To pick up the caching discussion again, I've started to experiment with a couple of different caching techniques. The headline results: a cold-start scan of entry points goes from about 4.5 seconds with no caching, to 0.45 seconds with caching. There's only a small difference (for me) between having a cache file per sys.path directory and a single per-user cache file. I've posted lots more details here: https://github.com/takluyver/entrypoints/issues/16 I keep my sys.path quite short, and the difference between the two approaches presumably increases the more sys.path entries you have. If anyone has time to test the different techniques, I'd be interested to know what your results look like. I've put (Linux) instructions on the same issue. If you'd like to try with a different format or caching mechanism, feel free to add some scripts and see how they compare. I haven't yet tried to integrate either technique into the entrypoints module. Thomas