On Wed, 2024-06-26 at 11:45 +0000, Graham Inggs wrote: > I suggest enabling the build time tests again, and request binary > removal on architectures where numba doesn't build. > > i.e. disable the "failsafe" introduced here: > https://salsa.debian.org/science-team/numba/-/commit/3d08b6e6fe43400e7640de51d8e5afb5942d79ed > > Look at the recent build log for s390x. It seems most tests are > failing. Is numba useful at all on this architecture in this state? > https://buildd.debian.org/status/fetch.php?pkg=numba&arch=s390x&ver=0.59.1%2Bdfsg-1&stamp=1719393837&raw=0
Numba has an extremely thorough testing suite, there's well over 10,000 different cases and running them all can take hours. Typically when running the full test suite on an unsupported architecture a modest number of test cases fail, but I don't know if 20 test failures out of 10,000 means the package is unusable for someone's specific use case? I personally wish libraries didn't have hard dependencies on numba given how frequently it breaks. I think the || true was to get the supported architectures to migrate out of unstable even though unsupported architectures can't succeed at the full suite. Because numba was running into buildd time limits I added --random 0.2 which randomly selects 20% of the tests to run. That also means builds are much more likely to finish on unsupported architectures, so I'll trying removing the || true and will try building with x86_64 and (emulated) arm64 and see how that goes. Reducing the number build time tests run does mean the autopkgtests (which is running most of the available test suite) are much more important to verifying that the package is working correctly. Diane
