Bug #18813
Updated by fxn (Xavier Noria) almost 2 years ago
## Introduction Let's consider ```ruby module M autoload :X, 'x' end ``` The constants API does not distinguish existing constants from potential constants: ```ruby M.constants(false) # => [:X] M.const_defined?(:X, false) # => true ``` ## Expectations For a `false` inherited flag, the documentation of `Module#constants` says > Returns an array of the names of the constants accessible in mod. As a Ruby programmer, that is telling me `X` belongs to `M`, and `M::X` is a valid reference. Therefore, for coherence, if loading `x.rb` does not define `M::X`, I'd expect that to be a programming error. I'd expect `Module#autoload` to raise `NameError` with an ad-hoc error message in the line of "file '/full/path/to/x.rb' failed to define the constant M::X`. This does not happen today, `Module#autoload` is a simple trigger that loads a file and execution resumes with whatever side-effects that happened to have. Similar to `Module#const_missing`. However, to me, `Module#autoload` is different from `Module#const_missing` in a fundamental way. If autoloads were not present in the constants API, both could be equal, but they are present. **For consistency with the constants API**, I believe there has to be a strict expectation. If the autoload does not define `M::X`, the programmer did a mistake and an exception should say so. Zeitwerk [does this](https://github.com/fxn/zeitwerk/blob/5a700a322495398730ed74c6bec8180a75ba404f/lib/zeitwerk/loader/callbacks.rb#L27) by hand nowadays. There is a patch implementing this in https://github.com/ruby/ruby/pull/5949. ## Backwards Compatibility Backwards compatibility has to be considered, because some people do, for example: ```ruby module MyGem autoload :OpenSSL, 'openssl' end ``` While that is technically allowed, in my opinion that idiom is an unnecessary abuse of autoloading. The autoload is not even scoped to your gem, because ```ruby class module MyGem::MyClass OpenSSL end ``` would not work. That autoload should be in the top-level, where `OpenSSL` is going to be defined (assuming the top-level nesting is empty, you know). My hunch is that such autoloads may technically exist, but this is a very edge feature and very few people know about it. And those that know, may still define autoloads in their natural place. I'd be surprised if it is widely used. Anyway, in case you'd like this proposal, whether this change deserves a deprecation cycle is totally your call.