Speeds up the end-to-end running time for Travis tests.
* Adds a new jobs to test_install* separately (using -k "[not] test_install")
* More generic cache ignore
* Move slower Python 3.5 tests before 3.4
* Factor tox envs
Since de-vendoring support exists only for downstream, and they need
to patch pip to get that support anyways, it seems reasonable to push
support for testing that configuration onto them. This is something
they need to do anyways, since they need to test their versions of the
vendored libraries.
See https://github.com/pypa/pip/pull/4657 for more information.
* Upgrade to Trusty on Travis
Before Travis pushes us forcefully, it makes sense to ensure that the flip won't break our builds.
* Show failures because travis misbehaves occasionally
* Try out newer PyPy versions on Travis
* Use pypy and pypy3
🎉
* Revert "Show failures because travis misbehaves occasionally"
This reverts commit 7434fd6103.
* Allow PyPy3 to fail
* Try using a different version of pypy
* Try everything
* Switch to simpler names
* Try using container build with newest PyPy
* Use container build everywhere
* Remove a useless job
The wheel based vendor tests should get us the same testing without
the issue of trying to upgrade something that setuptools itself is
depending on. It also matches the mechanism that downstream are
currently using better.
PyPy isn't often packages by downstream, and even when it it is
it's often not got a pip package to go with it, and if it does
then it's probably not very likely that this code will fail on
PyPy but not on CPython 2.7.
So we'll just drop them so that the tests complete faster.
We're targeting these changes at downstream re-distributors and it's
unlikely that one of them will be willing to have a new version of
pip without the most recent version of Python 3.
The Travis OSX builders are incredibly inconsistent for how long
they take to run our test suite. It is currently ranging anywhere
from ~15 minutes to well beyond 50 minutes, at which point it
times out.