[Ci-users] Strategy for scaling jenkins slaves ?

On 13/01/16 12:52, Karanbir Singh wrote:
> On 19/12/15 18:02, David Moreau Simard wrote:
>> The bulk of the jobs are offloaded to an ephemeral bare metal node but
>> ansible still runs from the slave. Some jobs do run locally, such as
>> smaller tox jobs.
>>
>> What's the strategy if we want to scale this ?
> 
> At the moment the slaves all run from VM's hosted near the admin and
> jenkins instances - and are setup manually, and managed manually - this
> was very much a stop gap arrangement till we can get a better
> virtualised setup in place. We've been looking at and trying to scope up
> getting an RDO cloud in place, that could then be used for 3 things :
> 
> 1) making an openstack api available for people who want to just consume
> VM's for their workloads
> 
[snip]
> 
> 3) offering up image backed resources for people looking at doing
> testing with other OSs, eg what the libvirt and libguestfs folks do at
> the moment.
> 
> We have a dedicated hardware slab ( ~ 24 phy machines worth ) dedicated
> to this task ( so as to not cut into the ci baremetal pools ), but are
> waiting on the RH facility folks to get it wired up and dial-toned.
> 
> Given the nature and impact of this setup, I am going to try and see if
> we can speed up delivery of that infra from the present timeline of end
> Feb '16.

I'd just like to chime in and say that the ability to provision VMs on
OpenStack (instead of C6/7 physical hosts) would be a big help for me on
migrating the Foreman tests to CentOS CI.  We currently run VMs on
Rackspace for CentOS, Debian, Fedora and Ubuntu so need a broad
selection of images which Duffy + bare metal hosts doesn't provide.

I've been trying to spin up VMs on the physical hosts for each OS, but
it's fiddly to set up compared to spinning up a pre-built VM image on
OpenStack.  (Little networking setup issues, virt-install workarounds,
bugs in our tests that assume certain package lists, and the complexity
of another provisioning layer.)

Looking forward to this, thanks for the update.

-- 
Dominic Cleal
dominic at cleal.org