bundle install get stuck for a specific gem under setup-ruby action + windows + ruby 2.6

[ntkme]

Although I have "a specific gem" in the title, but it does not appear to be the root cause of this problem. So, please bear with me for the details.

Here is a repository for reproduction: https://github.com/ntkme/debug-setup-ruby/blob/main/.github/workflows/build.yml

This is a matrix build:

     matrix:
        os: [macos-latest, ubuntu-latest, windows-latest]
        ruby-version: ['2.6', '2.7', '3.0']

The failed workflow result is here: https://github.com/ntkme/debug-setup-ruby/runs/3648601998?check_suite_focus=true

Every thing other than windows + ruby 2.6 passed really quick. I killed the windows + ruby 2.6 after 15 min, but in reality it get stuck forever and would get killed after 6 hours by GitHub actions.

What works and what doesn't

bundler-cache	cache-version	os	ruby	dependency (sass-embedded)	works?
true	0	macos	3.0	yes	yes
true	0	ubuntu	3.0	yes	yes
true	0	windows	3.0	yes	yes
true	0	macos	2.7	yes	yes
true	0	ubuntu	2.7	yes	yes
true	0	windows	2.7	yes	yes
true	0	macos	2.6	yes	yes
true	0	ubuntu	2.6	yes	yes
true	0	windows	2.6	yes	stuck forever 1
true	1	windows	2.6	yes	stuck forever 2
true	0	windows	2.6	no	yes 3
false	0	windows	2.6	yes	yes 4

1. The default matrix works for all combination except for windows + ruby 2.6.
2. Attempt to use a different cache-version does not fix the problem so it is not a cache corruption issue.
3. Remove dependency sass-embedded works. - However, this does not appear to be the root cause.
4. If I copy the bundle commands from the log of failed build and run them as a separate step, It works. This shows there is something weird going on with setup-ruby causing the failure. See below.

---

This is a diff showing the how I moved the bundle commands out of the ruby-setup action:

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index e368543..0f274e6 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -20,4 +20,11 @@ jobs:
       uses: ruby/setup-ruby@v1
       with:
         ruby-version: ${{ matrix.ruby-version }}
-        bundler-cache: true
+        bundler-cache: ${{ matrix.os != 'windows-latest' || matrix.ruby-version != '2.6' }}
+
+    - name: Bundle install
+      if: ${{ matrix.os == 'windows-latest' && matrix.ruby-version == '2.6' }}
+      run: |-
+        C:\Windows\system32\cmd.exe /D /S /C "C:\hostedtoolcache\windows\Ruby\2.6.8\x64\bin\bundle.bat config --local path D:\a\debug-setup-ruby\debug-setup-ruby\vendor\bundle"
+        C:\Windows\system32\cmd.exe /D /S /C "C:\hostedtoolcache\windows\Ruby\2.6.8\x64\bin\bundle.bat lock"
+        C:\Windows\system32\cmd.exe /D /S /C "C:\hostedtoolcache\windows\Ruby\2.6.8\x64\bin\bundle.bat install --jobs 4"

This build shows that running exactly the same bundle commands outside of setup-ruby works: https://github.com/ntkme/debug-setup-ruby/runs/3648668586?check_suite_focus=true

---

What I need help understanding is that - what is the difference running bundle commands inside setup-ruby action and running exactly same command as a separate step, and why setup-ruby action might cause bundle to stuck forever.

Thanks.

[MSP-Greg]

@ntkme

Sorry, I've been busy this morning. Is this working as expected using Ruby 2.7 and 3.0?

[ntkme]

@MSP-Greg Yes, you can see from here: https://github.com/ntkme/debug-setup-ruby/runs/3648601998?check_suite_focus=true

2.7 and 3.0 works under windows, ubuntu, and macos. 2.6 works for ubuntu and macos.

[ntkme]

The most strange part as I posted originally is that: if I copy the bundle commands printed by the setup-ruby actions, disable bundle-cache, and just run them in a different step, it would work under windows + ruby 2.6. Meanwhile, it is not a cache corruption as update cache-version does not work.

[MSP-Greg]

I repo'd the issue locally. I haven't had time to debug (and may not for a while), but in the ext/sass folder, 2.6 downloaded the zip files, but did not extracted them. Ruby 2.7, 3.0, and master all successfully installed, and all of their folders had the zip files extracted.

Maybe that helps...

[ntkme]

That did help. I managed to get a windows machine to test this and the root cause is insane.

It has to do with running cmd.exe /c del /f /q /s filename from Makefile.

Even after finding the root cause I still don't understand what kind of interpretation difference is happening causing the different behavior of different ruby versions, but apparently under windows ruby 2.6 (maybe due to the old gnu make it bundles), cmd.exe /c del /f /q /s turns out to be interpreted just as cmd.exe and all the arguments are magically getting ignored. Without the arguments, cmd.exe will hang there waiting for user input thus causing gem install to hang.

I did find a workaround that works for ruby 2.6 inside the Makefile: powershell -c "cmd.exe /c del /f /q /s" filename.

[ashmaroli]

apparently under windows ruby 2.6 (maybe due to the old gnu make it bundles), cmd.exe /c del /f /q /s turns out to be interpreted just as cmd.exe and all the arguments are magically getting ignored.

@larskanis Hello, is this something that could be addressed by RubyInstaller..?

[MSP-Greg]

maybe due to the old gnu make it bundles

GitHub Actions and my local system use one set of MSYS2 built tools. So, both in Actions and locally, the passing Ruby versions use the exact same set of MSYS2 tools as Ruby 2.6.