[release-1.12] Revert "Scheduler: Use a "scheduler" task for thread sleep (#57544)"

[kpamnany]

This reverts commit 640c4e1.

See discussion.

Cc: @KristofferC to confirm that this PR is done correctly (to release-1.12).

Cc: @Keno to confirm whether reverting this on 1.12 alone is fine or if you want this reverted on master.

[Keno]

I'm ok with 1.12 only

[kpamnany]

From the discussion in #58689:

Do we not properly try/catch/restart that task?

I was thinking of adding exactly this as an alternative to backing this out entirely. This patch can be considered a bugfix as it did close 2 issues.

We don't, but even if we did that, that would just result in us swallowing all InterruptExceptions.

The scheduler task should only be running if there are no other runnable tasks. If we add a try/catch to the scheduler task and swallow InterruptExceptions, then that seems to me like an adequate stopgap solution until there is action on #52291.

Thoughts @Keno and @vtjnash?

[gbaraldi]

Can you open a PR that does the try catch behaviour? Just so that we can see how it behaves?

[kpamnany]

Can you open a PR that does the try catch behaviour? Just so that we can see how it behaves?

See #58765.

[Keno]

I would pretty strongly advocate for a straight revert here. The problem with underdefined behavior is that people will have written their existing code to whatever happened to work, so even small changes are likely to break people's workflows. I don't see that the motivation to let the task be GC'ed slightly earlier is sufficiently strong to force this through and we don't have enough time left to really let people try a new behavior to assess impcat.

[fonsp]

From the Julia IDE perspective, I would really like to see interrupt testing in Julia CI to prevent issues like this in the future. Interrupt is a crucial feature for an interactive programming language. If interrupt works reliably, (new) users will feel much more comfortable with running and trying code, which speeds up the developer process.

I wrote a testable code snippet in #58798 (comment) that you can use.

Malt.jl also has many interrupt tests (about Distributed and Malt), and one reason for splitting this out of Pluto.jl was to get reliable coverage from PkgEval. So it's frustrating for me to see that PkgEval was not used in #57544 which would have caught this issue.

[nsajko]

I would really like to see interrupt testing in Julia CI to prevent issues like this in the future.

I guess you're right, but this tangent seems slightly off-topic here and might be forgotten. You probably want to open a new issue or PR, regarding regression testing specifically.

[Keno]

unrelated?

[kpamnany]

Yes. Sorry, will ensure that only the PR changes are reverted.

[kpamnany]

Huh, that was a PR change! 🤔

[kpamnany]

Oh, right, I fixed this test.

[KristofferC]

Changed PR to be against backport branch.