Moving coverage travis build from python2.7 to python3.5 - #5401

Wow, there is a massive decrease of lines covered, maybe @mhvk or @pllim you remember whether this is expected?

-0.04% does not look terribly massive. When I click on "coverage changed" tab in the report, seems like only astropy/_compiler.c has decreased coverage by one line. But it says "source not available" when I click on it to see which line. My suspicion is that it is probably a "skip this line if you are using Python 3" case. Did I interpret the report correctly?

As for the timing, comparing with the latest master build's coverage report:

platform linux2 -- Python 2.7.12, pytest-2.8.3, py-1.4.30, pluggy-0.3.1
Astropy version 1.3.dev1381
Numpy: 1.11.2
Scipy: 0.18.1
Matplotlib: 1.5.3
h5py: 2.6.0
Pandas: 0.19.0
11244 passed, 142 skipped, 74 xfailed in 455.48 seconds

And this PR:

platform linux -- Python 3.5.2, pytest-2.8.3, py-1.4.30, pluggy-0.3.1
Astropy version 1.3.dev1383
Numpy: 1.11.2
Scipy: 0.18.1
Matplotlib: 1.5.3
h5py: 2.6.0
Pandas: 0.19.0
8538 passed, 110 skipped, 69 xfailed in 512.58 seconds

Not super slow but still slower. No increase in performance. I wonder if it is due to all the if six.PY2 ... else ... (since the else part would take more time theoretically); If that is indeed the case, then this will be non-issue when we drop Python 2 support.

I remember that with python2 tests are run twice, once with unicode literals, once without, so that a decrease is expected (note the substantially smaller number of tests that were run in @pllim's example above).

Overall, my sense is to go for it now...

@pllim - There is some daily fluctuation in travis run times, so I think it's closer to the true difference to compare two coverage builds from the same travis run, e.g. using my testing branch:
py2 vs py3 is 512s vs 397s

https://travis-ci.org/bsipocz/astropy/jobs/166497098#L2550
https://travis-ci.org/bsipocz/astropy/jobs/166497099#L2538

@pllim - I think you're right, and I was the one who misread the report (I still think it's confusing). It's seems to be one line for almost every file, maybe there're the from __future__ lines?

@bsipocz , the two builds you linked to seem to both used Python 3.5...

I think the first one is python 2.7: https://travis-ci.org/bsipocz/astropy/jobs/166497098#L2109

@bsipocz , yes, you're right! I was confused by export PYTHON_VERSION=3.5 line at the beginning of the run. I apologize for the confusion. The time decrease in your reports is very encouraging! 👍

@pllim - no problem, I saw more travis log than my fair share to spot it :) The first exports are coming from the global settings, then they get overridden with the locals of the individual builds. I agree that it can be very misleading.

OK, I'll merge this, as I think it is good to be on python3 altogether, and it seems that there is no real downside left. Thanks, @bsipocz!