Closed odow closed 2 years ago
My experience has been from time to time some trivially small models fail on one OS over another for some numerical issue. This arrises in solver libraries (e.g. SCS is a common culprit) not the Julia code. So we have made a habit of testing on all platforms. That said, maybe the new JLL frameworks have mitigated this to some extent and we can be more restricted with what we test on.
Flakey tests here: https://github.com/lanl-ansi/Juniper.jl/pull/220#issuecomment-971156208
Closing because I think we need them all.
Do we really need to test 6 combinations? https://github.com/lanl-ansi/Juniper.jl/blob/df63a8d41b699be47d1eeecacef3ec2a6572c0e4/.github/workflows/ci.yml#L16-L22
Linux on the LTS and linux/Mac/windows on the latest point release should be okay.
I guess it's useful to test windows for the parallelism stuff, but there doesn't seem to be any OS-specific code (just the solver binaries), so I wonder if it's even worth testing on every platform. Has there ever been a case where things passed on linux but failed on Windows?