8:41 p.m.
On Mon, Aug 8, 2016 at 4:40 PM, Kevin Fenzi <kevin(a)scrye.com> wrote:
Greetings.
At the recent flock conference Infrastructure workshop, we had a nice
lively discussion on a number of items.
Sorry I missed it, the first Flock/FUDCon in a long time that I
haven't at least *planned* to attend :). Good news is that I'm in good
health and one piece still :D
* Containers in Fedora Infrastructure:
* We want to look at moving things that make sense to containers.
* A good initial candidate is the mirrorlist servers.
* Would use the existing OSBS build system to build them.
* Would run on proxies.
* Would have haproxy list their socket as primary and old
mirrorlists as secondary.
* The container would have mirrorlist-server wsgi in it along with
the pkl updated hourly.
* Could allow us to spin up more as needed, but also should allow
faster answers from proxies as they don't have to depend on or
query over the vpn.
Sounds like a decent proposal, but I'd like to challenge with the
question "why are we doing this?" Is it just because "containers, new
shiny things, yay!" or is there some real advantage that having this
in containers brings us? How is lifecycle management of these things
accomplished? What happens when a container host dies? Lots of
questions to be answered here. Not that I necessarily think that it's
a bad idea, but containers are disruptive to existing operational
models, and we have to be certain that we're prepared for that.
* Contributor resources in the fedorainfracloud
* Once our cloud is upgraded, we can use ipsilon to let users login
to the cloud and spin up instances for Fedora related needs.
* Outgoing restrictions would be added on port 25 and the like
* To start with users would only get 1 external floating ip.
* Initial rollout would enable qa and packager groups, need to see if
docs and i18n or other groups would have a use for it.
* would note that we can terminate any instance for any reason.
* Patrick would write some scripting to notify users after some time
and terminate if we didn't get an answer back.
Even with all of these restrictions, I think that the potential for
abuse exists. I'd like to see these instances (which I think that it's
a good idea to have) be locked down tighter than Fort Knox :). I think
that the only thing that they should be able to externally communicate
with should be the rest of Infrastructure (koji and the like) just
like anybody on the Internet communicates with those services. They
should be no more trusted than the computer that I'm writing this on
right now is, and perhaps even less so (if that's possible and makes
sense).
* Long term instances should be moved to persistent infra
playbooks.
Yep.
* Build setup and requirements for infrastructure applications.
* Will get releng to set us up some side tags that we can build from
src.rpm in.
* all prod builds to be done in koji.
* Up to maintainers what priority they place on getting into
EPEL/Fedora. Encouraged for many reasons.
Nothing earth-shattering here :). All good ideas :).
* FAS3 status
* Was running in staging, but we disabled for now until we can finish
a security audit.
* Need to get python-fedora changes lined up and ready/pushed out.
* Need to get fas3 fas_client packaged and ready to go.
* Need more testing in staging.
* Hopefully move production over after f25 is out.
* Fedora Infrastructure support setup
* Talked about on the list a fair bit.
* Support can be determined by looking at the domain:
fedoraproject.org - full 24x7 support, monitoring, uses RFR
stg.fedoraproject.org - 8x5 support, monitoring
fedoracommunity.org - some support, monitoring, uses simple RFR
fedorainfracloud.org - unsupported, apps run by contributors
+1
* Fedora CA and cert infrastructure.
* Current CA expires in 2018.
* Plans being worked on now to back fas3 with freeipa so we could
move to kerberos tickets for koji then
* Need to figure out what would need to happen to sigul for that.
* Wait and see pending freeipa/fas3 integration.
This sounds reasonable, but what happens if for some reason the
integration doesn't happen or doesn't happen in a timely manner? I
don't think that renewing the cert would be a huge deal, because the
users of the certs generated by that CA are a well-known quantity
(packagers and releng). The support burden of swapping out the CA cert
I wouldn't think would be *that* bad. I'm not sure off the top of my
head, how often do the user certs expire?