take for instance

module X
  module Y
  end
end

module A
  module B
  end
end

module X::Y
  module A::B
    # (3)
  end
end

The nesting in (3) consists of two module objects:

[A::B, X::Y]

So, it not only doesn't end in A, which does not even belong to the nesting, but it also contains X::Y, which is independent from A::B.

The nesting is an internal stack maintained by the interpreter, and it gets modified according to these rules:

• The class object following a class keyword gets pushed when its body is executed, and popped after it.

• The module object following a module keyword gets pushed when its body is executed, and popped after it.

• A singleton class opened with class << object gets pushed, and popped later.

• When instance_eval is called using a string argument, the singleton class of the receiver is pushed to the nesting of the eval'ed code. When class_eval or module_eval is called using a string argument, the receiver is pushed to the nesting of the eval'ed code.

• The nesting at the top-level of code interpreted by Kernel#load is empty unless the load call receives a true value as second argument, in which case a newly created anonymous module is pushed by Ruby.

It is interesting to observe that blocks do not modify the stack. In particular the blocks that may be passed to Class.new and Module.new do not get the class or module being defined pushed to their nesting. That's one of the differences between defining classes and modules in one way or another.

Class and Module Definitions are Constant Assignments

Let's suppose the following snippet creates a class (rather than reopening it):

class C
end

Ruby creates a constant C in Object and stores in that constant a class object. The name of the class instance is "C", a string, named after the constant.

That is,

class Project < ApplicationRecord
end

performs a constant assignment equivalent to

Project = Class.new(ApplicationRecord)

including setting the name of the class as a side-effect:

Project.name # => "Project"

Constant assignment has a special rule to make that happen: if the object being assigned is an anonymous class or module, Ruby sets the object's name to the name of the constant.

INFO. From then on, what happens to the constant and the instance does not matter. For example, the constant could be deleted, the class object could be assigned to a different constant, be stored in no constant anymore, etc. Once the name is set, it doesn't change.

Similarly, module creation using the module keyword as in

module Admin
end

performs a constant assignment equivalent to

Admin = Module.new

including setting the name as a side-effect:

Admin.name # => "Admin"

WARNING. The execution context of a block passed to Class.new or Module.new is not entirely equivalent to the one of the body of the definitions using the class and module keywords. But both idioms result in the same constant assignment.

Thus, an informal expression like "the String class" technically means the class object stored in the constant called "String". That constant, in turn, belongs to the class object stored in the constant called "Object".

String is an ordinary constant, and everything related to them such as resolution algorithms applies to it.

Likewise, in the controller

class PostsController < ApplicationController
  def index
    @posts = Post.all
  end
end

Post is not syntax for a class. Rather, Post is a regular Ruby constant. If all is good, the constant is evaluated to an object that responds to all.

That is why we talk about constant autoloading, Rails has the ability to load constants on the fly.

Constants are Stored in Modules

Constants belong to modules in a very literal sense. Classes and modules have a constant table; think of it as a hash table.

Let's analyze an example to really understand what that means. While common abuses of language like "the String class" are convenient, the exposition is going to be precise here for didactic purposes.

Let's consider the following module definition:

module Colors
  RED = '0xff0000'
end

First, when the module keyword is processed, the interpreter creates a new entry in the constant table of the class object stored in the Object constant. Said entry associates the name "Colors" to a newly created module object. Furthermore, the interpreter sets the name of the new module object to be the string "Colors".

Later, when the body of the module definition is interpreted, a new entry is created in the constant table of the module object stored in the Colors constant. That entry maps the name "RED" to the string "0xff0000".

In particular, Colors::RED is totally unrelated to any other RED constant that may live in any other class or module object. If there were any, they would have separate entries in their respective constant tables.

Pay special attention in the previous paragraphs to the distinction between class and module objects, constant names, and value objects associated to them in constant tables.

Resolution Algorithms

Resolution Algorithm for Relative Constants

At any given place in the code, let's define cref to be the first element of the nesting if it is not empty, or Object otherwise.

Without getting too much into the details, the resolution algorithm for relative constant references goes like this:

1. If the nesting is not empty the constant is looked up in its elements and in order. The ancestors of those elements are ignored.

2. If not found, then the algorithm walks up the ancestor chain of the cref.

3. If not found and the cref is a module, the constant is looked up in Object.

4. If not found, const_missing is invoked on the cref. The default implementation of const_missing raises NameError, but it can be overridden.

Rails autoloading does not emulate this algorithm, but its starting point is the name of the constant to be autoloaded, and the cref. See more in Relative References.

Resolution Algorithm for Qualified Constants

Qualified constants look like this:

Billing::Invoice

Billing::Invoice is composed of two constants: Billing is relative and is resolved using the algorithm of the previous section.

INFO. Leading colons would make the first segment absolute rather than relative: ::Billing::Invoice. That would force Billing to be looked up only as a top-level constant.

Invoice on the other hand is qualified by Billing and we are going to see its resolution next. Let's define parent to be that qualifying class or module object, that is, Billing in the example above. The algorithm for qualified constants goes like this:

1. The constant is looked up in the parent and its ancestors. In Ruby >= 2.5, Object is skipped if present among the ancestors. Kernel and BasicObject are still checked though.

2. If the lookup fails, const_missing is invoked in the parent. The default implementation of const_missing raises NameError, but it can be overridden.

INFO. In Ruby < 2.5 String::Hash evaluates to Hash and the interpreter issues a warning: "toplevel constant Hash referenced by String::Hash". Starting with 2.5, String::Hash raises NameError because Object is skipped.

As you see, this algorithm is simpler than the one for relative constants. In particular, the nesting plays no role here, and modules are not special-cased, if neither they nor their ancestors have the constants, Object is not checked.

Rails autoloading does not emulate this algorithm, but its starting point is the name of the constant to be autoloaded, and the parent. See more in Qualified References.

Vocabulary

Parent Namespaces

Given a string with a constant path we define its parent namespace to be the string that results from removing its rightmost segment.

For example, the parent namespace of the string "A::B::C" is the string "A::B", the parent namespace of "A::B" is "A", and the parent namespace of "A" is "".

The interpretation of a parent namespace when thinking about classes and modules is tricky though. Let's consider a module M named "A::B":

• The parent namespace, "A", may not reflect nesting at a given spot.

• The constant A may no longer exist, some code could have removed it from Object.

• If A exists, the class or module that was originally in A may not be there anymore. For example, if after a constant removal there was another constant assignment there would generally be a different object in there.

• In such case, it could even happen that the reassigned A held a new class or module called also "A"!

• In the previous scenarios M would no longer be reachable through A::B but the module object itself could still be alive somewhere and its name would still be "A::B".

The idea of a parent namespace is at the core of the autoloading algorithms and helps explain and understand their motivation intuitively, but as you see that metaphor leaks easily. Given an edge case to reason about, take always into account that by "parent namespace" the guide means exactly that specific string derivation.

Loading Mechanism

Rails autoloads files with Kernel#load when config.cache_classes is false, the default in development mode, and with Kernel#require otherwise, the default in production mode.

Kernel#load allows Rails to execute files more than once if constant reloading is enabled.

This guide uses the word "load" freely to mean a given file is interpreted, but the actual mechanism can be Kernel#load or Kernel#require depending on that flag.

Autoloading Availability

Rails is always able to autoload provided its environment is in place. For example the runner command autoloads:

$ rails runner 'p User.column_names'
["id", "email", "created_at", "updated_at"]

The console autoloads, the test suite autoloads, and of course the application autoloads.

By default, Rails eager loads the application files when it boots in production mode, so most of the autoloading going on in development does not happen. But autoloading may still be triggered during eager loading.

For example, given

class BeachHouse < House
end

if House is still unknown when app/models/beach_house.rb is being eager loaded, Rails autoloads it.

autoload_paths and eager_load_paths

As you probably know, when require gets a relative file name:

require 'erb'

Ruby looks for the file in the directories listed in $LOAD_PATH. That is, Ruby iterates over all its directories and for each one of them checks whether they have a file called "erb.rb", or "erb.so", or "erb.o", or "erb.dll". If it finds any of them, the interpreter loads it and ends the search. Otherwise, it tries again in the next directory of the list. If the list gets exhausted, LoadError is raised.

We are going to cover how constant autoloading works in more detail later, but the idea is that when a constant like Post is hit and missing, if there's a post.rb file for example in app/models Rails is going to find it, evaluate it, and have Post defined as a side-effect.

All right, Rails has a collection of directories similar to $LOAD_PATH in which to look up post.rb. That collection is called autoload_paths and by default it contains:

• All subdirectories of app in the application and engines present at boot time. For example, app/controllers. They do not need to be the default ones, any custom directories like app/workers belong automatically to autoload_paths.

• Any existing second level directories called app/*/concerns in the application and engines.

• The directory test/mailers/previews.

eager_load_paths is initially the app paths above

How files are autoloaded depends on eager_load and cache_classes config settings which typically vary in development, production, and test modes:

• In development, you want quicker startup with incremental loading of application code. So eager_load should be set to false, and Rails will autoload files as needed (see Autoloading Algorithms below) -- and then reload them when they change (see Constant Reloading below).
• In production, however, you want consistency and thread-safety and can live with a longer boot time. So eager_load is set to true, and then during boot (before the app is ready to receive requests) Rails loads all files in the eager_load_paths and then turns off auto loading (NB: autoloading may be needed during eager loading). Not autoloading after boot is a good thing, as autoloading can cause the app to be have thread-safety problems.
• In test, for speed of execution (of individual tests) eager_load is false, so Rails follows development behaviour.

What is described above are the defaults with a newly generated Rails app. There are multiple ways this can be configured differently (see Configuring Rails Applications. ). But using autoload_paths on its own in the past (before Rails 5) developers might configure autoload_paths to add in extra locations (e.g. lib which used to be an autoload path list years ago, but no longer is). However this is now discouraged for most purposes, as it is likely to lead to production-only errors. It is possible to add new locations to both config.eager_load_paths and config.autoload_paths but use at your own risk.

See also Autoloading in the Test Environment.

config.autoload_paths is not changeable from environment-specific configuration files.

The value of autoload_paths can be inspected. In a just-generated application it is (edited):

$ rails r 'puts ActiveSupport::Dependencies.autoload_paths'
.../app/assets
.../app/channels
.../app/controllers
.../app/controllers/concerns
.../app/helpers
.../app/jobs
.../app/mailers
.../app/models
.../app/models/concerns
.../activestorage/app/assets
.../activestorage/app/controllers
.../activestorage/app/javascript
.../activestorage/app/jobs
.../activestorage/app/models
.../actioncable/app/assets
.../actionview/app/assets
.../test/mailers/previews

INFO. autoload_paths is computed and cached during the initialization process. The application needs to be restarted to reflect any changes in the directory structure.

Autoloading Algorithms

Relative References

A relative constant reference may appear in several places, for example, in

class PostsController < ApplicationController
  def index
    @posts = Post.all
  end
end

all three constant references are relative.

Constants after the class and module Keywords

Ruby performs a lookup for the constant that follows a class or module keyword because it needs to know if the class or module is going to be created or reopened.

If the constant is not defined at that point it is not considered to be a missing constant, autoloading is not triggered.

So, in the previous example, if PostsController is not defined when the file is interpreted Rails autoloading is not going to be triggered, Ruby will just define the controller.

Top-Level Constants

On the contrary, if ApplicationController is unknown, the constant is considered missing and an autoload is going to be attempted by Rails.

In order to load ApplicationController, Rails iterates over autoload_paths. First it checks if app/assets/application_controller.rb exists. If it does not, which is normally the case, it continues and finds app/controllers/application_controller.rb.

If the file defines the constant ApplicationController all is fine, otherwise LoadError is raised:

unable to autoload constant ApplicationController, expected
<full path to application_controller.rb> to define it (LoadError)

INFO. Rails does not require the value of autoloaded constants to be a class or module object. For example, if the file app/models/max_clients.rb defines MAX_CLIENTS = 100 autoloading MAX_CLIENTS works just fine.

Namespaces

Autoloading ApplicationController looks directly under the directories of autoload_paths because the nesting in that spot is empty. The situation of Post is different, the nesting in that line is [PostsController] and support for namespaces comes into play.

The basic idea is that given

module Admin
  class BaseController < ApplicationController
    @@all_roles = Role.all
  end
end

to autoload Role we are going to check if it is defined in the current or parent namespaces, one at a time. So, conceptually we want to try to autoload any of

Admin::BaseController::Role
Admin::Role
Role

in that order. That's the idea. To do so, Rails looks in autoload_paths respectively for file names like these:

admin/base_controller/role.rb
admin/role.rb
role.rb

modulus some additional directory lookups we are going to cover soon.

INFO. 'Constant::Name'.underscore gives the relative path without extension of the file name where Constant::Name is expected to be defined.

Let's see how Rails autoloads the Post constant in the PostsController above assuming the application has a Post model defined in app/models/post.rb.

First it checks for posts_controller/post.rb in autoload_paths:

app/assets/posts_controller/post.rb
app/controllers/posts_controller/post.rb
app/helpers/posts_controller/post.rb
...
test/mailers/previews/posts_controller/post.rb

Since the lookup is exhausted without success, a similar search for a directory is performed, we are going to see why in the next section:

app/assets/posts_controller/post
app/controllers/posts_controller/post
app/helpers/posts_controller/post
...
test/mailers/previews/posts_controller/post

If all those attempts fail, then Rails starts the lookup again in the parent namespace. In this case only the top-level remains:

app/assets/post.rb
app/controllers/post.rb
app/helpers/post.rb
app/mailers/post.rb
app/models/post.rb

A matching file is found in app/models/post.rb. The lookup stops there and the file is loaded. If the file actually defines Post all is fine, otherwise LoadError is raised.

Qualified References

When a qualified constant is missing Rails does not look for it in the parent namespaces. But there is a caveat: when a constant is missing, Rails is unable to tell if the trigger was a relative reference or a qualified one.

For example, consider

module Admin
  User
end

and

Admin::User

If User is missing, in either case all Rails knows is that a constant called "User" was missing in a module called "Admin".

If there is a top-level User Ruby would resolve it in the former example, but wouldn't in the latter. In general, Rails does not emulate the Ruby constant resolution algorithms, but in this case it tries using the following heuristic:

If none of the parent namespaces of the class or module has the missing constant then Rails assumes the reference is relative. Otherwise qualified.

For example, if this code triggers autoloading

Admin::User

and the User constant is already present in Object, it is not possible that the situation is

module Admin
  User
end

because otherwise Ruby would have resolved User and no autoloading would have been triggered in the first place. Thus, Rails assumes a qualified reference and considers the file admin/user.rb and directory admin/user to be the only valid options.

In practice, this works quite well as long as the nesting matches all parent namespaces respectively and the constants that make the rule apply are known at that time.

However, autoloading happens on demand. If by chance the top-level User was not yet loaded, then Rails assumes a relative reference by contract.

Naming conflicts of this kind are rare in practice, but if one occurs, require_dependency provides a solution by ensuring that the constant needed to trigger the heuristic is defined in the conflicting place.

Automatic Modules

When a module acts as a namespace, Rails does not require the application to define a file for it, a directory matching the namespace is enough.

Suppose an application has a back office whose controllers are stored in app/controllers/admin. If the Admin module is not yet loaded when Admin::UsersController is hit, Rails needs first to autoload the constant Admin.

If autoload_paths has a file called admin.rb Rails is going to load that one, but if there's no such file and a directory called admin is found, Rails creates an empty module and assigns it to the Admin constant on the fly.

Generic Procedure

Relative references are reported to be missing in the cref where they were hit, and qualified references are reported to be missing in their parent (see Resolution Algorithm for Relative Constants at the beginning of this guide for the definition of cref, and Resolution Algorithm for Qualified Constants for the definition of parent).

The procedure to autoload constant C in an arbitrary situation is as follows:

if the class or module in which C is missing is Object
  let ns = ''
else
  let M = the class or module in which C is missing

  if M is anonymous
    let ns = ''
  else
    let ns = M.name
  end
end

loop do
  # Look for a regular file.
  for dir in autoload_paths
    if the file "#{dir}/#{ns.underscore}/c.rb" exists
      load/require "#{dir}/#{ns.underscore}/c.rb"

      if C is now defined
        return
      else
        raise LoadError
      end
    end
  end

  # Look for an automatic module.
  for dir in autoload_paths
    if the directory "#{dir}/#{ns.underscore}/c" exists
      if ns is an empty string
        let C = Module.new in Object and return
      else
        let C = Module.new in ns.constantize and return
      end
    end
  end

  if ns is empty
    # We reached the top-level without finding the constant.
    raise NameError
  else
    if C exists in any of the parent namespaces
      # Qualified constants heuristic.
      raise NameError
    else
      # Try again in the parent namespace.
      let ns = the parent namespace of ns and retry
    end
  end
end

require_dependency

Constant autoloading is triggered on demand and therefore code that uses a certain constant may have it already defined or may trigger an autoload. That depends on the execution path and it may vary between runs.

There are times, however, in which you want to make sure a certain constant is known when the execution reaches some code. require_dependency provides a way to load a file using the current loading mechanism, and keeping track of constants defined in that file as if they were autoloaded to have them reloaded as needed.

require_dependency is rarely needed, but see a couple of use-cases in Autoloading and STI and When Constants aren't Triggered.

WARNING. Unlike autoloading, require_dependency does not expect the file to define any particular constant. Exploiting this behavior would be a bad practice though, file and constant paths should match.

Constant Reloading

When config.cache_classes is false Rails is able to reload autoloaded constants.

For example, if you're in a console session and edit some file behind the scenes, the code can be reloaded with the reload! command:

> reload!

When the application runs, code is reloaded when something relevant to this logic changes. In order to do that, Rails monitors a number of things:

• config/routes.rb.

• Locales.

• Ruby files under autoload_paths.

• db/schema.rb and db/structure.sql.

If anything in there changes, there is a middleware that detects it and reloads the code.

Autoloading keeps track of autoloaded constants. Reloading is implemented by removing them all from their respective classes and modules using Module#remove_const. That way, when the code goes on, those constants are going to be unknown again, and files reloaded on demand.

INFO. This is an all-or-nothing operation, Rails does not attempt to reload only what changed since dependencies between classes makes that really tricky. Instead, everything is wiped.

Module#autoload isn't Involved

Module#autoload provides a lazy way to load constants that is fully integrated with the Ruby constant lookup algorithms, dynamic constant API, etc. It is quite transparent.

Rails internals make extensive use of it to defer as much work as possible from the boot process. But constant autoloading in Rails is not implemented with Module#autoload.

One possible implementation based on Module#autoload would be to walk the application tree and issue autoload calls that map existing file names to their conventional constant name.

There are a number of reasons that prevent Rails from using that implementation.

For example, Module#autoload is only capable of loading files using require, so reloading would not be possible. Not only that, it uses an internal require which is not Kernel#require.

Then, it provides no way to remove declarations in case a file is deleted. If a constant gets removed with Module#remove_const its autoload is not triggered again. Also, it doesn't support qualified names, so files with namespaces should be interpreted during the walk tree to install their own autoload calls, but those files could have constant references not yet configured.

An implementation based on Module#autoload would be awesome but, as you see, at least as of today it is not possible. Constant autoloading in Rails is implemented with Module#const_missing, and that's why it has its own contract, documented in this guide.

Common Gotchas

Nesting and Qualified Constants

Let's consider

module Admin
  class UsersController < ApplicationController
    def index
      @users = User.all
    end
  end
end

and

class Admin::UsersController < ApplicationController
  def index
    @users = User.all
  end
end

To resolve User Ruby checks Admin in the former case, but it does not in the latter because it does not belong to the nesting (see Nesting and Resolution Algorithms).

Unfortunately Rails autoloading does not know the nesting in the spot where the constant was missing and so it is not able to act as Ruby would. In particular, Admin::User will get autoloaded in either case.

Albeit qualified constants with class and module keywords may technically work with autoloading in some cases, it is preferable to use relative constants instead:

module Admin
  class UsersController < ApplicationController
    def index
      @users = User.all
    end
  end
end

Autoloading and STI

Single Table Inheritance (STI) is a feature of Active Record that enables storing a hierarchy of models in one single table. The API of such models is aware of the hierarchy and encapsulates some common needs. For example, given these classes:

# app/models/polygon.rb
class Polygon < ApplicationRecord
end

# app/models/triangle.rb
class Triangle < Polygon
end

# app/models/rectangle.rb
class Rectangle < Polygon
end

Triangle.create creates a row that represents a triangle, and Rectangle.create creates a row that represents a rectangle. If id is the ID of an existing record, Polygon.find(id) returns an object of the correct type.

Methods that operate on collections are also aware of the hierarchy. For example, Polygon.all returns all the records of the table, because all rectangles and triangles are polygons. Active Record takes care of returning instances of their corresponding class in the result set.

Types are autoloaded as needed. For example, if Polygon.first is a rectangle and Rectangle has not yet been loaded, Active Record autoloads it and the record is correctly instantiated.

All good, but if instead of performing queries based on the root class we need to work on some subclass, things get interesting.

While working with Polygon you do not need to be aware of all its descendants, because anything in the table is by definition a polygon, but when working with subclasses Active Record needs to be able to enumerate the types it is looking for. Let's see an example.

Rectangle.all only loads rectangles by adding a type constraint to the query:

SELECT "polygons".* FROM "polygons"
WHERE "polygons"."type" IN ("Rectangle")

Let's introduce now a subclass of Rectangle:

# app/models/square.rb
class Square < Rectangle
end

Rectangle.all should now return rectangles and squares:

SELECT "polygons".* FROM "polygons"
WHERE "polygons"."type" IN ("Rectangle", "Square")

But there's a caveat here: How does Active Record know that the class Square exists at all?

Even if the file app/models/square.rb exists and defines the Square class, if no code yet used that class, Rectangle.all issues the query

SELECT "polygons".* FROM "polygons"
WHERE "polygons"."type" IN ("Rectangle")

That is not a bug, the query includes all known descendants of Rectangle.

A way to ensure this works correctly regardless of the order of execution is to manually load the direct subclasses at the bottom of the file that defines each intermediate class:

# app/models/rectangle.rb
class Rectangle < Polygon
end
require_dependency 'square'

This needs to happen for every intermediate (non-root and non-leaf) class. The root class does not scope the query by type, and therefore does not necessarily have to know all its descendants.

Autoloading and require

Files defining constants to be autoloaded should never be required:

require 'user' # DO NOT DO THIS

class UsersController < ApplicationController
  ...
end

There are two possible gotchas here in development mode:

1. If User is autoloaded before reaching the require, app/models/user.rb runs again because load does not update $LOADED_FEATURES.

2. If the require runs first Rails does not mark User as an autoloaded constant and changes to app/models/user.rb aren't reloaded.

Just follow the flow and use constant autoloading always, never mix autoloading and require. As a last resort, if some file absolutely needs to load a certain file use require_dependency to play nice with constant autoloading. This option is rarely needed in practice, though.

Of course, using require in autoloaded files to load ordinary 3rd party libraries is fine, and Rails is able to distinguish their constants, they are not marked as autoloaded.

Autoloading and Initializers

Consider this assignment in config/initializers/set_auth_service.rb:

AUTH_SERVICE = if Rails.env.production?
  RealAuthService
else
  MockedAuthService
end

The purpose of this setup would be that the application uses the class that corresponds to the environment via AUTH_SERVICE. In development mode MockedAuthService gets autoloaded when the initializer runs. Let's suppose we do some requests, change its implementation, and hit the application again. To our surprise the changes are not reflected. Why?

As we saw earlier, Rails removes autoloaded constants, but AUTH_SERVICE stores the original class object. Stale, non-reachable using the original constant, but perfectly functional.

The following code summarizes the situation:

class C
  def quack
    'quack!'
  end
end

X = C
Object.instance_eval { remove_const(:C) }
X.new.quack # => quack!
X.name      # => C
C           # => uninitialized constant C (NameError)

Because of that, it is not a good idea to autoload constants on application initialization.

In the case above we could implement a dynamic access point:

# app/models/auth_service.rb
class AuthService
  if Rails.env.production?
    def self.instance
      RealAuthService
    end
  else
    def self.instance
      MockedAuthService
    end
  end
end

and have the application use AuthService.instance instead. AuthService would be loaded on demand and be autoload-friendly.

require_dependency and Initializers

As we saw before, require_dependency loads files in an autoloading-friendly way. Normally, though, such a call does not make sense in an initializer.

One could think about doing some require_dependency calls in an initializer to make sure certain constants are loaded upfront, for example as an attempt to address the gotcha with STIs.

Problem is, in development mode autoloaded constants are wiped if there is any relevant change in the file system. If that happens then we are in the very same situation the initializer wanted to avoid!

Calls to require_dependency have to be strategically written in autoloaded spots.

When Constants aren't Missed

Relative References

Let's consider a flight simulator. The application has a default flight model

# app/models/flight_model.rb
class FlightModel
end

that can be overridden by each airplane, for instance

# app/models/bell_x1/flight_model.rb
module BellX1
  class FlightModel < FlightModel
  end
end

# app/models/bell_x1/aircraft.rb
module BellX1
  class Aircraft
    def initialize
      @flight_model = FlightModel.new
    end
  end
end

The initializer wants to create a BellX1::FlightModel and nesting has BellX1, that looks good. But if the default flight model is loaded and the one for the Bell-X1 is not, the interpreter is able to resolve the top-level FlightModel and autoloading is thus not triggered for BellX1::FlightModel.

That code depends on the execution path.

These kind of ambiguities can often be resolved using qualified constants:

module BellX1
  class Plane
    def flight_model
      @flight_model ||= BellX1::FlightModel.new
    end
  end
end

Also, require_dependency is a solution:

require_dependency 'bell_x1/flight_model'

module BellX1
  class Plane
    def flight_model
      @flight_model ||= FlightModel.new
    end
  end
end

Qualified References

WARNING. This gotcha is only possible in Ruby < 2.5.

Given

# app/models/hotel.rb
class Hotel
end

# app/models/image.rb
class Image
end

# app/models/hotel/image.rb
class Hotel
  class Image < Image
  end
end

the expression Hotel::Image is ambiguous because it depends on the execution path.

As we saw before, Ruby looks up the constant in Hotel and its ancestors. If app/models/image.rb has been loaded but app/models/hotel/image.rb hasn't, Ruby does not find Image in Hotel, but it does in Object:

$ rails r 'Image; p Hotel::Image' 2>/dev/null
Image # NOT Hotel::Image!

The code evaluating Hotel::Image needs to make sure app/models/hotel/image.rb has been loaded, possibly with require_dependency.

In these cases the interpreter issues a warning though:

warning: toplevel constant Image referenced by Hotel::Image

This surprising constant resolution can be observed with any qualifying class:

2.1.5 :001 > String::Array
(irb):1: warning: toplevel constant Array referenced by String::Array
 => Array

WARNING. To find this gotcha the qualifying namespace has to be a class, Object is not an ancestor of modules.

Autoloading within Singleton Classes

Let's suppose we have these class definitions:

# app/models/hotel/services.rb
module Hotel
  class Services
  end
end

# app/models/hotel/geo_location.rb
module Hotel
  class GeoLocation
    class << self
      Services
    end
  end
end

If Hotel::Services is known by the time app/models/hotel/geo_location.rb is being loaded, Services is resolved by Ruby because Hotel belongs to the nesting when the singleton class of Hotel::GeoLocation is opened.

But if Hotel::Services is not known, Rails is not able to autoload it, the application raises NameError.

The reason is that autoloading is triggered for the singleton class, which is anonymous, and as we saw before, Rails only checks the top-level namespace in that edge case.

An easy solution to this caveat is to qualify the constant:

module Hotel
  class GeoLocation
    class << self
      Hotel::Services
    end
  end
end

Autoloading in BasicObject

Direct descendants of BasicObject do not have Object among their ancestors and cannot resolve top-level constants:

class C < BasicObject
  String # NameError: uninitialized constant C::String
end

When autoloading is involved that plot has a twist. Let's consider:

class C < BasicObject
  def user
    User # WRONG
  end
end

Since Rails checks the top-level namespace User gets autoloaded just fine the first time the user method is invoked. You only get the exception if the User constant is known at that point, in particular in a second call to user:

c = C.new
c.user # surprisingly fine, User
c.user # NameError: uninitialized constant C::User

because it detects that a parent namespace already has the constant (see Qualified References).

As with pure Ruby, within the body of a direct descendant of BasicObject use always absolute constant paths:

class C < BasicObject
  ::String # RIGHT

  def user
    ::User # RIGHT
  end
end

Autoloading in the Test Environment

When configuring the test environment for autoloading you might consider multiple factors.

For example it might be worth running your tests with an identical setup to production (config.eager_load = true, config.cache_classes = true) in order to catch any problems before they hit production (this is compensation for the lack of dev-prod parity). However this will slow down the boot time for individual tests on a dev machine (and is not immediately compatible with spring see below). So one possibility is to do this on a CI machine only (which should run without spring).

On a development machine you can then have your tests running with whatever is fastest (ideally config.eager_load = false).

With the Spring pre-loader (included with new Rails apps), you ideally keep config.eager_load = false as per development. Sometimes you may end up with a hybrid configuration (config.eager_load = true, config.cache_classes = true AND config.enable_dependency_loading = true), see spring issue. However it might be simpler to keep the same configuration as development, and work out whatever it is that is causing autoloading to fail (perhaps by the results of your CI test results).

Occasionally you may need to explicitly eager_load by using Rails .application.eager_load! in the setup of your tests -- this might occur if your tests involve multithreading.

Troubleshooting

Tracing Autoloads

Active Support is able to report constants as they are autoloaded. To enable these traces in a Rails application, put the following two lines in some initializer:

ActiveSupport::Dependencies.logger = Rails.logger
ActiveSupport::Dependencies.verbose = true

Where is a Given Autoload Triggered?

If constant Foo is being autoloaded, and you'd like to know where is that autoload coming from, just throw

puts caller

at the top of foo.rb and inspect the printed stack trace.

Which Constants Have Been Autoloaded?

At any given time,

ActiveSupport::Dependencies.autoloaded_constants

has the collection of constants that have been autoloaded so far.