mark florisson, 27.03.2012 13:20:
I tested the release in my own branch and jenkins was blue, but the release build seems to disagree.
The release (and master) branch is tested against the "-ext" builds of CPython, which have some external packages installed, including NumPy. You have to change the config of your "tests" job to use those instead of the plain CPython builds. The build job can stay as it is. As for the inner workings, there's a normal "pyXY-hg" job to build CPython and a corresponding "pyXY-ext-hg" job that takes the build and installs a list of packages into it, then creates a new install archive. You can then reference either of the two archives in your build/test jobs by asking for a "pyXY" or "pyXY-ext" Python.
The py32 C++ build shows some refcount error:
numpy_memoryview.cpp:16922: warning: ‘void __pyx_pw_5numpy_7ndarray_3__releasebuffer__(PyObject*, Py_buffer*)’ defined but not used
These are legitimate warnings that are worth fixing (at some point), I think. They seem to originate from the buffer implementation in numpy.pxd. Those sort-of-external special methods shouldn't lead to the generation of a Python wrapper function.
python: Modules/gcmodule.c:327: visit_decref: Assertion `gc->gc.gc_refs != 0' failed.
It's surprising that that only occurs in one out of four Py3 test configurations. Maybe there's something indeterministic in those tests?
and some of the other python 3 tests are also failing.
Yep, some of them look really funny (complaining about getting exactly the expected output), others are the typical Py3 problems (e.g. printing bytes). Note that the reason the py3k tests are not impacted is that it does not have NumPy. So the tests would equally fail in all Py3 versions. Stefan