TST: Update travis.yml

[charris]

• Use 3.8 instead of 3.8-dev
• Use Python 3.7 for most tests.
• Re-enable --durations except for debug runs
• Add some blank lines for clarity

Perhaps we should spread the other tests across more Python versions?

[mattip]

i think it is good to thoroughly test one version. Is there a way to collapse all the python3.7 tests into one environment with many jobs to make it clear this is on purpose?

[charris]

The different tests are run with different environment variables so NumPy still needs to be rebuilt. I think building and testing takes up most of the time. The pickle test should be covered by default in 3.8, so that probably doesn't need to be there. There may be some other tests that can be eliminated or combined but I don't want to experiment with those things in this PR. This is just to get the tests ready for the 1.8 branch.

[charris]

I've posted on the list for a discussion of tests we might want to eliminate.

[charris]

A bunch of resize tests are failing on Python3.7. Any idea why?

E           ValueError: cannot resize an array that references or is referenced
E           by another array in this way.
E           Use the np.resize function or refcheck=False

Seems to have come in in #12121.

[charris]

My guess is that the refcheck in the code is failing in 3.7. Trying now with 3.8 to see if that is still the case.

[mattip]

I think there is some strange interaction between pytest --coverage, the resize tests, and python > 3.6. Still trying to hunt it down.

[charris]

Just to summarize, the resize tests fail on 3.7 and 3.8, but pass on 3.6. The pytest, wheel, setup, etc. versions are all the same. The official wheel builds pass on all versions.

@mattip Yeah, it is probably one of the "extras" in that test run.

[mattip]

It is --coverage which changes the allocators and messes with refcounts.

[charris]

@mattip Thanks for tracking that down. How did you determine that?

[mattip]

git clean -xfd; CFLAGS='-O0 -g3' gdb --args python -b runtests.py --coverage -t numpy/core/tests/test_regression.py, then setting a breakpoint in shape.c:113 where we emit the error. The strange thing is, python sys.getrefcount(x) is reporting 2, but inside the debugger PyArray_REFCOUNT(self) is 3, so maybe coverage is also messing with sys.getrefcount?

[mattip]

It must be some kind of meta-programming, I don't see where coverage changes any of what I claim it does, but the refcount is definitely 3 in the gdb debugger

[mattip]

Here is an issue CPython had with coverage and refcounts, but they closed it in 2011

[charris]

Given the Python dependency, I'm going to guess that something in sys.getrefcount changed. Or maybe something else used by coverage.

[charris]

The short term fix looks to be using 3.6 for that test, that gives us some time to deal with the problem.

[charris]

Anyone know why these environmental variables were added? I don't think we use any of those.

[charris]

Rebased.