Skip to content
/ zeitwerk Public
forked from fxn/zeitwerk

Efficient and thread-safe code loader for Ruby

License

MIT license
0 stars 118 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

leehericks/zeitwerk

BranchesTags
 
 

Repository files navigation

Zeitwerk

Gem Version Build Status

Introduction

Zeitwerk is an efficient and thread-safe code loader for Ruby.

Given a conventional file structure, Zeitwerk is able to load your project's classes and modules on demand (autoloading), or upfront (eager loading). You don't need to write require calls for your own files, rather, you can streamline your programming knowing that your classes and modules are available everywhere. This feature is efficient, thread-safe, and matches Ruby's semantics for constants.

Zeitwerk is also able to reload code, which may be handy while developing web applications. Coordination is needed to reload in a thread-safe manner. The documentation below explains how to do this.

The gem is designed so that any project, gem dependency, application, etc. can have their own independent loader, coexisting in the same process, managing their own project trees, and independent of each other. Each loader has its own configuration, inflector, and optional logger.

Internally, Zeitwerk issues require calls exclusively using absolute file names, so there are no costly file system lookups in $LOAD_PATH. Technically, the directories managed by Zeitwerk do not even need to be in $LOAD_PATH.

Furthermore, Zeitwerk does at most one single scan of the project tree, and it descends into subdirectories lazily, only if their namespaces are used.

Synopsis

Main interface for gems:

# lib/my_gem.rb (main file)

require "zeitwerk"
loader = Zeitwerk::Loader.for_gem
loader.setup # ready!

module MyGem
  # ...
end

loader.eager_load # optionally

Main generic interface:

loader = Zeitwerk::Loader.new
loader.push_dir(...)
loader.setup # ready!

The loader variable can go out of scope. Zeitwerk keeps a registry with all of them, and so the object won't be garbage collected.

You can reload if you want to:

loader = Zeitwerk::Loader.new
loader.push_dir(...)
loader.enable_reloading # you need to opt-in before setup
loader.setup
...
loader.reload

and you can eager load all the code:

loader.eager_load

It is also possible to broadcast eager_load to all instances:

Zeitwerk::Loader.eager_load_all

File structure

The idea: File paths match constant paths

To have a file structure Zeitwerk can work with, just name files and directories after the name of the classes and modules they define:

lib/my_gem.rb         -> MyGem
lib/my_gem/foo.rb     -> MyGem::Foo
lib/my_gem/bar_baz.rb -> MyGem::BarBaz
lib/my_gem/woo/zoo.rb -> MyGem::Woo::Zoo

You can tune that a bit by collapsing directories, or by ignoring parts of the project, but that is the main idea.

Inner simple constants

While a simple constant like HttpCrawler::MAX_RETRIES can be defined in its own file:

# http_crawler/max_retries.rb
HttpCrawler::MAX_RETRIES = 10

that is not required, you can also define it the regular way:

# http_crawler.rb
class HttpCrawler
  MAX_RETRIES = 10
end

The first example needs a custom inflection rule:

loader.inflector.inflect("max_retries" => "MAX_RETRIES")

Otherwise, Zeitwerk would expect the file to define MaxRetries.

In the second example, no custom rule is needed.

Root directories and root namespaces

Every directory configured with push_dir is called a root directory, and they represent root namespaces.

The default root namespace is Object

By default, the namespace associated to a root directory is the top-level one: Object.

For example, given

loader.push_dir("#{__dir__}/models")
loader.push_dir("#{__dir__}/serializers"))

these are the expected classes and modules being defined by these files:

models/user.rb                 -> User
serializers/user_serializer.rb -> UserSerializer

Custom root namespaces

While Object is by far the most common root namespace, you can associate a different one to a particular root directory. The method push_dir accepts a non-anonymous class or module object in the optional namespace keyword argument.

For example, given:

require "active_job"
require "active_job/queue_adapters"
loader.push_dir("#{__dir__}/adapters", namespace: ActiveJob::QueueAdapters)

a file defining ActiveJob::QueueAdapters::MyQueueAdapter does not need the conventional parent directories, you can (and have to) store the file directly below adapters:

adapters/my_queue_adapter.rb -> ActiveJob::QueueAdapters::MyQueueAdapter

Please, note that the given root namespace must be non-reloadable, though autoloaded constants in that namespace can be. That is, if you associate app/api with an existing Api module, that module should not be reloadable. However, if the project defines and autoloads the class Api::Deliveries, that one can be reloaded.

Nested root directories

Root directories should not be ideally nested, but Zeitwerk supports them because in Rails, for example, both app/models and app/models/concerns belong to the autoload paths.

Zeitwerk detects nested root directories, and treats them as roots only. In the example above, concerns is not considered to be a namespace below app/models. For example, the file:

app/models/concerns/geolocatable.rb

should define Geolocatable, not Concerns::Geolocatable.

Implicit namespaces

If a namespace is just a simple module with no code, you do not need to define it in a file: Directories without a matching Ruby file get modules created automatically on your behalf.

For example, if a project has an admin directory:

app/controllers/admin/users_controller.rb -> Admin::UsersController

and does not have a file called admin.rb, Zeitwerk automatically creates an Admin module on your behalf the first time Admin is used.

For this to happen, the directory has to contain non-ignored Ruby files with extension .rb, directly or recursively, otherwise it is ignored. This condition is evaluated again on reloads.

Explicit namespaces

Classes and modules that act as namespaces can also be explicitly defined, though. For instance, consider

app/models/hotel.rb         -> Hotel
app/models/hotel/pricing.rb -> Hotel::Pricing

There, app/models/hotel.rb defines Hotel, and thus Zeitwerk does not autovivify a module.

The classes and modules from the namespace are already available in the body of the class or module defining it:

class Hotel < ApplicationRecord
  include Pricing # works
  ...
end

An explicit namespace must be managed by one single loader. Loaders that reopen namespaces owned by other projects are responsible for loading their constants before setup.

Collapsing directories

Say some directories in a project exist for organizational purposes only, and you prefer not to have them as namespaces. For example, the actions subdirectory in the next example is not meant to represent a namespace, it is there only to group all actions related to bookings:

booking.rb                -> Booking
booking/actions/create.rb -> Booking::Create

To make it work that way, configure Zeitwerk to collapse said directory:

loader.collapse("#{__dir__}/booking/actions")

This method accepts an arbitrary number of strings or Pathname objects, and also an array of them.

You can pass directories and glob patterns. Glob patterns are expanded when they are added, and again on each reload.

To illustrate usage of glob patterns, if actions in the example above is part of a standardized structure, you could use a wildcard:

loader.collapse("#{__dir__}/*/actions")

Testing compliance

When a managed file is loaded, Zeitwerk verifies the expected constant is defined. If it is not, Zeitwerk::NameError is raised.

So, an easy way to ensure compliance in the test suite is to eager load the project:

begin
  loader.eager_load(force: true)
rescue Zeitwerk::NameError => e
  flunk e.message
else
  assert true
end

Usage

Setup

Generic

Loaders are ready to load code right after calling setup on them:

loader.setup

This method is synchronized and idempotent.

Customization should generally be done before that call. In particular, in the generic interface you may set the root directories from which you want to load files:

loader.push_dir(...)
loader.push_dir(...)
loader.setup

for_gem

Zeitwerk::Loader.for_gem is a convenience shortcut for the common case in which a gem has its entry point directly under the lib directory:

lib/my_gem.rb         # MyGem
lib/my_gem/version.rb # MyGem::VERSION
lib/my_gem/foo.rb     # MyGem::Foo

Neither a gemspec nor a version file are technically required, this helper works as long as the code is organized using that standard structure.

Conceptually, for_gem translates to:

# lib/my_gem.rb

require "zeitwerk"
loader = Zeitwerk::Loader.new
loader.tag = File.basename(__FILE__, ".rb")
loader.inflector = Zeitwerk::GemInflector.new(__FILE__)
loader.push_dir(__dir__)

If the main module references project constants at the top-level, Zeitwerk has to be ready to load them. Their definitions, in turn, may reference other project constants. And this is recursive. Therefore, it is important that the setup call happens above the main module definition:

# lib/my_gem.rb (main file)

require "zeitwerk"
loader = Zeitwerk::Loader.for_gem
loader.setup

module MyGem
  # Since the setup has been performed, at this point we are already able
  # to reference project constants, in this case MyGem::MyLogger.
  include MyLogger
end

Loaders returned by Zeitwerk::Loader.for_gem issue warnings if lib has extra Ruby files or directories.

For example, if the gem has Rails generators under lib/generators, by convention that directory defines a Generators Ruby module. If generators is just a container for non-autoloadable code and templates, not acting as a project namespace, you need to setup things accordingly.

If the warning is legit, just tell the loader to ignore the offending file or directory:

loader.ignore("#{__dir__}/generators")

Otherwise, there's a flag to say the extra stuff is OK:

Zeitwerk::Loader.for_gem(warn_on_extra_files: false)

This method is idempotent when invoked from the same file, to support gems that want to reload (unlikely).

If the entry point of your gem lives in a subdirectory of lib because it is reopening a namespace defined somewhere else, please use the generic API to setup the loader, and make sure you check the section Reopening third-party namespaces down below.

Autoloading

After setup, you are able to reference classes and modules from the project without issuing require calls for them. They are all available everywhere, autoloading loads them on demand. This works even if the reference to the class or module is first hit in client code, outside your project.

Let's revisit the example above:

# lib/my_gem.rb (main file)

require "zeitwerk"
loader = Zeitwerk::Loader.for_gem
loader.setup

module MyGem
  include MyLogger # (*)
end

That works, and there is no require "my_gem/my_logger". When (*) is reached, Zeitwerk seamlessly autoloads MyGem::MyLogger.

If autoloading a file does not define the expected class or module, Zeitwerk raises Zeitwerk::NameError, which is a subclass of NameError.

Eager loading

Zeitwerk instances are able to eager load their managed files:

loader.eager_load

That skips ignored files and directories.

In gems, the method needs to be invoked after the main namespace has been defined, as shown in Synopsis.

Eager loading is synchronized and idempotent.

Attempting to eager load without previously calling setup raises Zeitwerk::SetupRequired.

Eager load exclusions

You can tell Zeitwerk that certain files or directories are autoloadable, but should not be eager loaded:

db_adapters = "#{__dir__}/my_gem/db_adapters"
loader.do_not_eager_load(db_adapters)
loader.setup
loader.eager_load # won't eager load the database adapters

However, that can be overridden with force:

loader.eager_load(force: true) # database adapters are eager loaded

Which may be handy if the project eager loads in the test suite to ensure project layout compliance.

The force flag does not affect ignored files and directories, those are still ignored.

Eager load directories

The method Zeitwerk::Loader#eager_load_dir eager loads a given directory, recursively:

loader.eager_load_dir("#{__dir__}/custom_web_app/routes")

This is useful when the loader is not eager loading the entire project, but you still need some subtree to be loaded for things to function properly.

Both strings and Pathname objects are supported as arguments. If the argument is not a directory managed by the receiver, the method raises Zeitwerk::Error.

Eager load exclusions, ignored files and directories, and shadowed files are not eager loaded.

Zeitwerk::Loader#eager_load_dir is idempotent, but compatible with reloading. If you eager load a directory and then reload, eager loading that directory will load its (current) contents again.

The method checks if a regular eager load was already executed, in which case it returns fast.

Nested root directories which are descendants of the argument are skipped. Those subtrees are considered to be conceptually apart.

Attempting to eager load a directory without previously calling setup raises Zeitwerk::SetupRequired.

Eager load namespaces

The method Zeitwerk::Loader#eager_load_namespace eager loads a given namespace, recursively:

loader.eager_load_namespace(MyApp::Routes)

This is useful when the loader is not eager loading the entire project, but you still need some namespace to be loaded for things to function properly.

The argument has to be a class or module object and the method raises Zeitwerk::Error otherwise.

If the namespace is spread over multiple directories in the receiver's source tree, they are all eager loaded. For example, if you have a structure like

root_dir1/my_app/routes
root_dir2/my_app/routes
root_dir3/my_app/routes

where root_directory{1,2,3} are root directories, eager loading MyApp::Routes will eager load the contents of the three corresponding directories.

There might exist external source trees implementing part of the namespace. This happens routinely, because top-level constants are stored in the globally shared Object. It happens also when deliberately reopening third-party namespaces. Such external code is not eager loaded, the implementation is carefully scoped to what the receiver manages to avoid side-effects elsewhere.

This method is flexible about what it accepts. Its semantics have to be interpreted as: "If you manage this namespace, or part of this namespace, please eager load what you got". In particular, if the receiver does not manage the namespace, it will simply do nothing, this is not an error condition.

Eager load exclusions, ignored files and directories, and shadowed files are not eager loaded.

Zeitwerk::Loader#eager_load_namespace is idempotent, but compatible with reloading. If you eager load a namespace and then reload, eager loading that namespace will load its (current) descendants again.

The method checks if a regular eager load was already executed, in which case it returns fast.

If root directories are assigned to custom namespaces, the method behaves as you'd expect, according to the namespacing relationship between the custom namespace and the argument.

Attempting to eager load a namespace without previously calling setup raises Zeitwerk::SetupRequired.

Eager load namespaces shared by several loaders

The method Zeitwerk::Loader.eager_load_namespace broadcasts eager_load_namespace to all loaders.

Zeitwerk::Loader.eager_load_namespace(MyFramework::Routes)

This may be handy, for example, if a framework supports plugins and a shared namespace needs to be eager loaded for the framework to function properly.

Please, note that loaders only eager load namespaces they manage, as documented above. Therefore, this method does not allow you to eager load namespaces not managed by Zeitwerk loaders.

This method does not require that all registered loaders have setup already invoked, since that is out of your control. If there's any in that state, it is simply skipped.

Global eager load

If you want to eager load yourself and all dependencies that use Zeitwerk, you can broadcast the eager_load call to all instances:

Zeitwerk::Loader.eager_load_all

This may be handy in top-level services, like web applications.

Note that thanks to idempotence Zeitwerk::Loader.eager_load_all won't eager load twice if any of the instances already eager loaded.

This method does not accept the force flag, since in general it wouldn't be a good idea to force eager loading in 3rd party code.

This method does not require that all registered loaders have setup already invoked, since that is out of your control. If there's any in that state, it is simply skipped.

Loading individual files

The method Zeitwerk::Loader#load_file loads an individual Ruby file:

loader.load_file("#{__dir__}/custom_web_app/routes.rb")

This is useful when the loader is not eager loading the entire project, but you still need an individual file to be loaded for things to function properly.

Both strings and Pathname objects are supported as arguments. The method raises Zeitwerk::Error if the argument is not a Ruby file, is ignored, is shadowed, or is not managed by the receiver.

Zeitwerk::Loader#load_file is idempotent, but compatible with reloading. If you load a file and then reload, a new call will load its (current) contents again.

If you want to eager load a directory, Zeitwerk::Loader#eager_load_dir is more efficient than invoking Zeitwerk::Loader#load_file on its files.

Reloading

Configuration and usage

Zeitwerk is able to reload code, but you need to enable this feature:

loader = Zeitwerk::Loader.new
loader.push_dir(...)
loader.enable_reloading # you need to opt-in before setup
loader.setup
...
loader.reload

There is no way to undo this, either you want to reload or you don't.

Enabling reloading after setup raises Zeitwerk::Error. Attempting to reload without having it enabled raises Zeitwerk::ReloadingDisabledError. Attempting to reload without previously calling setup raises Zeitwerk::SetupRequired.

Generally speaking, reloading is useful while developing running services like web applications. Gems that implement regular libraries, so to speak, or services running in testing or production environments, won't normally have a use case for reloading. If reloading is not enabled, Zeitwerk is able to use less memory.

Reloading removes the currently loaded classes and modules and resets the loader so that it will pick whatever is in the file system now.

It is important to highlight that this is an instance method. Don't worry about project dependencies managed by Zeitwerk, their loaders are independent.

Thread-safety

In order to reload safely, no other thread can be autoloading or reloading concurrently. Client code is responsible for this coordination.

For example, a web framework that serves each request in its own thread and has reloading enabled could create a read-write lock on boot like this:

require "concurrent/atomic/read_write_lock"

MyFramework::RELOAD_RW_LOCK = Concurrent::ReadWriteLock.new

You acquire the lock for reading for serving each individual request:

MyFramework::RELOAD_RW_LOCK.with_read_lock do
  serve(request)
end

Then, when a reload is triggered, just acquire the lock for writing in order to execute the method call safely:

MyFramework::RELOAD_RW_LOCK.with_write_lock do
  loader.reload
end

On reloading, client code has to update anything that would otherwise be storing a stale object. For example, if the routing layer of a web framework stores reloadable controller class objects or instances in internal structures, on reload it has to refresh them somehow, possibly reevaluating routes.

Inflection

Each individual loader needs an inflector to figure out which constant path would a given file or directory map to. Zeitwerk ships with two basic inflectors, and you can define your own.

Zeitwerk::Inflector

Each loader instantiated with Zeitwerk::Loader.new has an inflector of this type by default.

This is a very basic inflector that converts snake case to camel case:

user             -> User
users_controller -> UsersController
html_parser      -> HtmlParser

The camelize logic can be overridden easily for individual basenames:

loader.inflector.inflect(
  "html_parser"   => "HTMLParser",
  "mysql_adapter" => "MySQLAdapter"
)

The inflect method can be invoked several times if you prefer this other style:

loader.inflector.inflect "html_parser" => "HTMLParser"
loader.inflector.inflect "mysql_adapter" => "MySQLAdapter"

Overrides need to be configured before calling setup.

The inflectors of different loaders are independent of each other. There are no global inflection rules or global configuration that can affect this inflector. It is deterministic.

Zeitwerk::GemInflector

Each loader instantiated with Zeitwerk::Loader.for_gem has an inflector of this type by default.

This inflector is like the basic one, except it expects lib/my_gem/version.rb to define MyGem::VERSION.

The inflectors of different loaders are independent of each other. There are no global inflection rules or global configuration that can affect this inflector. It is deterministic.

Custom inflector

The inflectors that ship with Zeitwerk are deterministic and simple. But you can configure your own:

# frozen_string_literal: true

class MyInflector < Zeitwerk::Inflector
  def camelize(basename, abspath)
    if basename =~ /\Ahtml_(.*)/
      "HTML" + super($1, abspath)
    else
      super
    end
  end
end

The first argument, basename, is a string with the basename of the file or directory to be inflected. In the case of a file, without extension. In the case of a directory, without trailing slash. The inflector needs to return this basename inflected. Therefore, a simple constant name without colons.

The second argument, abspath, is a string with the absolute path to the file or directory in case you need it to decide how to inflect the basename. Paths to directories don't have trailing slashes.

Then, assign the inflector:

loader.inflector = MyInflector.new

This needs to be done before calling setup.

If a custom inflector definition in a gem takes too much space in the main file, you can extract it. For example, this is a simple pattern:

# lib/my_gem/inflector.rb
module MyGem
  class Inflector < Zeitwerk::GemInflector
    ...
  end
end

# lib/my_gem.rb
require "zeitwerk"
require_relative "my_gem/inflector"

loader = Zeitwerk::Loader.for_gem
loader.inflector = MyGem::Inflector.new(__FILE__)
loader.setup

module MyGem
  # ...
end

Since MyGem is referenced before the namespace is defined in the main file, it is important to use this style:

# Correct, effectively defines MyGem.
module MyGem
  class Inflector < Zeitwerk::GemInflector
    # ...
  end
end

instead of:

# Raises uninitialized constant MyGem (NameError).
class MyGem::Inflector < Zeitwerk::GemInflector
  # ...
end

Callbacks

The on_setup callback

The on_setup callback is fired on setup and on each reload:

loader.on_setup do
  # Ready to autoload here.
end

Multiple on_setup callbacks are supported, and they run in order of definition.

If setup was already executed, the callback is fired immediately.

The on_load callback

The usual place to run something when a file is loaded is the file itself. However, sometimes you'd like to be called, and this is possible with the on_load callback.

For example, let's imagine this class belongs to a Rails application:

class SomeApiClient
  class << self
    attr_accessor :endpoint
  end
end

With on_load, it is easy to schedule code at boot time that initializes endpoint according to the configuration:

# config/environments/development.rb
loader.on_load("SomeApiClient") do |klass, _abspath|
  klass.endpoint = "https://api.dev"
end

# config/environments/production.rb
loader.on_load("SomeApiClient") do |klass, _abspath|
  klass.endpoint = "https://api.prod"
end

Some uses cases:

  • Doing something with a reloadable class or module in a Rails application during initialization, in a way that plays well with reloading. As in the previous example.
  • Delaying the execution of the block until the class is loaded for performance.
  • Delaying the execution of the block until the class is loaded because it follows the adapter pattern and better not to load the class if the user does not need it.

on_load gets a target constant path as a string (e.g., "User", or "Service::NotificationsGateway"). When fired, its block receives the stored value, and the absolute path to the corresponding file or directory as a string. The callback is executed every time the target is loaded. That includes reloads.

Multiple callbacks on the same target are supported, and they run in order of definition.

The block is executed once the loader has loaded the target. In particular, if the target was already loaded when the callback is defined, the block won't run. But if you reload and load the target again, then it will. Normally, you'll want to define on_load callbacks before setup.

Defining a callback for a target not managed by the receiver is not an error, the block simply won't ever be executed.

It is also possible to be called when any constant managed by the loader is loaded:

loader.on_load do |cpath, value, abspath|
  # ...
end

The block gets the constant path as a string (e.g., "User", or "Foo::VERSION"), the value it stores (e.g., the class object stored in User, or "2.5.0"), and the absolute path to the corresponding file or directory as a string.

Multiple callbacks like these are supported, and they run in order of definition.

There are use cases for this last catch-all callback, but they are rare. If you just need to understand how things are being loaded for debugging purposes, please remember that Zeitwerk::Loader#log! logs plenty of information.

If both types of callbacks are defined, the specific ones run first.

Since on_load callbacks are executed right after files are loaded, even if the loading context seems to be far away, in practice the block is subject to circular dependencies. As a rule of thumb, as far as loading order and its interdependencies is concerned, you have to program as if the block was executed at the bottom of the file just loaded.

The on_unload callback

When reloading is enabled, you may occasionally need to execute something before a certain autoloaded class or module is unloaded. The on_unload callback allows you to do that.

For example, let's imagine that a Country class fetches a list of countries and caches them when it is loaded. You might want to clear that cache if unloaded:

loader.on_unload("Country") do |klass, _abspath|
  klass.clear_cache
end

on_unload gets a target constant path as a string (e.g., "User", or "Service::NotificationsGateway"). When fired, its block receives the stored value, and the absolute path to the corresponding file or directory as a string. The callback is executed every time the target is unloaded.

on_unload blocks are executed before the class is unloaded, but in the middle of unloading, which happens in an unspecified order. Therefore, that callback should not refer to any reloadable constant because there is no guarantee the constant works there. Those blocks should rely on objects only, as in the example above, or regular constants not managed by the loader. This remark is transitive, applies to any methods invoked within the block.

Multiple callbacks on the same target are supported, and they run in order of definition.

Defining a callback for a target not managed by the receiver is not an error, the block simply won't ever be executed.

It is also possible to be called when any constant managed by the loader is unloaded:

loader.on_unload do |cpath, value, abspath|
  # ...
end

The block gets the constant path as a string (e.g., "User", or "Foo::VERSION"), the value it stores (e.g., the class object stored in User, or "2.5.0"), and the absolute path to the corresponding file or directory as a string.

Multiple callbacks like these are supported, and they run in order of definition.

If both types of callbacks are defined, the specific ones run first.

Technical details

Zeitwerk uses the word "unload" to ease communication and for symmetry with on_load. However, in Ruby you cannot unload things for real. So, when does on_unload technically happen?

When unloading, Zeitwerk issues Module#remove_const calls. Classes and modules are no longer reachable through their constants, and on_unload callbacks are executed right before those calls.

Technically, though, the objects themselves are still alive, but if everything is used as expected and they are not stored in any non-reloadable place (don't do that), they are ready for garbage collection, which is when the real unloading happens.

Logging

Zeitwerk is silent by default, but you can ask loaders to trace their activity. Logging is meant just for troubleshooting, shouldn't normally be enabled.

The log! method is a quick shortcut to let the loader log to $stdout:

loader.log!

If you want more control, a logger can be configured as a callable

loader.logger = method(:puts)
loader.logger = ->(msg) { ... }

as well as anything that responds to debug:

loader.logger = Logger.new($stderr)
loader.logger = Rails.logger

In both cases, the corresponding methods are going to be passed exactly one argument with the message to be logged.

It is also possible to set a global default this way:

Zeitwerk::Loader.default_logger = method(:puts)

If there is a logger configured, you'll see traces when autoloads are set, files loaded, and modules autovivified. While reloading, removed autoloads and unloaded objects are also traced.

As a curiosity, if your project has namespaces you'll notice in the traces Zeitwerk sets autoloads for directories. That's a technique used to be able to descend into subdirectories on demand, avoiding that way unnecessary tree walks.

Loader tag

Loaders have a tag that is printed in traces in order to be able to distinguish them in globally logged activity:

Zeitwerk@9fa54b: autoload set for User, to be loaded from ...

By default, a random tag like the one above is assigned, but you can change it:

loader.tag = "grep_me"

The tag of a loader returned by for_gem is the basename of the root file without extension:

Zeitwerk@my_gem: constant MyGem::Foo loaded from ...

Ignoring parts of the project

Zeitwerk ignores automatically any file or directory whose name starts with a dot, and any files that do not have extension ".rb".

However, sometimes it might still be convenient to tell Zeitwerk to completely ignore some particular Ruby file or directory. That is possible with ignore, which accepts an arbitrary number of strings or Pathname objects, and also an array of them.

You can ignore file names, directory names, and glob patterns. Glob patterns are expanded when they are added and again on each reload.

There is an edge case related to nested root directories. Conceptually, root directories are independent source trees. If you ignore a parent of a nested root directory, the nested root directory is not affected. You need to ignore it explictly if you want it ignored too.

Let's see some use cases.

Use case: Files that do not follow the conventions

Let's suppose that your gem decorates something in Kernel:

# lib/my_gem/core_ext/kernel.rb

Kernel.module_eval do
  # ...
end

Kernel is already defined by Ruby so the module cannot be autoloaded. Also, that file does not define a constant path after the path name. Therefore, Zeitwerk should not process it at all.

The extension can still coexist with the rest of the project, you only need to tell Zeitwerk to ignore it:

kernel_ext = "#{__dir__}/my_gem/core_ext/kernel.rb"
loader.ignore(kernel_ext)
loader.setup

You can also ignore the whole directory:

core_ext = "#{__dir__}/my_gem/core_ext"
loader.ignore(core_ext)
loader.setup

Now, that file has to be loaded manually with require or require_relative:

require_relative "my_gem/core_ext/kernel"

and you can do that anytime, before configuring the loader, or after configuring the loader, does not matter.

Use case: The adapter pattern

Another use case for ignoring files is the adapter pattern.

Let's imagine your project talks to databases, supports several, and has adapters for each one of them. Those adapters may have top-level require calls that load their respective drivers:

# my_gem/db_adapters/postgresql.rb
require "pg"

but you don't want your users to install them all, only the one they are going to use.

On the other hand, if your code is eager loaded by you or a parent project (with Zeitwerk::Loader.eager_load_all), those require calls are going to be executed. Ignoring the adapters prevents that:

db_adapters = "#{__dir__}/my_gem/db_adapters"
loader.ignore(db_adapters)
loader.setup

The chosen adapter, then, has to be loaded by hand somehow:

require "my_gem/db_adapters/#{config[:db_adapter]}"

Note that since the directory is ignored, the required adapter can instantiate another loader to manage its subtree, if desired. Such loader would coexist with the main one just fine.

Use case: Test files mixed with implementation files

There are project layouts that put implementation files and test files together. To ignore the test files, you can use a glob pattern like this:

tests = "#{__dir__}/**/*_test.rb"
loader.ignore(tests)
loader.setup

Shadowed files

In Ruby, if you have several files called foo.rb in different directories of $LOAD_PATH and execute

require "foo"

the first one found gets loaded, and the rest are ignored.

Zeitwerk behaves in a similar way. If foo.rb is present in several root directories (at the same namespace level), the constant Foo is autoloaded from the first one, and the rest of the files are not evaluated. If logging is enabled, you'll see something like

file #{file} is ignored because #{previous_occurrence} has precedence

(This message is not public interface and may change, you cannot rely on that exact wording.)

Even if there's only one foo.rb, if the constant Foo is already defined when Zeitwerk finds foo.rb, then the file is ignored too. This could happen if Foo was defined by a dependency, for example. If logging is enabled, you'll see something like

file #{file} is ignored because #{constant_path} is already defined

(This message is not public interface and may change, you cannot rely on that exact wording.)

Shadowing only applies to Ruby files, namespace definition can be spread over multiple directories. And you can also reopen third-party namespaces if done orderly.

Edge cases

Explicit namespaces like Trip here:

# trip.rb
class Trip
  include Geolocation
end

# trip/geolocation.rb
module Trip::Geolocation
  ...
end

have to be defined with the class/module keywords, as in the example above.

For technical reasons, raw constant assignment is not supported:

# trip.rb
Trip = Class.new { ... }  # NOT SUPPORTED
Trip = Struct.new { ... } # NOT SUPPORTED

This only affects explicit namespaces, those idioms work well for any other ordinary class or module.

Beware of circular dependencies

In Ruby, you can't have certain top-level circular dependencies. Take for example:

# c.rb
class C < D
end

# d.rb
class D
  C
end

In order to define C, you need to load D. However, the body of D refers to C.

Circular dependencies like those do not work in plain Ruby, and therefore do not work in projects managed by Zeitwerk either.

Reopening third-party namespaces

Projects managed by Zeitwerk can work with namespaces defined by third-party libraries. However, they have to be loaded in memory before calling setup.

For example, let's imagine you're writing a gem that implements an adapter for Active Job that uses AwesomeQueue as backend. By convention, your gem has to define a class called ActiveJob::QueueAdapters::AwesomeQueue, and it has to do so in a file with a matching path:

# lib/active_job/queue_adapters/awesome_queue.rb
module ActiveJob
  module QueueAdapters
    class AwesomeQueue
      # ...
    end
  end
end

It is very important that your gem reopens the modules ActiveJob and ActiveJob::QueueAdapters instead of defining them. Because their proper definition lives in Active Job. Furthermore, if the project reloads, you do not want any of ActiveJob or ActiveJob::QueueAdapters to be reloaded.

Bottom line, Zeitwerk should not be managing those namespaces. Active Job owns them and defines them. Your gem needs to reopen them.

In order to do so, you need to make sure those modules are loaded before calling setup. For instance, in the entry file for the gem:

# Ensure these namespaces are reopened, not defined.
require "active_job"
require "active_job/queue_adapters"

require "zeitwerk"
# By passing the flag, we acknowledge the extra directory lib/active_job
# has to be managed by the loader and no warning has to be issued for it.
loader = Zeitwerk::Loader.for_gem(warn_on_extra_files: false)
loader.setup

With that, when Zeitwerk scans the file system and reaches the gem directories lib/active_job and lib/active_job/queue_adapters, it detects the corresponding modules already exist and therefore understands it does not have to manage them. The loader just descends into those directories. Eventually will reach lib/active_job/queue_adapters/awesome_queue.rb, and since ActiveJob::QueueAdapters::AwesomeQueue is unknown, Zeitwerk will manage it. Which is what happens regularly with the files in your gem. On reload, the namespaces are safe, won't be reloaded. The loader only reloads what it manages, which in this case is the adapter itself.

Introspection

The method Zeitwerk::Loader#dirs returns an array with the absolute paths of the root directories as strings:

loader = Zeitwerk::Loader.new
loader.push_dir(Pathname.new("/foo"))
loader.dirs # => ["/foo"]

This method accepts an optional namespaces keyword argument. If truthy, the method returns a hash table instead. Keys are the absolute paths of the root directories as strings. Values are their corresponding namespaces, class or module objects:

loader = Zeitwerk::Loader.new
loader.push_dir(Pathname.new("/foo"))
loader.push_dir(Pathname.new("/bar"), namespace: Bar)
loader.dirs(namespaces: true) # => { "/foo" => Object, "/bar" => Bar }

By default, ignored root directories are filtered out. If you want them included, please pass ignored: true.

These collections are read-only. Please add to them with Zeitwerk::Loader#push_dir.

Encodings

Zeitwerk supports projects whose files and file system are in UTF-8. The encoding of the file system can be checked this way:

% ruby -e "puts Encoding.find('filesystem')"
UTF-8

The test suite passes on Windows with codepage Windows-1252 if all the involved absolute paths are ASCII. Other supersets of ASCII may work too, but you have to try.

Rules of thumb

  1. Different loaders should manage different directory trees. It is an error condition to configure overlapping root directories in different loaders.

  2. Think the mere existence of a file is effectively like writing a require call for them, which is executed on demand (autoload) or upfront (eager load).

  3. In that line, if two loaders manage files that translate to the same constant in the same namespace, the first one wins, the rest are ignored. Similar to what happens with require and $LOAD_PATH, only the first occurrence matters.

  4. Projects that reopen a namespace defined by some dependency have to ensure said namespace is loaded before setup. That is, the project has to make sure it reopens, rather than define. This is often accomplished just loading the dependency.

  5. Objects stored in reloadable constants should not be cached in places that are not reloaded. For example, non-reloadable classes should not subclass a reloadable class, or mixin a reloadable module. Otherwise, after reloading, those classes or module objects would become stale. Referring to constants in dynamic places like method calls or lambdas is fine.

  6. In a given process, ideally, there should be at most one loader with reloading enabled. Technically, you can have more, but it may get tricky if one refers to constants managed by the other one. Do that only if you know what you are doing.

Debuggers

Zeitwerk works fine with debug.rb and Break.

Byebug is compatible except for an edge case explained in deivid-rodriguez/byebug#564.

Pronunciation

"Zeitwerk" is pronounced this way.

Supported Ruby versions

Zeitwerk works with CRuby 2.5 and above.

On TruffleRuby all is good except for thread-safety. Right now, in TruffleRuby Module#autoload does not block threads accessing a constant that is being autoloaded. CRuby prevents such access to avoid concurrent threads from seeing partial evaluations of the corresponding file. Zeitwerk inherits autoloading thread-safety from this property. This is not an issue if your project gets eager loaded, or if you lazy load in single-threaded environments. (See oracle/truffleruby#2431.)

JRuby 9.3.0.0 is almost there. As of this writing, the test suite of Zeitwerk passes on JRuby except for three tests. (See jruby/jruby#6781.)

Testing

In order to run the test suite of Zeitwerk, cd into the project root and execute

bin/test

To run one particular suite, pass its file name as an argument:

bin/test test/lib/zeitwerk/test_eager_load.rb

Furthermore, the project has a development dependency on minitest-focus. To run an individual test mark it with focus:

focus
test "capitalizes the first letter" do
  assert_equal "User", camelize("user")
end

and run bin/test.

Motivation

Kernel#require is brittle

Since require has global side-effects, and there is no static way to verify that you have issued the require calls for code that your file depends on, in practice it is very easy to forget some. That introduces bugs that depend on the load order.

Also, if the project has namespaces, setting things up and getting client code to load things in a consistent way needs discipline. For example, require "foo/bar" may define Foo, instead of reopen it. That may be a broken window, giving place to superclass mismatches or partially-defined namespaces.

With Zeitwerk, you just name things following conventions and done. Things are available everywhere, and descend is always orderly. Without effort and without broken windows.

Rails autoloading was brittle

Autoloading in Rails was based on const_missing up to Rails 5. That callback lacks fundamental information like the nesting or the resolution algorithm being used. Because of that, Rails autoloading was not able to match Ruby's semantics, and that introduced a series of issues. Zeitwerk is based on a different technique and fixed Rails autoloading starting with Rails 6.

Awards

Zeitwerk has been awarded an "Outstanding Performance Award" Fukuoka Ruby Award 2022.

Thanks

I'd like to thank @matthewd for the discussions we've had about this topic in the past years, I learned a couple of tricks used in Zeitwerk from him.

Also, would like to thank @Shopify, @rafaelfranca, and @dylanahsmith, for sharing this PoC. The technique Zeitwerk uses to support explicit namespaces was copied from that project.

Jean Boussier (@casperisfine, @byroot) deserves special mention. Jean migrated autoloading in Shopify when Zeitwerk integration in Rails was yet unreleased. His work and positive attitude have been outstanding, and thanks to his feedback the interface and performance of Zeitwerk are way, way better. Kudos man ❤️.

Finally, many thanks to @schurig for recording an audio file with the pronunciation of "Zeitwerk" in perfect German. 💯

License

Released under the MIT License, Copyright (c) 2019–ω Xavier Noria.

About

Efficient and thread-safe code loader for Ruby

Resources

Readme

License

MIT license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

53 tags

Packages

No packages published

Languages

  • Ruby 100.0%