Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-11-27
14:53:05 efried gibi: I think what I'm asking (and maybe dansmith also could weigh in on this) is: In general if it's possible simply to use the flavor/image-meta to construct the request groups, like what these guys do:
14:53:05 efried https://opendev.org/openstack/nova/src/branch/master/nova/scheduler/utils.py#L141-L143
14:53:05 efried ...then would there be any other reason to use RequestSpec.requested_resources instead?
14:53:26 dansmith efried: that was supposed to be a hack
14:53:53 efried dansmith: which was? the thing I linked?
14:53:55 dansmith because the requested_resources stuff wasn't ready for ... something
14:53:59 dansmith efried: doing it via the flavor
14:54:53 gibi efried: how do you use flavor/image-meta to construct the bandwidth related request groups?
14:55:14 efried gibi: oh, you wouldn't necessarily for bw, I don't know.
14:55:21 dansmith efried: https://opendev.org/openstack/nova/src/branch/master/nova/scheduler/request_filter.py#L188
14:55:27 efried presumably that one is more involved, requires going back to neutron API etc.
14:56:01 efried dansmith: that makes it sound like using the *request spec* is the hacky one.
14:56:32 efried i.e. if we want to make sure we don't persist the "interpreted request groups", they're better off being put in the ResourceRequest constructor.
14:56:48 dansmith efried: it's using the *flavor* in the request_spec, instead of requested_resources
14:57:08 gibi efried: there is a gap, an itch, something. We have flavor to describe the nova related resource requst, we have the neutron port that describes the networking related resource request. And we need a common, internal place where we collect these. And that should be the RequestSpec. But today we have different ways to collect the resource request there
14:57:30 dansmith efried: there was a big thing when we merged that code, something that couldn't be done with requested_resources because of traits in the unnumbered group or some such, so that flavor-based approach was the workaround until that got resolved
14:57:38 dansmith efried: I'm sure mriedem remembers, but he's out
14:57:43 gibi we store a flavor in the RequestSpec but also a list of RequestGroups
14:58:40 gibi dansmith, efried: at the moment the requested_resources list does not have _all_ the request groups represented as resources reqeusted throigh the flavor are not added to the requested_resources
14:59:00 gibi so there is no unnumbered group there either
14:59:14 dansmith gibi: I don't remember the details, I could try to find that review
14:59:18 efried Right -- no matter what, ResourceRequest.__init__ is going to parse the flavor.
14:59:26 dansmith gibi: but we couldn't request the trait via requested_resources
14:59:58 gibi dansmith: yes, as you would need to add that to the unnumbered group but there is no unnumbered group in the requested_resources
15:00:24 dansmith ah okay I see
15:00:32 gibi efried: yes, at the end ResourceRequest.__init__ uses the flavor,image and the requested_resources to construc the a_c query
15:01:06 dansmith https://review.opendev.org/#/c/656413/2/nova/scheduler/request_filter.py@105
15:01:38 dansmith anyway, since the (ab)use of the flavor is only transient, not stored or exposed anywhere, it makes it fairly obscure what is going on, IMHO
15:01:55 gibi https://github.com/openstack/nova/blob/23995b4f9945aa242f6a9e7bd5bc700c27f31eb2/nova/objects/request_spec.py#L491
15:02:05 gibi here is a TODO that is hard to fix ^^
15:02:19 gibi as per my mail http://lists.openstack.org/pipermail/openstack-discuss/2019-May/005920.html
15:02:51 openstackgerrit Merged openstack/nova stable/stein: Add --dry-run option to heal_allocations CLI https://review.opendev.org/693199
15:07:33 efried it looks like
15:07:33 efried --> in _provision_instances we populate the RequestSpec.requested_resources, but only with part of the info available
15:07:33 efried --> from select_destinations we call resources_from_request_spec which
15:07:33 efried ---> calls ResourceRequest.__init__ with the RequestSpec, (that ignores the requested_resources and only processes the flavor/image)
15:07:33 efried ---> *then* appends RequestSpec.requested_resources into the ResourceRequest
15:07:34 efried then we call GET /a_c
15:08:32 gibi yepp that sounds like it
15:08:53 efried _provision_instances happens on the conductor side?
15:09:28 gibi I think it happens at the nova-api side
15:09:52 efried sorry, yeah, this whole thing happens on the controller side, duh.
15:10:21 gibi it is called from nova.api.openstack.compute.servers.ServersController.create during boot
15:38:54 efried gibi, bauzas, stephenfin: do y'all want to have a meeting tomorrow or should I cancel? (US folk will be out)
15:39:07 stephenfin I don't have anything to bring up
15:39:15 gibi me neither but I can be present
15:39:38 efried let me see if anyone has put topics on the agenda. If not, I'll cancel.
15:40:55 efried it looks like shilpasd wants to talk about disk provider stuff.
15:41:08 gibi I reviewd the sharing disk spec today it looks good to me
15:41:19 gibi but somebody else should check at least the API impact part
15:41:20 stephenfin Because it's stuck?
15:41:29 stephenfin Maybe we can unstick it now
15:41:39 efried I don't know. I saw gibi's review but haven't dug in
15:41:39 gibi the API impact part needs eyes I would say
15:42:01 gibi bauzas was involved but he is busy downstream recently
15:42:08 efried If nothing else it gives her a scheduled time to talk to a couple of cores. So, gibi would you mind running that meeting?
15:42:19 gibi efried: I can do that, yes
15:42:26 efried thank you sir.
15:53:38 bauzas gibi: efried: thanks
15:53:47 bauzas sorry for being busy at the moment :(
15:53:52 gibi bauzas: no worries
15:54:14 efried gibi, dansmith: stephenfin: can you see any problem with putting the appending of RequestSpec.resource_requests [1] directly into ResourceRequest.__init__ [2]?
15:54:14 efried [1] https://opendev.org/openstack/nova/src/branch/master/nova/scheduler/utils.py#L546-L551
15:54:14 efried [2] https://opendev.org/openstack/nova/src/branch/master/nova/scheduler/utils.py#L60
15:54:43 efried Since we're passing in the RequestSpec anyway, it makes sense to me to consolidate that.
15:55:32 gibi efried: that refactoring make sense to me
15:56:07 gibi efried: is it part of a bigger surgery series?
15:56:28 efried gibi: not specifically, no. It's a standalone "make this shit make more sense to efried" patch.
15:56:40 gibi efried: ack. go for it!
15:57:07 melwitt gibi: ack thanks
15:57:08 sean-k-mooney efried: sorry i was not following but im going to rework my prefilter patches soon
15:57:19 sean-k-mooney what is the tldr of what is beign changed
15:57:44 sean-k-mooney are we ditching adding the stuff to the falvor temportaly and moveing to use requested_resouces
15:57:51 sean-k-mooney which i know was the long term plan
15:58:20 stephenfin efried: Looks sane. I'm not sure why I didn't do that when I refactored all that so might be worth giving git-blame a quick whirl
15:58:32 stephenfin but otherwise LGTM
15:59:21 efried stephenfin: it makes sense why you didn't do it when you refactored ResourceRequest, because (code stack wise) resources_from_request_spec is not actually "near" ResourceRequest.__init__.
15:59:26 efried they're sort of cousins
15:59:51 stephenfin ah, makes sense
16:00:24 efried and theoretically there would have been more extensive test fallout. Though practically I'm not sure there will be any test impact, as the places that use ResourceRequest.__init__ either don't care about RequestSpec.resource_requests, or *should* care and are currently ignoring it.
16:00:25 sean-k-mooney efried: is the requested_resources in the request spec still just resocue request and not traits
16:00:45 efried sean-k-mooney: It's a list of RequestGroup
16:00:49 efried RequestGroup includes traits
16:01:01 efried well, theoretically it does.
16:01:07 efried I don't know if it actually does for bw
16:01:10 efried but it will for accels
16:01:29 sean-k-mooney right but can i create a resouce group with no resouce requests and just traits
16:01:41 efried um
16:03:11 efried I don't see any reason code-wise that that shouldn't work
16:03:24 sean-k-mooney so in my prefilter changes i have this add_trait funciton https://review.opendev.org/#/c/665775/21/nova/scheduler/request_filter.py@52
16:04:12 sean-k-mooney which when i wrote it i wanted to refactor all the exsiting prefilters to use
16:04:23 efried sean-k-mooney: ...except that I don't think nova is yet using the placement microversion that allows a resourceless request group.
16:04:55 sean-k-mooney but at least in the prefilter case for image type or the device modeles which im addeing supprot for there are not resouce requests
16:04:58 sean-k-mooney right
16:05:02 efried So: I thought there might be some places in the nova side where we check and forbid resourceless request groups, but I don't see them on a quick skim
16:05:02 efried But: placement is going to bounce it if you try.
16:05:51 sean-k-mooney right so i woudl either have to continue to modify the flavor for the traits case so they get included in the unnumbered group
16:06:06 sean-k-mooney or we woudl need to increase the microverion we use
16:06:22 openstackgerrit Alexandre arents proposed openstack/nova stable/train: Do not update root_device_name during guest config https://review.opendev.org/696339
16:06:44 efried which would be nbd, and we're going to want to do it very soon anyway.

Earlier   Later