On Travis, follow package dependency graph to expand package list.

[dhermes]

To make computing the dependency graph easier I moved the requirements of each package into a requirements.txt file.

I'm happy to instead write a custom setup.py parser but decided that the route of less code was "better" in some sense.

[daspecster]

@dhermes can you fix the conflicts?
Otherwise this LGTM.

[tseaver]

I am -1 on the "parse requirements.txt" approach:

• The requirements.txt format is effectively an internal implementation detail of pip: they provide no public API for parsing it.
• In the words of the PyPA docs and @dstufft, it is intended to support "application deployments", not as a substitute driver for install_requires / "library requirements."

I believe we would be better off leaving install_requires as the definitive version spec for the library, and using the pkg_resources API in our scripts to get the dependency metadata we need.

/cc @jonparrott

[dhermes]

@tseaver SGTM. @jonparrott and I discussed and I view pkg_resources as inferior to parsing because it requires having those packages installed somewhere before we decide we want to run those tests.

I'll whip up a proof of concept that parses our setup.py files, assuming some kind of format, and enforces the assumptions with errors.

[tseaver]

@dhermes To do the analysis on Travis, we are going to have them installed somewhere, likely into a tox environment. If we add a console_script (maybe as an extra) which is installed into the environment, then we can get the metadata from it.

[dhermes]

@tseaver PTAL at #2498