R P Herrold wrote:
That is, why does yum do something different after a 'yum clean all'? Shouldn't it be trying all the mirrors anyway if it fails to get a file for any reason? Otherwise, what's the point of having the list that generally screws up caching?
I do not see that yum failover is **not** working; indeed it seems to be working just fine in my testing against a 'as designed' "centos-release" package as we ship it. The outcome _I_ see when I hit 'centos.mirror.nac.net' is the failure, a failover, and a success on a later listed peer. The 127.0.0.2 workaround will permit you to test this as well (simulating a dead mirrorlist entry); transparent proxies are out of our control by definition.
Dead mirrors are different from live mirrors that don't have a file. I can't simulate the latter, which seemed to be the source of the problem.
Most of my x86 updates required this:
yum -y --exclude tk --exclude linuxwacom --exclude wxGTK --exclude nss-devel --exclude 'rpm-*' --exclude popt --exclude kernel* update
on the first pass or they failed with one or more missing packages.
After this completed and I did a 'yum clean all' a subsequent 'yum update' picked up all of the rest, including the previously missing package(s). Proxies were involved in most cases, but not transparent proxies.
I am not so interested in trying slow motion debug via mailing list of what a person's setup is, and will (and do) read bugs.centos.org for reports from people who file a formal report
I don't know how to submit a sensible bug report about something that is not repeatable. I have an odd mix of 3rd party repositories enabled but that shouldn't cause a 'missing' package in the first place and didn't change between runs. It also seemed to be the same thing happening to others, since I got most of the --exclude's from the mail list postings. The only thing visibly different between runs was that after the "yum clean all" it went through the motions of recomputing the fastest mirror.
I don't think the answer is to expect the repos to be perfect but rather to make the clients recover without intervention (and without killing the good repos too...).
patches to yum upstream are still welcome, I assume. Feel free to ask Seth, et al.
Seth doesn't like my ranting any more than you do. Usually it is about the mirrorlist rotation screwing up caching - which isn't even his fault, but this time the download speeds were fast enough that I didn't care about the duplication. So your mirrors have great performance even if some contents are missing.