On 8/15/19 11:03 AM, Ian Jackson wrote:
> (off-list)
> Ian.

You replied off-list, so I wont quote you, though I would like my answer
to be public.

I probably should have mentioned that my remark was not so much related
to the crash, but more to what I experienced using Salsa's CI. It's
currently super slow, as apparently, it's using GCE using the smallest
instance possible (maybe because we have a limited credit?). As a
result, adding the salsa CI team's script just fails in some packages
like for example OpenVSwitch, because of the slowness. In the extreme
case of OpenVSwitch, during build, mostly all tests are timing-out. I
found the overall experience frustrating, and the Salsa CI not very
helpful because too slow.

So, instead of what I experienced, I thought using Zuul + Nodepool like
in the OpenStack CI would probably solve the slowness and resource
starvation issues. What's done there, is that OpenStack based cloud
providers are sponsoring a pool of VMs (. They are maintained using
nodepool, that "pre-provision" the VMs. That's a bit like with apache
prefork, where processes are spawned in advance to be ready for the next
connections, except here, we're talking about VMs ready to accept CI
jobs. Then Zuul does the job queue and scheduling. Overall, this is a
very efficient CI, which is able to spawn thousands of jobs, on many
different cloud providers.

I'm not sure exactly what the Salsa issue was. Whatever it was, I
strongly believe that Zuul + Nodepool would help:

1/ CI Jobs going faster
2/ Have a way more workers
3/ Don't have all our eggs on the same Google basket
4/ Use free software platforms instead of GCE
5/ Stop being bound to a single VM provider [1]

Not only OpenStack uses Zuul, big companies like Netflix and MediaWiki
foundations too.

I've read that this year, there was some efforts to add Gitlab support
to Zuul. I don't know what the status is though, but I know that it's
possible to add Gerrit in front, and then have Gerrit plugged to Zuul.
I've used Gerrit previously to do OpenStack packaging (the version of
OpenStack in Stretch was done this way), and I was very happy of it. I'd
love if I could have the opportunity to do this again, but with Salsa
and the Debian infra this time.


Thomas Goirand (zigo)

P.S: If you want an idea of what it is like with nodepool in action, go
here: http://grafana.openstack.org then click on "Home" and select
nodepool. There you'll see graphs of spawned instances.

[1] The OpenStack CI is currently configured to use VMs from:
Fortnebula, OVH (2 regions), Rackspace (2 regions), and Vexxhost, though
it used to have a large amount of sponsors.

Reply via email to