> * "-j1" would explicitly turn off multiprocessing
Running tests "sequentially" but run them in one subprocess per test
file is interesting for tests isolation. Runing tests one by one
reduces the risk of triggering a race condition (test only failing
when the system load is high).
-jN was always documented as "use multiprocessing".
Maybe we need a new option to explicitly disable multiprocessing instead?
vstinner at apu$ ./python -m test
Run tests sequentially
vs
vstinner at apu$ ./python -m test -j1
Run tests in parallel using 1 child processes
By the way, Python 2.7 behaves differently and it's annoying:
vstinner at apu$ ./python -m test -j0
Run tests sequentially
I'm in favor of modifying Python 2.7 to detect the number of cores for
-j0, as Python 3.6 does, and run tests in parallel. Python 3.6:
vstinner at apu$ ./python -m test -j0
Run tests in parallel using 10 child processes
About the default: run tests in parallel or -j1 are the two most
reliable options. While -j0 is faster, sometimes it triggers race
conditions. I'm not sure that it's safe to change that, at least maybe
don't do that in stable branches but only master?
Note: Obviously, I'm strongly in favor of fixing all race conditions.
I'm doing that for years. We are better today, but we are still not
race-condition-free yet.
Victor