Coincidentally, we talked about this on the Julia HPC call yesterday (@jblaschke @staticfloat). The thing with one interpreter per thread is that you get in the “distributed business” of having to do something like @everywhere pyimport("numpy") and so on.