IIUC the difference in the verification script is that we construct the wheel file name, download it, and install the file, rather than have pip query a repo and download the latest: https://github.com/apache/arrow/blob/master/dev/release/verify-release-candidate-wheels.bat#L69-L72
Verification failed first with a 404 on the wget, and then when the file path was fixed to reflect the actual file name (the first time around, with an extra m), python said that there was no package matching the ABI. If we were testing pip installing from a repository, this would have failed as soon as we started building wheels for py38. Neal On Fri, Feb 14, 2020 at 11:22 AM Krisztián Szűcs <szucs.kriszt...@gmail.com> wrote: > Actually we are verifying the wheels after building them, the manylinux > wheels in fresh environments using docker. > We should first investigate why didn't the windows wheel issue surface > during its verification [1]. > > [1]: > https://github.com/apache/arrow/blob/master/dev/tasks/python-wheels/win-build.bat#L102 > > On Fri, Feb 14, 2020 at 5:16 PM Neal Richardson > <neal.p.richard...@gmail.com> wrote: > > > > In terms of wheel verification, I added > > https://issues.apache.org/jira/browse/ARROW-7853 to the 0.16.1 tag. It's > > for CI using pip to install the wheels we build nightly. Obviously this > is > > not required for 0.16.1 but it would make the release verification > simpler > > because we would be running that logic every day and wouldn't have to > wait > > until release time to find issues. > > > > Neal > > > > On Fri, Feb 14, 2020 at 7:49 AM Wes McKinney <wesmck...@gmail.com> > wrote: > > > > > It seems inevitable that a few annoying regressions will pop up as > > > 0.16.0 becomes more widely deployed, e.g. ARROW-7841 was just reported > > > and patched. I created a 0.16.1 Fix Version in JIRA so that we can tag > > > issues that we may want to cherry pick into a maint-0.16.x branch. We > > > probably would want to wait a couple of weeks before cutting the patch > > > release, though. > > > > > > Also, as a matter of easing release management for patch releases, I > > > wonder if it makes sense to focus RC verification efforts on the > > > binary artifacts (especially the Python wheels). Should we write down > > > a "reduced" set of tasks for release managers to track for patch > > > releases? > > > > > > Thanks, > > > Wes > > > >
