-
Directories are processed in lexicographic order.
Different file systems may list directories in different order, and with this change we ensure that client code eager loads consistently across platforms, for example.
-
Before this release, subdirectories of root directories always represented namespaces (unless ignored or collapsed). From now on, to be considered namespaces they also have to contain at least one non-ignored Ruby file with extension
.rb
, directly or recursively.If you know beforehand a certain directory or directory pattern does not represent a namespace, it is intentional and more efficient to tell Zeitwerk to ignore it.
However, if you don't do so and have a directory
tasks
that only contains Rake files, arguably that directory is not meant to represent a Ruby module. Before, Zeitwerk would define a top-levelTasks
module after it; now, it does not.This feature is also handy for projects that have directories with auxiliary resources mixed in the project tree in a way that is too dynamic for an ignore pattern to be practical. See #216.
In the unlikely case that an existing project has an empty directory for the sole purpose of defining a totally empty module (no code, and no nested classes or modules), such module has now to be defined in a file.
Directories are scanned again on reloads.
-
On setup, loaders created with
Zeitwerk::Loader.for_gem
issue warnings iflib
has extra, non-ignored Ruby files or directories.This is motivated by existing gems with directories under
lib
that are not meant to define Ruby modules, like directories for Rails generators, for instance.This warning can be silenced in the unlikely case that the extra stuff is actually autoloadable and has to be managed by Zeitwerk.
Please, check the documentation for further details.
This method returns an instance of a private subclass of
Zeitwerk::Loader
now, but you cannot rely on the type, just on the interface.
- If a file did not define the expected constant, there was a reload, and there were
on_unload
callbacks, Zeitwerk still tried to access the constant during reload, which raised. This has been corrected.
- The change introduced in 2.5.2 implied a performance regression that was particularly dramatic in Ruby 3.1. We'll address #198 in a different way.
- When
Module#autoload
triggers the autovivification of an implicit namespace,$LOADED_FEATURES
now gets the correspoding directory pushed. This is just a tweak to Zeitwerk'sKernel#require
decoration. That way it acts more like the original, and cooperates better with other potentialKernel#require
wrappers, like Bootsnap's.
- Restores support for namespaces that are not hashable. For example namespaces that override the
hash
method with a different arity as shown in #188.
-
Requires Ruby 2.5.
-
Deletes the long time deprecated preload API. Instead of:
loader.preload("app/models/user.rb")
just reference the constant on setup:
loader.on_setup { User }
If you want to eager load a namespace, use the constants API:
loader.on_setup do Admin.constants(false).each { |cname| Admin.const_get(cname) } end
-
Fixes a bug in which a certain valid combination of overlapping trees managed by different loaders and ignored directories was mistakenly reported as having conflicting directories.
-
Detects external namespaces defined with
Module#autoload
. If your project reopens a 3rd party namespace, Zeitwerk already detected it and did not consider the namespace to be managed by the loader (automatically descends, ignored for reloads, etc.). However, the loader did not do that if the namespace had only an autoload in the 3rd party code yet to be executed. Now it does.
-
Implements
Zeitwerk::Loader#on_setup
, which allows you to configure blocks of code to be executed on setup and on each reload. When the callback is fired, the loader is ready, you can refer to project constants in the block.See the documentation for further details.
-
There is a new catch-all
Zeitwerk::Loader#on_load
that takes no argument and is triggered for all loaded objects:loader.on_load do |cpath, value, abspath| # ... end
Please, remember that if you want to trace the activity of a loader,
Zeitwerk::Loader#log!
logs plenty of information.See the documentation for further details.
-
The block of the existing
Zeitwerk::Loader#on_load
receives also the value stored in the constant, and the absolute path to its corresponding file or directory:loader.on_load("Service::NotificationsGateway") do |klass, abspath| # ... end
Remember that blocks can be defined to take less arguments than passed. So this change is backwards compatible. If you had
loader.on_load("Service::NotificationsGateway") do Service::NotificationsGateway.endpoint = ... end
That works.
-
Implements
Zeitwerk::Loader#on_unload
, which allows you to configure blocks of code to be executed before a certain class or module gets unloaded:loader.on_unload("Country") do |klass, _abspath| klass.clear_cache end
These callbacks are invoked during unloading, which happens in an unspecified order. Therefore, they should not refer to reloadable constants.
You can also be called for all unloaded objects:
loader.on_unload do |cpath, value, abspath| # ... end
Please, remember that if you want to trace the activity of a loader,
Zeitwerk::Loader#log!
logs plenty of information.See the documentation for further details.
-
Performance improvements.
-
Documentation improvements.
-
The method
Zeitwerk::Loader#eager_load
accepts aforce
flag:loader.eager_load(force: true)
If passed, eager load exclusions configured with
do_not_eager_load
are not honoured (but ignored files and directories are).This may be handy for test suites that eager load in order to ensure all files define the expected constant.
-
Eliminates internal use of
File.realpath
. One visible consequence is that in logs root dirs are shown as configured if they contain symlinks. -
When an autoloaded file does not define the expected constant, Ruby clears state differently starting with Ruby 3.1. Unloading has been revised to be compatible with both behaviours.
-
Logging prints a few new traces.
-
Implements
Zeitwerk::Loader#on_load
, which allows you to configure blocks of code to be executed after a certain class or module have been loaded:# config/environments/development.rb loader.on_load("SomeApiClient") do SomeApiClient.endpoint = "https://api.dev" # config/environments/production.rb loader.on_load("SomeApiClient") do SomeApiClient.endpoint = "https://api.prod" end
See the documentation for further details.
- Use
__send__
instead ofsend
internally.
-
Zeitwerk::Loader#push_dir
supports an optionalnamespace
keyword argument. Pass a class or module object if you want the given root directory to be associated with it instead ofObject
. Said class or module object cannot be reloadable. -
The default inflector is even more performant.
-
Saves some unnecessary allocations made internally by MRI. See #125, by @casperisfine.
-
Documentation improvements.
-
Internal code base maintenance.
-
Adds support for collapsing directories.
For example, if
booking/actions/create.rb
is meant to defineBooking::Create
because the subdirectoryactions
is there only for organizational purposes, you can tell Zeitwerk withcollapse
:loader.collapse("booking/actions")
The method also accepts glob patterns to support standardized project structures:
loader.collapse("*/actions")
Please check the documentation for more details.
-
Eager loading is idempotent, but now you can eager load again after reloading.
Zeitwerk::NameError#name
has the name of the missing constant now.
-
Zeitwerk raised
NameError
when a managed file did not define its expected constant. Now, it raisesZeitwerk::NameError
instead, so it is possible for client code to distinguish that mismatch from a regularNameError
.Regarding backwards compatibility,
Zeitwerk::NameError
is a subclass ofNameError
.
-
The default inflectors have API to override how to camelize selected basenames:
loader.inflector.inflect "mysql_adapter" => "MySQLAdapter"
This addresses a common pattern, which is to use the basic inflectors with a few straightforward exceptions typically configured in a hash table or
case
expression. You no longer have to define a custom inflector if that is all you need. -
Documentation improvements.
- Raises
Zeitwerk::NameError
with a better error message when a managed file or directory has a name that yields an invalid constant name when inflected.Zeitwerk::NameError
is a subclass ofNameError
.
-
Preloading is soft-deprecated. The use case it was thought for is no longer. Please, if you have a legit use case for it, drop me a line.
-
Root directory conflict detection among loaders takes ignored directories into account.
-
Supports classes and modules with overridden
name
methods. -
Documentation improvements.
- Fixes eager loading nested root directories. The new approach in 2.1.7 introduced a regression.
-
Prevent the inflector from deleting parts un multiword constants whose capitalization is the same. For example,
point_2d
should be inflected asPoint2d
, rather thanPoint
. While the inflector is frozen, this seems to be just wrong, and the refinement should be backwards compatible, since those constants were not usable. -
Make eager loading consistent with auto loading with regard to detecting namespaces that do not define the matching constant.
-
Documentation improvements.
-
Fixed: If an eager load exclusion contained an autoload for a namespace also present in other branches that had to be eager loaded, they could be skipped.
-
loader.log!
is a convenient shortcut to get traces to$stdout
. -
Allocates less strings.
-
Failed autoloads raise
NameError
as always, but with a more user-friendly message instead of the original generic one from Ruby. -
Eager loading uses
const_get
now rather thanrequire
. A file that does not define the expected constant could be eager loaded, but not autoloaded, which would be inconsistent. Thanks to @casperisfine for reporting this one and help testing the alternative.
-
Supports deletion of root directories in disk after they've been configured.
push_dir
requires root directories to exist to prevent misconfigurations, but after that Zeitwerk no longer assumes they exist. This might be convenient if you removed one in a web application while a server was running.
- Documentation improvements.
- Internal work.
- Calling
reload
with reloading disabled raisesZeitwerk::ReloadingDisabledError
.
- Internal performance work.
loaded_cpaths
is gone, you can ask if a constant path is going to be unloaded instead withloader.to_unload?(cpath)
. Thanks to this refinement, Zeitwerk is able to consume even less memory. (Change included in a minor upgrade because the introspection API is not documented, and it still isn't, needs some time to settle down).
- Reloading is disabled by default. In order to be able to reload you need to opt-in by calling
loader.enable_reloading
before setup. The motivation for this breaking change is twofold. On one hand, this is a design decision at the interface/usage level that reflects that the majority of use cases for Zeitwerk do not need reloading. On the other hand, if reloading is not enabled, Zeitwerk is able to use less memory. Notably, this is more optimal for large web applications in production.
- Faster reload. If you're using
bootsnap
, requires at least version 1.4.2.
- Includes an optimization.
- Fixes concurrent autovivifications.
-
Trace point optimization for singleton classes by @casperisfine. See the use case, explanation, and patch in #24.
-
Zeitwerk::Loader#do_not_eager_load
provides a way to have autoloadable files and directories that should be skipped when eager loading.
- Files shadowed by previous occurrences defining the same constant path were being correctly skipped when autoloading, but not when eager loading. This has been fixed. This mimicks what happens when there are two files in
$LOAD_PATH
with the same relative name, only the first one is loaded byrequire
.
- Bug fix by @casperisfine: If the superclass or one of the ancestors of an explicit namespace
N
has an autoload set for constantC
, andn/c.rb
exists, the autoload forN::C
proper could be missed.
-
Improved documentation.
-
Zeitwerk creates at most one trace point per process, instead of one per loader. This is more performant when there are multiple gems managed by Zeitwerk.
- After module vivification, the tracer could trigger one unnecessary autoload walk.
- In addition to callables, loggers can now also be any object that responds to
debug
, which accepts one string argument.
- Use
pretty_print
in the exception message for conflicting directories.
- Two different loaders cannot be managing the same files. Now,
Zeitwerk::Loader#push_dir
raisesZeitwerk::ConflictingDirectory
if it detects a conflict.
- New class attribute
Zeitwerk::Loader.default_logger
, inherited by newly instantiated loaders. Default isnil
. - Traces include the loader tag in the prefix to easily distinguish them.
- Loaders now have a tag.
- Documentation improvements.
- Documentation improvements.
Zeitwerk::Loader#ignore
accepts glob patterns.- New read-only introspection method
Zeitwerk::Loader.all_dirs
. - New read-only introspection method
Zeitwerk::Loader#dirs
. - New introspection predicate
Zeitwerk::Loader#loaded?(cpath)
.
do_not_eager_load
has been removed, please useignore
to opt-out.- Documentation improvements.
- Pronunciation section in the README, linking to sample audio file.
- All logged messages have a "Zeitwerk:" prefix for easy grepping.
- On reload, the logger also traces constants and autoloads removed.
- Initial beta release.