Actually use the astropy nightly wheel.
#795
Conversation
github-actions
bot
added
the
dependencies
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## main #795 +/- ##
==========================================
- Coverage 77.07% 77.07% -0.01%
==========================================
Files 94 94
Lines 5680 5684 +4
==========================================
+ Hits 4378 4381 +3
- Misses 1302 1303 +1
*This pull request uses carry forward flags. Click here to find out more.
☔ View full report in Codecov by Sentry. |
Due to ASDF issues related to changes in how `numpy` 2.0 represents for the purposes of printing floats (used by ASDF), PyYAML can no longer handle the new representation gracefully. This change simply forces `numpy` 2.0.dev+ versions of numpy to use the `numpy` 1.25 representation format, which is supported by PyYAML. Note that this "fix" only works when running tests, not in general.
WilliamJamieson
force-pushed
the
bugfix/jenkins_devdeps
branch
from
74d8f30 to
f0a7653
Compare
|
Proof that it is now using a prebuilt dev wheel: https://github.com/spacetelescope/romancal/actions/runs/5674451888/job/15378017238#step:10:269 (line 268 of the tox output) |
|
This passes the regression tests: https://plwishmaster.stsci.edu:8081/job/RT/job/Roman-Developers-Pull-Requests/298/ Note that I did not run the Roman-devdeps tests on this because that would require me to modify the configuration of that test suite. Instead I modified the Roman-Developers-Pull-Requests job to both point at this version of Observe that the dev version of astropy (6.0.dev...) does in fact make it into the Jenkins regression tests job too:
|
| from astropy.utils import minversion | ||
|
|
||
| # HACK: This is a temporary workaround for ASDF not being able to handle how | ||
| # numpy 2.0.dev+ represents (prints) floating point numbers. This simply | ||
| # forces numpy to use the old printing method while running the tests only. | ||
| if minversion(np, "2.0.dev"): | ||
| np.set_printoptions(legacy="1.25") |
There was a problem hiding this comment.
This has been added only so that the devdeps CI and regression tests can be demonstrated to fully work by covering over a known issue in ASDF.
| @@ -1,7 +1,7 @@ | |||
| git+https://github.com/asdf-format/asdf | |||
| git+https://github.com/asdf-format/asdf-transform-schemas | |||
| git+https://github.com/asdf-format/asdf-wcs-schemas | |||
| git+https://github.com/astropy/astropy | |||
| --extra-index-url https://pypi.anaconda.org/astropy/simple astropy --pre | |||
There was a problem hiding this comment.
The extra-index-url listed in the tox.ini really ought to be moved here.
This is not a work around. Astropy builds a new dev wheel every week or so and uploads it here. It is my understanding that @nden wanted us to move all the pipeline development testing from building wheels ourselves to using the prebuilt development wheels. I do not think that we should be in the business of attempting to support building libraries that we do not have to. Moreover, this is a correctly functioning version of the astropy wheel building changes proposed by #695, which was approved without issue. |
|
I think the goal is to test working with the latest available dev version of astropy. We can assume they are testing their build system. Let's use the wheels for packages where available to speed up testing, since what we are testing is new functionality changes in dependencies, not their build systems. This is not a new way of running the devdeps tests afaik. |
|
Hi, It is my advice that you use the dev wheel that we provide instead of trying to build it yourself. We generate wheels for all the major architectures on versions that we support. If you switch to using our wheel and the system is unable to find one, it might be an indication that you have to upgrade stuff on your end. In addition, using wheel better reflects how your users will actually run the whole stack. No one sane is going to try to build Hope this helps! |
Currently, the Jenkins devdeps job is failing because something on the Jenkins machine has changed enough so that it can no longer build
astropywheels natively (some sort of compiler/base-library issue). Unlike the previous attempts at using theastropynightly wheel this PR actually uses that wheel.Also as a workaround to #793, this PR forces
numpy2.0.dev+ versions to use the legacynumpy1.25 floating point (print) representation. Note that this will only work around this issue for the unit tests. ASDF will need to resolve this issue in order to have a general solution.Checklist
CHANGES.rstunder the corresponding subsection