run-mirror does not populate enough versions

[Robert Collins]

See bug 1222306 and bug 1222308 for examples.

What's happening is that these projects have tighter constraints than the global requirements, and so fail to install in a fresh mirror built just from the global requirements. This means that our production pypi.openstack.org mirror is now actually precious state - not just a performance / reliability optimisation. It also means that folk building their own mirrors have special failures :(.

[Robert Collins (lifeless)]

Note that the python-cinderclient issue is from python-cinderclient 1.0.5 - the most recent released version at time of print.

[Clark Boylan (cboylan)]

The trouble here is we install the current state of the requirements, rsync that to the mirror, then do that every day. Over time we build up a mirror that has historical versions of packages that allow us to work even when some projects don't follow the global requirements. I think this is something we should try to address.

Also, there is a proposed summit session to talk about the requirements repo and mirror to sort out what works and doesn't and how we can make it better. This bug seems like the sort of thing we should discuss there.

Finally, we are trying to keep projects more closely in sync by automatically proposing changes to the individual projects when the requirements repo updates.

[Jeremy Stanley (fungi)]

I believe this will be solved for any projects covered in the requirements integration test, between that and requirements update proposals once they all get more in sync?

[James E. Blair (corvus)]

This should no longer be an issue due to the switch to bandersnatch.