Open Stack

Hi Chris,

I wanted to chime in as somebody who operates an environment where users make advanced reservations for bare metal nodes (we have ~500 in total) via Blazar, which are provisioned by Ironic. The way this used to work (pre-Stein) was that Blazar would create host aggregates in Nova associated with a reservation ID, and then a special Blazar scheduler filter in Nova would narrow the list of hosts to only the ones tied to a reservation ID (specified as a scheduler hint, I believe.) This worked well enough but ran a bit against the grain of how Nova was designed and we often experience issues with large launches due to Nova's design, whereby one design assumption is that there is significant headroom available, and thus retries are a sane and practical mechanism to step around provision failures. In our use-case, trying to provision 50 nodes against an aggregate with 50 hosts results in many provisioning errors because Nova isn't doing inventory management and will try to launch 2 bare metal instances on one host. Correct me if I'm wrong, but I think the Placement API addresses this better.

Anyways, there was a lot of work[1] done in Stein to change the approach for one aspect of Blazar reservations: instance reservations (VMs). With the new approach, which uses the Placement API, a custom resource class is created to represent the reservation. I believe it is the intent of the Blazar core team to implement this for host reservations, whereby a user reserves an entire compute host; this is particularly relevant for Ironic.

Advanced reservations are a thorny problem especially when you are dealing with more complex use-cases. To be honest the division of responsibility between Blazar, Nova, and Placement is still not clear to me. So, more conversation between these working groups, which you are trying to facilitate, feels like a good idea.

I'm a bit of an OpenStack newbie still, so I'm still working on wrapping my head around all the pieces and how they ideally fit together, so some of this may be nonsense, and that would be why.

Cheers,
/Jason

Chameleon Cloud
https://www.chameleoncloud.org/


[1]: https://review.openstack.org/#/q/topic:bp/placement-api
________________________________
From: Chris Dent <cdent+os at anticdent.org>
Sent: Tuesday, April 9, 2019 06:56
To: OpenStack-discuss at lists.openstack.org
Subject: [placement][ironic][blazar][ptg] "direct baremetal reservations" with ironic+blazar


The etherpad [1] doesn't have any additional notes besides the topic
but the general questions are:

* Do people want to do baremetal reservations with some combination
   of Ironic, Placement and Blazar, now or ever?

* What about Ironic and Placement, sans-Nova?

* If so, what would need to happen to make those things possible?

Based on my observations at the recent OpenInfra Days UK, I'd say
there's definitely an underserved need for this kind of
functionality. From the Placement side of things I would guess
there's little that would need to be done; it's more a matter of
getting other pieces and parts getting data in and out of Placement.

Discuss...

[1] https://etherpad.openstack.org/p/placement-ptg-train

--
Chris Dent                       ٩◔̯◔۶           https://anticdent.org/
freenode: cdent                                         tw: @anticdent
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-discuss/attachments/20190409/61fcfca2/attachment.html>