New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Revert adding loaded specs to Gem::Specification.stubs
and Gem::Specification.stubs_for
#3455
Conversation
9f1db4b
to
dba12a3
Compare
6f04477
to
5b90223
Compare
…ecification.stubs_for` The rationale is that: * The change has caused realworld issues. See for example ruby/did_you_mean#117 and specifically [this comment](ruby/did_you_mean#117 (comment)) for a great explanation of the issue it caused for `did_you_mean`. * The change also causes problems for our development workflows. For example, because of it, our `bundler` specs cannot currently be run with `bin/rake` and we have to use `bin/rspec` or `bin/parallel_spec` directly. The explanation for this is: - Our specs install test dependencies to `tmp` before running specs. - `rake` is one of these test dependencies. - Before installing each test dependency, we check whether it has matching installed specs: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/bundler/spec/support/rubygems_ext.rb#L109-L114. - Normally, if `rake` has not yet been installed to `tmp`, this check fails and `rake` is installed, but since the loaded specs are now added to `Gem::Specification.stubs` and `rake`'s specification _is_ loaded because we're running through `bin/rake`, the check incorrectly assumes that `rake` is already installed to `tmp` and skips installation. - At a later point the specs check whether `rake` is actually installed and fail if it's not: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/bundler/spec/support/builders.rb#L372-L383 Essentially, both of the issues are the same. If at runtime we change the location of gems, we'll _want_ to not consider loaded specifications when dealing with the new gem location, because the loaded specifications have not been loaded from there. Loaded specifications is something different from installed stub specifications and those should not be mixed. The PR still seemed to have fixed an issue, so I did my archaeology job and investigated the original issue to double check if reverting is ok. The logs for the original error can be found here: https://ci.appveyor.com/project/rubygems/rubygems/build/1172/job/ogubyucpljcv22ux. So I installed ruby 2.4.4, checked out the commit reference before the offending PR, and the exact error reproduced. 🎉 ``` $ rake test /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:231:in `search_for': Unable to resolve dependency: user requested 'bundler (= 1.16.2)' (Gem::UnsatisfiableDependencyError) from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:283:in `block in sort_dependencies' from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:277:in `each' from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:277:in `sort_by' from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:277:in `with_index' from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:277:in `sort_dependencies' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/delegates/specification_provider.rb:52:in `block in sort_dependencies' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/delegates/specification_provider.rb:69:in `with_no_such_dependency_error_handling' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/delegates/specification_provider.rb:51:in `sort_dependencies' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/resolution.rb:165:in `initial_state' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/resolution.rb:106:in `start_resolution' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/resolution.rb:64:in `resolve' from /home/deivid/Code/rubygems/lib/rubygems/resolver/molinillo/lib/molinillo/resolver.rb:42:in `resolve' from /home/deivid/Code/rubygems/lib/rubygems/resolver.rb:188:in `resolve' from /home/deivid/Code/rubygems/lib/rubygems/request_set.rb:396:in `resolve' from /home/deivid/Code/rubygems/lib/rubygems/request_set.rb:408:in `resolve_current' from /home/deivid/Code/rubygems/lib/rubygems.rb:243:in `finish_resolve' from /home/deivid/Code/rubygems/lib/rubygems/rdoc.rb:13:in `<top (required)>' from /home/deivid/Code/rubygems/lib/rubygems/core_ext/kernel_require.rb:54:in `require' from /home/deivid/Code/rubygems/lib/rubygems/core_ext/kernel_require.rb:54:in `require' from /home/deivid/Code/rubygems/lib/rubygems/test_case.rb:1563:in `<top (required)>' from /home/deivid/Code/rubygems/test/rubygems/test_bundled_ca.rb:2:in `require' from /home/deivid/Code/rubygems/test/rubygems/test_bundled_ca.rb:2:in `<top (required)>' from /home/deivid/.rbenv/versions/2.4.4/lib/ruby/gems/2.4.0/gems/rake-12.0.0/lib/rake/rake_test_loader.rb:15:in `require' from /home/deivid/.rbenv/versions/2.4.4/lib/ruby/gems/2.4.0/gems/rake-12.0.0/lib/rake/rake_test_loader.rb:15:in `block in <main>' from /home/deivid/.rbenv/versions/2.4.4/lib/ruby/gems/2.4.0/gems/rake-12.0.0/lib/rake/rake_test_loader.rb:4:in `select' from /home/deivid/.rbenv/versions/2.4.4/lib/ruby/gems/2.4.0/gems/rake-12.0.0/lib/rake/rake_test_loader.rb:4:in `<main>' rake aborted! Command failed with status (1) Tasks: TOP => test ``` Now the explanation of the error: * Rubygems base `TestCase` class requires `bundler` because some tests use `bundler`: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/lib/rubygems/test_case.rb#L26 * That `require` (our custom rubygems require) would activate the default bundler spec (1.16.1 for ruby 2.4.4) but then overwrite it with a 1.16.2 version (the locally provided bundler those days) due to [this old hack](https://github.com/rubygems/bundler/blob/9f7bf0ac3ab8d995e3a274cec3c292a5203f4534/lib/bundler/version.rb#L7-L23). * Rubygems base `TestCase` class requires `rubygems/rdoc`: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/lib/rubygems/test_case.rb#L1536 * And that file ends up calling `Gem.finish_resolve`: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/lib/rubygems/rdoc.rb#L13 * `Gem.finish_resolve` adds the currently loaded specs to the resolution: https://github.com/rubygems/rubygems/blob/2bbcdcde08b90d4ef03da8fb1f7a8a3313e13bb8/lib/rubygems.rb#L235 * That means it would try to resolve bundler 1.16.2, but no specification for that version was installed since the default was 1.16.1. That explains why upgrading to rubygems 2.7.7 fixed the issue, since it provided bundler 1.16.2 by default so there was not bundler version discrepancy. After understanding the error, I conclude that: * Only this part of the original patch was actually needed to resolve the error, not any of the changes in `Gem::Specification.stubs` and `Gem::Specification.stubs_for`: ```diff diff --git a/lib/rubygems/test_case.rb b/lib/rubygems/test_case.rb index f1cd3d274c..92c848e870 100644 --- a/lib/rubygems/test_case.rb +++ b/lib/rubygems/test_case.rb @@ -13,6 +13,15 @@ else require 'rubygems' end +# If bundler gemspec exists, add to stubs +bundler_gemspec = File.expand_path("../../../bundler/bundler.gemspec", __FILE__) +if File.exist?(bundler_gemspec) + Gem::Specification.dirs.unshift File.dirname(bundler_gemspec) + Gem::Specification.class_variable_set :@@stubs, nil + Gem::Specification.stubs + Gem::Specification.dirs.shift +end + begin gem 'minitest' rescue Gem::LoadError ``` So, I propose to revert adding loaded specification to `Gem::Specification.stubs` and `Gem::Specification.stubs_for` because I think it's safe, it fixes the issues caused by their addition, and it simplifies `Gem::Specification` code, which is already complicated enough.
5b90223
to
5269cd6
Compare
I ended up keeping the hack to add the local bundler gemspec stub to the list of loaded stubs at the beginning of the base test case file. It's a big hack, but it makes sense to me and I think it might fix some issues in cases where you have the exact same version as the one being tested installed globally. Plus, the goal of this PR was removing loaded specs from |
Description:
In this PR I'm proposing to partially revert #2309. In that PR,
Gem::Specification.stubs
andGem::Specification.stubs_for
were changed to also include loaded specifications.Rationale
The change has caused realworld issues. See for example No DidYouMean constants in Travis CI ruby/did_you_mean#117 and specifically this comment for a great explanation of the issue it caused for
did_you_mean
.The change also causes problems for our development workflows. For example, because of it, our
bundler
specs cannot currently be run withbin/rake
and we have to usebin/rspec
orbin/parallel_spec
directly. The explanation for this is:tmp
before running specs.rake
is one of these test dependencies.rubygems/bundler/spec/support/rubygems_ext.rb
Lines 109 to 114 in 2bbcdcd
rake
has not yet been installed totmp
, this check fails andrake
is installed, but since the loaded specs are now added toGem::Specification.stubs
andrake
's specification is loaded because we're running throughbin/rake
, the check incorrectly assumes thatrake
is already installed totmp
and skips installation.rake
is actually installed and fail if it's not:rubygems/bundler/spec/support/builders.rb
Lines 372 to 383 in 2bbcdcd
Essentially, both of the issues are the same. If at runtime we change the location of gems, we'll want to not consider loaded specifications when dealing with the new gem location, because the loaded specifications have not been loaded from there. Loaded specifications is something different from installed stub specifications and those should not be mixed.
Investigation of the original issue
The PR still seemed to have fixed an issue, so I did my archaeology job and investigated the original issue to double check if reverting is ok. The logs for the original error can be found here: https://ci.appveyor.com/project/rubygems/rubygems/build/1172/job/ogubyucpljcv22ux.
So I installed ruby 2.4.4, checked out the commit reference before the offending PR, and the exact error reproduced. 🎉
Now the explanation of the error:
Rubygems base
TestCase
class requiresbundler
because some tests usebundler
:rubygems/lib/rubygems/test_case.rb
Line 26 in 2bbcdcd
That
require
(our custom rubygems require) would activate the default bundler spec (1.16.1 for ruby 2.4.4) but then overwrite it with a 1.16.2 version (the locally provided bundler those days) due to this old hack.Rubygems base
TestCase
class requiresrubygems/rdoc
:rubygems/lib/rubygems/test_case.rb
Line 1536 in 2bbcdcd
And that file ends up calling
Gem.finish_resolve
:rubygems/lib/rubygems/rdoc.rb
Line 13 in 2bbcdcd
Gem.finish_resolve
adds the currently loaded specs to the resolution:rubygems/lib/rubygems.rb
Line 235 in 2bbcdcd
That means it would try to resolve bundler 1.16.2, but no specification for that version was installed since the default was 1.16.1. That explains why upgrading to rubygems 2.7.7 fixed the issue, since it provided bundler 1.16.2 by default so there was not bundler version discrepancy.
Conclusions
After understanding the error, I conclude that:
Gem::Specification.stubs
andGem::Specification.stubs_for
So, I propose to revert adding loaded specification to
Gem::Specification.stubs
andGem::Specification.stubs_for
because I think it's safe, it fixes the issues caused by their addition, and it simplifiesGem::Specification
code, which is already complicated enough.Tasks:
I will abide by the code of conduct.