Test suite improvements #329
Conversation
nosetests will invoke the functions setup_module() and teardown_module(), if defined, before and after running all test cases within the module. In general, numpy arithmetic errors (e.g. integer or floating-point overflow, or division by zero) should not happen under normal circumstances, and the package should neither assume that these errors will raise an exception, nor assume that they won't. Therefore, while running the test suite, set the numpy error handling mode to "raise" so the test will fail if such an error occurs.
bemoody
force-pushed
the
arch-tests
branch
5 times, most recently
from
b1d119c to
bdd41c3
Compare
In addition to running the test suite on the "latest" OS images provided by GitHub, with the latest dependencies installed from pip, also run the test suite on 32-bit Debian 10 (via Docker), with stable dependencies installed from Debian. This should help to catch incompatibilities with older package versions, as well as errors due to differing CPU architectures.
|
Regarding the i386 tests: GitHub apparently has no problem running i386 containers with docker; however, the Of course it's possible to install amd64 libraries using But for purposes of this test I actually prefer not to have any amd64 libraries installed, to be sure we're not inadvertently relying on them and to keep things simpler and faster. So I've used |
|
Looks good, thanks! |
While running the test suite, set the numpy error handling mode to raise exceptions on arithmetic errors (as opposed to the default behavior of showing a warning, which doesn't cause the test to fail and is difficult to debug.)
In addition to the current matrix of tests, run tests in a 32-bit Debian 10 (oldstable) environment, in order to catch any errors that might result from different default integer sizes (is numpy.dtype(int) an int32 or an int64?) or different FP instruction sets, and also to check backward compatibility with stable versions of numpy/scipy/pandas/matplotlib/requests.