I classify this between a gotcha and a bug.

description:

invoking pp.server() with no args, lauches the cluster on the local machine.  It places the current working directory at the head of the python path.

invoking pp.server() with args uses remote machines.  There a different python path is used since the current working directory makes no sense.

the problem:
1) this is different behavior than how standard python behaves.  Normal python adds the path to the source code location to the python path not the current working directory.

Result:  Any program that tries to import modules will import different modules:
    the master program will import the module from the source code directory
    the jobs on smp nodes will import the module from the current working directory of the shell
    the jobs on remote nodes will use yet another directory.



Example of how this leads to a Gotcha:

imagine invoking a python program that is located outside the current working directory (/user/home)  like this:

$/user/home>  python  /some/path/myprog.py

inside myprog.py is the lines

import foo

js = pp.Server()
job = js.submit(func,modules = ('foo',)


What happens?

well myprog.py will look for foo in the subdirectory of the source code:   /some/path

and the parallel server will import foo from the current working directory /user/foo if there happens to be a module coincidentally named foo.py in the current working directory.  These can easily be different modules that share a name.   This happens a lot if you are developing code and your working directory has modules in it.

Solution:

the problem is NOT that parallel python failed to add /some/path to the python path on the server, (the way normal python does when you run a program in a different directory.)

This is not the problem because if one is running remotely then /some/path probably does not even exist.  The problem is that the server process is being launched in a fork that knows the current working directory.

two possible solution:
the server process should change it's current working directory to some empty temporary directory.  This will mimic how one would expect pp to behave if one were running on remote servers.

Alternatively, you could add a command flag to set the cwd of the server explicitly.