I recently had need to run a piece of Python code before every invocation of the interpreter. In my case, it was to enable coverage testing of processes spawned by a test suite. Certain tests would run Python code in a subprocess, and I wanted to measure coverage within that process.

I added a function to coverage.py that is suitable for executing on startup. It examines the environment to determine whether to start measuring coverage. (This code isn’t released yet, but if you want to try it, it’s on the trunk at bitbucket.)

Perl has a command line switch (-M) for this, as does Ruby (-r), but Python doesn’t. I came up with two ways to do it for Python, both depending on the site module which is imported automatically on startup.

The first way is to create a sitecustomize.py file on the path somewhere. This module will be imported automatically by site.py. Any lines you add to that file are executed when Python starts:

import coverage
coverage.process_startup()

This is simple, but editing an existing sitecustomize.py could be awkward, and you may not even have permissions to do it.

The second way is to create a .pth file with an executable line. Usually lines in .pth files are directory names, but if the line starts with “import “, it is executed (what a hack!), so you can create a .pth file like this:

import coverage; coverage.process_startup()

This has the advantage that you can create a new file without worrying about touching existing files, all .pth files will be examined and interpreted. Unfortunately, you might have to give the file a strange name, since they are executed alphabetically, and a package you need may only be available after its .pth file has been seen. In my case, I named my .pth file zzz_coverage_process_start.pth to be sure it ran after coverage was added to the path.

Neither of these feels clean or elegant. I’m wondering if there are better techniques I’m overlooking? Python has a clean way to register functions to be called at process exit (atexit), which has been designed so that multiple authors of exit functions don’t need to cooperate to avoid interfering with each other. It’d be nice to have a similar facility for startup functions.