/autoload_reloader

Experimental implementation of code reloading using Ruby's autoload

Primary LanguageRubyMIT LicenseMIT

AutoloadReloader

An alternative to the rails autoloader (ActiveSupport::Dependencies) that is based on Module#autoload instead of const_missing. As the Rails Autoloading and Reloading Constants guide admits:

An implementation based on Module#autoload would be awesome

Installation

Add this line to your application's Gemfile to use it by itself

gem 'autoload_reloader', github: 'Shopify/autoload_reloader'

Or add this line to the application's Gemfile to replace the const_missing based autoloading in rails' activesupport gem.

gem 'autoload_reloader', github: 'Shopify/autoload_reloader', require: 'autoload_reloader/active_support_ext'

And then execute:

$ bundle

Usage

This library can be used by itself as follows

require 'autoload_reloader'

File.write "foo.rb", "Foo = 1"

# sets up autoloads by scanning the file system
AutoloadReloader::Paths.push(Dir.pwd)

# start using autoloaded constants
Foo # => 1

File.write "foo.rb", "Foo = 2"

# unload constants and re-scan paths to autoload them again
AutoloadReloader.reload
Foo # => 2

# load all autoloadable constants in the paths
AutoloadReloader.eager_load

How It Works

The Module#autoload isn't Involved section of the Rails Autoloading and Reloading Constants guide makes it sound like this gem shouldn't work

An implementation based on Module#autoload would be awesome but, as you see, at least as of today it is not possible.

so let's dispell some myths and explain how this gem works.

When the paths are changed through Array like methods on AutoloadReloader::Paths, this gem walks the file system as the rails guide suggested

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.

So let's start addressing the problems previously ran into with this approach

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.

require will load a file again if the file's path is removed from $LOADED_FEATURES, so this is done when AutoloadReloader.reload is called.

Not only that, it uses an internal require which is not Kernel#require.

This was true before MRI ruby 2.3.0, but it was changed to use a normal ruby require method call , so Kernel#require can be wrapped to hook into constant autoloading through Module#autoload.

Then, it provides no way to remove declarations in case a file is deleted.

A constant is considered defined (in a special autoload state) after using Module#autoload before the file itself it loaded, so the autoload can be removed using Module#remove_const when the file is deleted.

If a constant gets removed with Module#remove_const its autoload is not triggered again.

Ruby doesn't keep track of the autoload after the constant is loaded, so Module#autoload must be used again after removing the constant for reloading.

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.

Since we can't register autoloads under an autoload namespace with Module#autoload, this gem keeps track of the autoloads under an unloaded namespace to register when that namespace is loaded.

A TracePoint for the :class event is used to detect namespace loaded using the module or class keywords. This TracePoint will miss constants defined using Class.new or Module.new that are assigned as a constant, in which case this gem detects that the constant is loaded at the end of the require for that constant using this gem's require wrapper.

Why is using Module#autoload awesome compared to const_missing

Using Module#autoload avoids several common gotchas from using const_missing .

In general, using const_missing leads to various forms of load dependent code when using namespaced code.

Avoid Accidentally using a Constant in a Parent Namespace

For example, if you have the code

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

and there are autoloadable ::User and ::Admin::User constants, then without using Module#autoload, the code would use ::User if it has been loaded and ::Admin::User has not yet been loaded.

This isn't a problem with Module#autoload, because ruby finds the ::Admin::User constant when doing the constant resolution and autoloads the constant rather than using ::User.

Rails recommends using a qualified constant reference (Admin::User in this case) to avoid the ambiguity, but failing to do so results in subtle load order dependent bug. Qualified constant references also don't work well with constant privacy.

Works with Constant Privacy

When a constant with the same name is in a namespace and a parent namespace, then Rails recommends to use a qualified constant to avoid accidentally using the constant in the parent namespace. However, this doesn't work if the constant in the inner namespace is a private constant, because it is internal to that inner namespace.

For example, you may make Admin::User private

module Admin
  class User < ApplicationRecord
  end
  private_constant :User
end

but now the code

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

will get a NameError: private constant Admin::User referenced error from the constant reference Admin::User if Admin::User is already loaded, since ruby prevents private constants from being referenced in this way.

When using this gem, Admin::User.all will consistently get the above mentioned NameError and can simply be replaced with User.all and it will work reliably.

When using Rails' const_missing based autoloader, Admin::User reference would work the first time if neither ::User nor Admin::User were already loaded, regardless of the namespace it is called from, since const_missing doesn't know how the constant was referenced. This differing behaviour on first use adds extra developer confusion. The next time it comes to the constant reference it will get the above mentioned NameError, but then it can't be fixed by changing the code to User.all since that could end up doing ::User.all if ::User has been loaded by Admin::User hasn't.

Always Respects Module Nesting

Ruby relative constant resolution works by looking for the named constant directly under each module in Module.nesting, which is affected by the nesting from module or class keywords.

For example,

module Admin
  class UsersTest < ActiveSupport::TestCase
    Module.nesting # => [Admin::UsersTest, Admin]
    test "new" do
      assert User.new
    end
  end
end

would use Admin::User if it were defined and Admin::UsersTest::User weren't also defined, but the following code

class Admin::UsersTest < ActiveSupport::TestCase
  Module.nesting # => [Admin::UsersTest]
  test "new" do
    assert User.new
  end
end

would not look for Admin::User in ruby's constant resolution since the only constant opened with class or module is Admin::UsersTest.

const_missing can't use Module.nesting to reproduce ruby's algorithm since it would return the nesting for the code in const_missing rather than the code referencing the constant before const_missing was called. As such, rails assumes and recommends the former style is used, so const_missing would return Admin::User even in the latter case inside class Admin::UserTest.

Avoid Preventing Autoload by Opening a Namespace

When autoloading based on const_missing, opening a namespace can define a constant and prevent it from being autoloaded.

For example, by following the Rails recommendation on using nested namespaces in tests can lead to code like

require 'test_helper'

module Admin
  class UsersTest < ActiveSupport::TestCase
    # ...
  end
end

If there were an admin.rb file on the autoload path, then loading the above test file with the Rails' autoloader would cause module Admin to define the constant and the Admin constant would never be missing to load admin.rb through const_missing.

In contrast, with Module#autoload based autoloading, ruby will autoload admin.rb when encoutering module Admin then will re-open the constant defined in admin.rb.

Works with require

Rails warns you to never require an autoloaded constant, since it prevents the constant from being reloaded. This isn't a problem with this gem. If you want to require a file without an absolute path (e.g. require 'user') you simply need to make sure it can be found on the $LOAD_PATH which isn't done automatically by using AutoloadReloader::Paths but is done for autoload paths in rails.

With this gem's Active Support integration, require_dependency is supported for compatibility, but can be safely replaced with require where still needed (e.g. for Single Table Inheritance ).

Limitations

There are some limitations of this gem to be aware of.

Non-MRI Ruby Implementation Support

Although MRI no longer uses an internal require to load autoload constants, other ruby implementations still appear to use an internal require as older version of MRI ruby did.

I have noticed this internal require autoload behaviour for:

  • JRuby 9.1.16.0 (although support is on the way)
  • Rubinius 3.100
  • TruffleRuby 0.33

Non-CamelCased Constants

The Rails' autoloader infers the filename from the constant name (using String#underscore) so is able to work with UPPER_CASE named constants.

In contrast, this gem infers the constant name from the filename, so filenames like foo_bar.rb should always define FooBar and not FOO_BAR. This seems fine in practice though, since module and class naming convention is already to use CamelCase and value constants are generally too small to merit their own file. If this is problem in practice, then please open an issue describing the use case.

Nested References in File Assigning Namespace Constant

For example, if the autoloaded foo.rb file contains

Foo = Module.new
Foo.module_eval do
  Foo::Bar
  # ...
end

then it won't have a chance to setup autoloads on the Foo namespace before Foo::Bar is referenced. This is because Foo = Module.new doesn't trigger a :class TracePoint event. Using module Foo to define the namespace avoids this problem.

However, this is a contrived example, so if you have a real example then please open an issue describing the use case.

Development

After checking out the repo, run bin/setup to install dependencies. Then, run rake test to run the tests. You can also run bin/console for an interactive prompt that will allow you to experiment.

To install this gem onto your local machine, run bundle exec rake install. To release a new version, update the version number in version.rb, and then run bundle exec rake release, which will create a git tag for the version, push git commits and tags, and push the .gem file to rubygems.org.

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/Shopify/autoload_reloader.

License

The gem is available as open source under the terms of the MIT License.