Bug #18813: Let Module#autoload be strict about the autoloaded constant

• Status: Closed
• Assignee: -
• Target version: -
• Backport: 2.7: UNKNOWN, 3.0: UNKNOWN, 3.1: UNKNOWN

[fxn (Xavier Noria)]

Introduction

Let's consider

module M
  autoload :X, 'x'
end

The constants API does not distinguish existing constants from potential constants:

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 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:

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

class 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.

[byroot (Jean Boussier)]

Thanks for opening the ticket @fxn (Xavier Noria).

I'm in favor of this change, however I believe that a deprecation cycle is preferable.

This behavior is a bit of a cruft, but at the same time it's not really breaking much things, and it's not in the way of an optimization or anything like that.

So a warning seems enough to ensure this mis-use will be eliminated over time, a hard break doesn't had value in my opinion.

[fxn (Xavier Noria)]

• Description updated (diff)

[ioquatix (Samuel Williams)]

I support raising an error when this fails to load the right constant. I think we can add a warning in 3.2 and make it an error in 3.3.

[byroot (Jean Boussier)]

• Tracker changed from Misc to Bug
• Backport set to 2.7: UNKNOWN, 3.0: UNKNOWN, 3.1: UNKNOWN

Changing this from Misc to bug, and I'll add it to the upcoming dev-meeting.

[matz (Yukihiro Matsumoto)]

I am against the behavior change, because of the compatibility concern. It's OK for me to give warning instead of raising exceptions.

Matz.

[byroot (Jean Boussier)]

• Status changed from Open to Closed

Applied in changeset git|eca31d24d606a73def3674938112dc3c5b79c445.

---

[Bug #18813] Warn when autoload has to lookup in parent namespace

This is a verbose mode only warning.

[fxn (Xavier Noria)]

Hi @matz (Yukihiro Matsumoto)! I'd like to work on some documentation. Would it be OK for you to say something along the line that declaring autoloads for constants not defined in the receiver is discouraged?