I don’t really do any Python development myself, but at work I do support some automated testing infrastructure (in this particular case, I’m talking about a CruiseControl.rb installation), and some of the projects that use that infrastructure use Python. The setup is so that the tests are actually executed in the CC.rb server, so I have to have Python installed there, and it has to have some basic dependencies to be able to run the tests.
A couple of times something strange happened: suddenly, those tests would start failing with no apparent reason, and looking at the logs, it looked like some dependencies were not installed (error messages such as
ImportError: No module named sqlalchemy). Of course that didn’t make any sense, because SQLAlchemy is needed for the tests and they were working like a charm for weeks. I was totally and completely confused by the error message, and I tried to install SQLAlchemy again. That solved the problem, luckily, so I decided to forget about it because it wasn’t my thing anyway.
But the problems appeared again. And again. And I got another error message that was really confusing, because it looked like Python was using some old version of some module (a version that wasn’t there anymore, because the code had been updated from SVN). So I just got tired of not knowing what was going on, and decided to investigate enough to find out the root of the problem. And I found something surprising.
What I found is that the famous
python setup.py develop (that everyone told me to use) actually adds the “current directory” to the list of paths where Python searches for modules, so you can develop your module and use it from anywhere. I had heard some comment on that, but I didn’t quite get what it meant, and I don’t think the person that said it realised either.
The fun thing with
setup.py develop is that when you have several branches of the same project in the same machine, and you use that to make the modules available… well, I guess that knowing which versions of which modules Python will use becomes an interesting question to say the least. I’m not saying that the way it works is necessarily wrong, but I do think it is dangerous, and people shouldn’t think of it as the “normal” way of developing modules in Python. It should be used with care.
After having realised that and thought about it a bit, I still don’t understand why those modules simply “dissappeared”, but it seems that there was some corruption of
/usr/lib/python2.5/site-packages/easy_install.pth or similar (that file seems to be modified when you install packages with
easy_install, and it had references to the directories I ran
setup.py develop from, so that’s my main suspect for now). At least I know now that I could backup a working
easy_install.pth file, and restore when we have problems again, but I’m far from happy with that “solution”
Also, I’m wondering what the hell should I do in the future to prevent more problems, because using
setup.py develop sounds like a terrible idea to me. I tried to set
PYTHONPATH instead, but apparently I failed. Any suggestions?
EDIT: I’m finally using
PYTHONPATH. I have no idea what I tried last time, but using it was easy, quick and clean. I still have no idea why the hell Python sometimes forgets where some modules are, though.