I kind of don't like the idea of this, but if you really feel like you need it then go ahead and merge.

@jonparrott Here is the problem I am trying to solve:

• If local deps change, tox does not re-install them, which leads to inconsistent envs and forces tox -e ${ENV} --recreate any time core is touched

How big is the cost of tox --recreate? wheelcache should help significantly, unless for some reason some dependency can't be wheeled?

@jonparrott On my old-ish laptop with 16GB RAM the cost is prohibitive (it triples runtime):

$ cd ${GIT_ROOT}/datastore
$ rm -fr .tox
$ time tox -e py27
real    0m11.698s
user    0m9.456s
sys     0m0.804s
$ time tox -e py27
real    0m3.815s
user    0m3.412s
sys     0m0.344s
$ time tox -e py27
real    0m3.847s
user    0m3.448s
sys     0m0.308s
$ time tox -e py27 --recreate
real    0m9.772s
user    0m8.836s
sys     0m0.864s
$ time tox -e py27 --recreate
real    0m10.145s
user    0m9.184s
sys     0m0.796s

The whole point of unit tests is that they can be run during development, and 10s runs are not dev-user-friendly

TBH, 3.8s runs are also not dev-user-friendly, especially when the fresh sdist on every tox run eats most of that. Without tox:

$ source .tox/py27/bin/activate
(py27) $ time .tox/py27/bin/py.test unit_tests
real    0m0.685s
user    0m0.644s
sys     0m0.036s

For comparison, here are the times on this branch:

$ rm -fr .tox
$ time tox -e py27
real    0m6.930s
user    0m6.344s
sys     0m0.504s
$ time tox -e py27
real    0m2.079s
user    0m1.908s
sys     0m0.140s
$ time tox -e py27
real    0m2.196s
user    0m1.952s
sys     0m0.208s

Alrighty. I guess I can live with it until setuptools can get things fixed.

Thanks for weighing in. FWIW I'm not happy about this hack, but I'm even less happy about the Python ecosystem support for namespace packages.

@tseaver Would love to hear what you think

@dhermes please follow pypa/setuptools#250 and chime in if their solution doesn't seem to fix the issues here.

OK this has now been rebased on top of the simple fix in #2705, but I'd still like to avoid the massive build-time penalty if possible.

@tseaver PTAL

@jonparrott @dhermes @tseaver so it looks like this will never be fixed for Python 2.7 or possibly even 3.4.

Limiting support for this to 3.5+ only seems like a reasonable decision to me.

It seems like not having this PR, or something like it, could be a barrier to contributors.