jafrog's dev blog

Update: As @yves_senn and @fxn pointed out there’s a Rails Guide now on the subject of autloading and constants reloading. Kudos to @fxn!

It all started with one innocent refactoring. We had two classes defined in the same file. Say Foo and Foo::Bar. Naturally, as Foo lives in app/models/foo.rb the place for Foo::Bar is to be app/models/foo/bar.rb. As app/models/ is listed among autoload_paths I thought nothing could go wrong. As soon as Bar is referenced from inside the Foo its path will be resolved and it would be loaded from app/models/foo/bar.rb.

The change was tested and successfully deployed to production. The hit came from an unexpected side - Rake task using this code failed with Uninitialized constant Bar.

Rails eager loading

The first question was “Why does it work in the UI but fails in the Rake task?” After some investigation I found out two facts:

• Rails redefines const_missing (more on that later);
• Rails const_missing can’t locate its own ass let alone my poor Foo::Bar class.

Given that, the fact that the app still worked when accessed from the web interface could mean only one thing - resolving Foo::Bar never had to hit const_missing in the first place. All thanks to eager loading.

Rails’s eager loading feature allows files (and constants defined in them) to be loaded on startup instead of loading “on demand” when a constant is first referenced. This means that when in production environment Foo::Bar constant is first encountered it’s already loaded and doesn’t need to be found with const_missing.

But if production Rake tasks run in the same production environment where eager loading is enabled then why it doesn’t work for them? For Rails 3.2 the reason could be found in finisher.rb file:

initializer :eager_load! do
  if config.cache_classes && !(defined?($rails_rake_task) && $rails_rake_task)
    ActiveSupport.run_load_hooks(:before_eager_load, self)
    eager_load!
  end
end

Note the condition. Checking for config.cache_classes is fair enough - it probably doesn’t make much sense to load everything on startup if you’re going to reload it on every change. But it also checks for the global variable $rails_rake_task which if defined indicates that the code is run inside the Rake task!

Interestingly there are no such conditions in Rails 4.2. Instead you explicitly say whether or not to eager load classes with config.eager_load option:

initializer :eager_load! do
  if config.eager_load
    ActiveSupport.run_load_hooks(:before_eager_load, self)
    config.eager_load_namespaces.each(&:eager_load!)
  end
end

The solution I came up with was to require bar.rb in foo.rb to make sure Foo::Bar is already loaded when it’s needed. I tested it in the development environment where eager loading was disabled and it seemed to work. Except it didn’t.

As soon as I changed anything at all in the app and reloaded the page I got same old Undefined constant Bar exception. After thinking about it for a while I realized that it happened because of the combination of two factors:

• When Rails reloads classes it removes all constants first;
• require in Ruby requires files only once. If the file was already loaded it will not be loaded again.

So when I changed code in the app Rails removed all constants, but require on top of the foo.rb didn’t do anything because files had already been loaded. One way to solve this is to use load in place of require which will load a file every time the code is run. But I decided to get to the core of the const_missing issue instead. Why can’t it find Bar constant when it’s referenced from inside the scope it was defined in?

Rails const_missing

First, read this brilliant piece on Ruby and Rails constant lookup and autoloading. It explains everything you should know about Rails autoload process.

Now let’s dig into const_missing as it’s defined in ActiveSupport::Dependencies module:

def const_missing(const_name, nesting = nil)
  klass_name = name.presence || "Object"

  unless nesting
    # We'll assume that the nesting of Foo::Bar is ["Foo::Bar", "Foo"]
    # even though it might not be, such as in the case of
    # class Foo::Bar; Baz; end
    nesting = []
    klass_name.to_s.scan(/::|$/) { nesting.unshift $` }
  end

  # If there are multiple levels of nesting to search under, the top
  # level is the one we want to report as the lookup fail.
  error = nil

  nesting.each do |namespace|
    begin
      return Dependencies.load_missing_constant Inflector.constantize(namespace), const_name
    rescue NoMethodError then raise
    rescue NameError => e
      error ||= e
    end
  end

  # Raise the first error for this set. If this const_missing came from an
  # earlier const_missing, this will result in the real error bubbling
  # all the way up
  raise error
end

The method implementation is quite a bit different in Rails 4, but the process is essentially the same:

1. Get parent scope name.
2. Get nesting from the nesting parameter. I’m not sure how exactly this parameter is passed, if ever. So it’s safe to assume that nesting is nil. Note that nesting is also a special method in pry which makes it a little tricky to inspect.
3. If nesting is nil try to derive nesting from the parent scope name - klass_name.
4. Construct constant “full name” from the const_name and nesting.
5. Try to find missing constant based on the constructed name.

Now here’s the rough example of how the troublesome code looks:

# app/models/foo.rb

module Foo
  class << self
    def foo
        Bar.new.bar
      end
    end
  end
end

# app/models/foo/bar.rb

module Foo
  class Bar
    def bar
      "bar"
    end
  end
end

What puzzled me is that Bar is referenced from inside the scope where it’s defined, why const_missing can’t detect that? And then it hit me. Look closer at const_missing first line: klass_name = name.presence || "Object". What is name?

Turns out name is a built in Ruby method that

1
Returns the name of the module... Returns nil for anonymous modules.

And if we inspect self at the point where name is called it will show us #<Class:Foo::Bar> which is a class object because Baz is referenced from inside class << self scope! And as class object is an anonymous module its name is nil and const_missing is trying to find top-level Baz constant instead of Foo::Bar. If instead of class << self the method was defined as self.baz the problem wouldn’t occur.

Conclusion

I don’t know if there’s a catchall solution to this kind of problems. Explicitly loading all dependencies seems to be to tedious for large Rails projects. Referencing constants by the full name all the time breaks encapsulation. IMO The best way to avoid constant lookup issues in all your environments is to

• understand how constant lookup works and what code at each point of your program;
• keep an eye on scopes where constant is referenced.

I hope this will save you some hair I had to pull off while trying to debug this issue.

TL;DR: Look out for classes referenced from inside class << self.