[Starlingx-discuss] Broken build due to removal of irqbalance

Dean Troyer dtroyer at gmail.com
Fri Aug 3 14:45:16 UTC 2018


On Fri, Aug 3, 2018 at 9:31 AM, Scott Little <scott.little at windriver.com> wrote:
> http://zuul.openstack.org/
>
> So it seems there can be multiple queues per git.  Withing a queue there are
> dependencies, but different queues can make progress independently?  Or is
> there a queue of queues for the git?

[0] is the Zuul concept page, but basically it has a set of pipelines
for different job types, the two we care the most about are check and
gate.  Check jobs run on every Gerrit submission, gate jobs run after
Workflow +1 is set.  There is overlap between those job sets (check
may have run last week) but some things like non-voting jobs don't run
in the gate.

The other pipeline we'll be using is experimental, for on-demand runs,
that's where I plan to put the initial py3 jobs for example, so we can
see where we are but not waste resource running them all the time.

> On top of that, it seems like there are a finite number of execution
> engines.  If no execution engines are available, none of your queues will
> progress.  Is that about right?

Not just about, that is it exactly.  An older Zuul status showed the
VM allocation graphs at the bottom, I'm not sure where those went
after the Zuul v3 upgrade...

All of OpenStack CI is run on donated cloud resources (mostly
single-use VMs) from places like Rackspace, Vexxhost, OVH, Dreamcloud
and about 7 more that I don't remember offhand.  We have a quota on
each cloud and the load during North American working hours usually
puts us way over.  This is a big part of why Zuul was born, to
dynamically manage that ever-changing pool of test resources (VMs).
It turns out that hosting OpenStack CI is an excellent cloud load
test, we've found more than a few scaling problems this way.

dt

[0] https://zuul-ci.org/docs/zuul/user/concepts.html

-- 

Dean Troyer
dtroyer at gmail.com



More information about the Starlingx-discuss mailing list