mktempdir cleanup regression(?) in 1.8.3

[KristofferC]

The tests in Ferrite.jl started failing on Julia 1.8.3 on Windows with IOError: unlink("file"): resource busy or locked (EBUSY). An example CI run: https://github.com/Ferrite-FEM/Ferrite.jl/actions/runs/3569162068/jobs/5998825660#step:5:435

This used to pass on Julia 1.8.2. Maybe the libuv upgrade had something to do with it? It is possible that something bad is done in that CI but it is surprising that 1.8.3 made it fail.

[barucden]

Might be related to #39457?

[kagalenko-m-b]

This also affects me on 1.9.2, no fix nor workaround everywhere I searched. Can't precompile necessary for work packages on Windows as a result.

[mkitti]

See https://learn.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-gettemppathw

So perhaps changing the environment variable TMP or TEMP might help.

Also libuv xref:
https://docs.libuv.org/en/v1.x/misc.html#c.uv_os_tmpdir

[stevengj]

Could this be related to #34700? i.e. the problem is not mktempdir, the problem is that rm(dir, recursive=true) has sporadic failures on Windows because of race conditions.

[kagalenko-m-b]

Could this be related to #34700? i.e. the problem is not mktempdir, the problem is that rm(dir, recursive=true) has sporadic failures on Windows because of race conditions.

The error indeed is the failure to remove the temporary files created by the precompiling process. If the cause is some race condition, perhaps, I could add a sleep() in the downloaded sources somewhere that will slow down precompilation, but at least would allow me to work?

[mkitti]

Did things work with some earlier Julia version? Does this problem still occur if you start julia via julia --pkgimages=no?