Bugfix/install missing compiler from buildcache#17536
Conversation
|
@haampie For me this fixes the issue you reported on slack, where your attempt to bootstrap compilers to build readline failed. Feel free to test it out and let me know if it solves that issue for you. |
94ba784 to
cfd4a1d
Compare
|
Whoops, I guess I had a gpg key trusted locally that prevented me from seeing the issue revealed here. But I removed it, and updated the buildcache creation and installation therefrom so they should work with out any need for a key. I guess we'll see though... |
|
To make everything work I need the following patches currently:
Hopefully they can all be merged. -- Edit: maybe that was a false positive actually... I just realized the pipeline succeeded because all dependencies were in the binary cache already. Let me retry. Edit 2: it still does not solve the problem :( details in #17544 |
|
Sadly, issues seem to remain with compiler bootstrapping pipeline, though this patch was needed in order to reveal the next problem. I'll continue to investigate. |
|
FYI @zackgalbreath and @tgamblin The workflow I set up for the spack-tutorial-container project depends on compiler bootstrapping, so this is important at the moment. I'm still trying to figure out exactly what we need for it to work again, but @haampie and I have been iterating on several issues that have cropped up while we weren't running any of those pipelines. |
tldahlgren
left a comment
There was a problem hiding this comment.
Confirmed tests pass on an LC machine.
|
Thank you for your contribution. |
Yes, slate started failing this week after the cuda 11 update went in, and although we fixed that in E4S by making slate depend specifically on the previous version of cuda, I didn't push this branch again after I made that fix. Thanks for noticing though, I'm glad folks are looking at those builds 😄
Awesome, thanks taking the time to check that. And thanks for the review and merge! |
Ensure compilers installed from buildcache are registered.
This PR fixes an issue I (and others) have seen when running compiler bootstrapping pipelines using the
spack cisub-commands. The first commit here adds a test that mimics what we do in pipelines, where the bootstrapped compiler is installed and put into a binary mirror in some stage of the pipeline, and then in a later stage, we attempt to useinstall-missing-compilers: trueto have that compiler installed from the binary mirror during installation of the spec that needs it.Without the fix in the second commit, the test in the first commit fails with the message:
spack.installer.InstallError: Cannot proceed with b: 1 uninstalled dependency: gcc... which corresponds to the issue I've seen in pipeline logs.