Installation issue: py-jupyter-server build error when using different gcc backend for Intel about spack

[climbfuji]

Thanks for your help. Let's keep this open for now. I've got a release to roll out where I am using my duct-tape solution to get it done. After that I'll fix it properly and submit a PR to spack develop.

[climbfuji]

@adamjstewart One workaround I found was to install node.js manually outside of spack with the correct gcc compiler loaded. That was surprisingly simple and allowed the build to proceed (just make sure the spack compiler config loads the node.js module).

[adamjstewart]

Should we add node.js as a dependency? Does that fix the issue?

[climbfuji]

Should we add node.js as a dependency? Does that fix the issue?

I assume this will work. My quick-and-dirty approach was to simply add the node.js paths into the compiler config on the system. I was a bit hesitant making it a dependency because node.js doesn't build out of the box in spack (where I tried it), but of course an external package should do.

[adamjstewart]

Hmm, I'm able to build py-jupyter-server 2.6.0 and 1.21.0 without node/npm, so we definitely don't want to make it a required dep. Maybe it's only used if it's found? Do you have the same issue with py-jupyter-server 2.6.0?

[climbfuji]

No, 2.6.0 builds fine. The package I am after py-cylc-uiserver has a constraint on py-jupyter-server<2 in the package config. I removed that for testing if py-cylc-uiserver worked fine with [email protected], but unfortunately it didn't.

[adamjstewart]

Looks like that constraint was dropped in py-cylc-uiserver 1.4 if you want to update that recipe.

[adamjstewart]

The best alternative is to patch out whatever calls node in jupyter-server 1

[climbfuji]

@adamjstewart I finally took the time to sort this out. The story behind this is: py-jupyter-server@:1 uses py-jupyter-packaging as build tool (v2 uses something else).

There is an optional feature to build the typescript extension when npm is found, but unfortunately this is all hardcoded to be auto-detected in the py-jupyter-server build and the underlying py-jupyter-packaging tool. The CLI of py-jupyter-packaging provides an option to skip npm, but this is not available when you import the Python modules as py-jupyter-server does.

I found a way around this, which is to add a variant typescript that is False by default. For ~typescript, I am patching (hacking?) pyproject.toml to pass a nonsense build argument npm=/dev/null to the npm_builder, which means it can't find it and gives up. For +typescript, I am adding the npm dependency explicitly to the package.

I tested both installs +/~typescript on the two systems that had errors with py-jupyter-server (the error that lead to the creation of this issue, and also a weird macOS where npm was found, but node was not), and they all worked.

This is all in PR #43279