Document OS X Python bug and submit a minimal test case

[bocajnotnef]

debugging led by @mr-c revealed the error plauging the OSX builder is in fact a python bug, solved by upgrading stock python (2.7.6) using python.org's 2.7.10

The problem appeared to be that the exception message was optimized out. We attempted to work around this by explicitly setting the exception text in read_parsers.cc instead of using the default constructor.

so, to do:

• install pythons 2.7.7+ (official releases) until we find which one fixes the problem
  • scan the release notes to see what they fix

if release note grokking fails we could disable that test for OSX.

@luizirber Got a good way to download/install multiple python versions (and/or make them coexist nicely?)

Bonus: test Python installed via Brew & Anaconda; also Python 3.3+ on OS X

[kdm9]

@bocajnotnef pythonz?

[bocajnotnef]

Ooh, I like.

@mr-c would that work for a jenkins environment?

[mr-c]

• split test into two: one that looks for the exception and the other that tests for the specific value exception message
• add a @attr to the second test to prevent running on OS X
• make a PR for the above while the rest of this issue is being investigated

[bocajnotnef]

Workaround up in #1192

[bocajnotnef]

python.org python 2.7.7 also passes.

Whee

[bocajnotnef]

python.org python 2.7.6 passes

What version of python ships with OSX, you ask? Why, 2.7.6 of course!

[mr-c]

@kdmurray91 alas pythonz compiles new Python; we are interested in testing the various binary distributions of Python

Since python.org's Python doesn't have the issue I declare this to be a problem with Apple's customization.

A very enterprising person could:

• construct an irreducible test case and file a bug

[mr-c]

The test case at 3450290#diff-0e869b7a17ce1f9ac85bc00b330c804fR327 triggers the bug

[kdm9]

Note that this bug or a similar one appears to affect Debian unstable too. See #885 for build logs etc.

Some gitter notes from @mr-c and myself

I'd just quilt patch as it is only a problem for Debian Unstable right now

A minimal test case + report upstream needs to be made at some point

It's not anything to do with how we store exception strings is it? does using char * instead of std::string change things perhaps?
I.e. dtors delete-ing the string after PyErr_SetString(msg.c_str()); (or whatever it is)

[ctb]

Fixed in #1192, I think.