an example of py-pde error for which mpi4py.COMM_WORLD.Abort uses exit code 0 (hiding the problem downstream)

[Grzyb33k]

here's a minimal reproducer:

$ mpiexec -n 2 python -c "import pde;pde.PDE({'c':'laplace(c)'}).solve(pde.ScalarField.random_uniform(pde.UnitGrid([2,2]), 0, 1), solver='explicit_mpi',t_range=1)"
NotImplementedError: ("Cannot handle local boundary <class 'pde.grids.boundaries.local._MPIBC'>. Did you mean to use the `numba_mpi` backend?", 'Problems in boundary condition `auto_periodic_neumann` for operator `laplace` in PDE for `c`')
--------------------------------------------------------------------------
MPI_ABORT was invoked on rank 1 in communicator MPI_COMM_WORLD
with errorcode 0.

NOTE: invoking MPI_ABORT causes Open MPI to kill all MPI processes.
You may or may not see output from other processes, depending on
exactly when Open MPI kills them.
--------------------------------------------------------------------------
[lateralus:60536] 1 more process has sent help message help-mpi-api.txt / mpi-abort
[lateralus:60536] Set MCA parameter "orte_base_help_aggregate" to 0 to see all help / error messages
$ echo $?
0

(caught while debugging numba-mpi CI problem where a breaking change in py-pde API has gone unnoticed despite the above error in the logs)

Hope that helps

[Grzyb33k]

relevant mpi4py docs entry: https://mpi4py.readthedocs.io/en/latest/reference/mpi4py.MPI.Comm.html#mpi4py.MPI.Comm.Abort

[david-zwicker]

Thanks for the report! I think we fixed this problem in a recent update, which has not yet been released.

[david-zwicker]

I just published a new version, which should fix the problem

[slayoo]

Thanks!
Perhaps the distilled minimal reproducer could be used in a CI workflow (with an expectation of non-zero exit code) to ensure it works and to prevent regressions in future?

[david-zwicker]

I generally agree that this would be good to test for, but I genuinely don't know how to write a test.

First, the minimal example does not raise an error anymore, since the appropriate backend is now selected automatically (this should always have worked, but for some reason didn't). It's not difficult to come up with an faulty code, though: mpiexec -n 2 python -c "import pde;pde.PDE({'c':'undefined_rhs'}).solve(pde.ScalarField(pde.UnitGrid([2])), solver='explicit_mpi', t_range=1)"

Second, and more importantly, I don't know how to easily write a test that determines how MPI crashes and whether this is caught appropriately. I might need to do some research to understand the problem better, but since it works now the priority of this is quite low :)