Open Stack

On 2016-10-03 16:11:25 +0000 (+0000), Jeremy Stanley wrote:
[...]
> I think we can bring this thread to a close now. Upstream deleted
> the broken wheel from PyPI a little over an hour ago, and within
> about 5 minutes it disappeared from our CI system mirrors as well
> (PyPI deletions propagate automatically through our mirroring). At
> this point things _should_ be back to normal, and projects can
> probably start working on reverting or abandoning any temporary
> workarounds.

I just learned that we (unnecessarily) also copy available wheels
from PyPI into our separate wheel mirrors when building
architecture-specific wheels as a side effect of trivially reusing
the wheel cache to populate it. Since those mirrors are effectively
append-only we ended up persisting a copy of the broken wheel there
even after it vanished from our PyPI mirror, and so it continued to
be found by jobs in our CI system.

In the past few minutes I've cleared out vestigial copies of it
there, and we're reflecting on ways we can enhance our wheel
building jobs to omit duplication of wheels which are already
present on PyPI (which would be more intuitive and avoid situations
like this one). As far as I can tell things seem to be working in
the gate behind the constraints revert[*] now.

[*] https://review.openstack.org/381192
-- 
Jeremy Stanley