Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2023-11-08
11:01:58 sean-k-mooney as it can break flavors or iamges
11:02:02 sean-k-mooney with requried traits requests
11:02:09 bauzas new computes can stop reporting a trait, but the prefilter needs to speak old language until next SLURP
11:02:33 sean-k-mooney no that woudl still break explcit traits request in teh flavor/image
11:02:43 bauzas sean-k-mooney: right, but dvo-plv's spec is not exposing the trait itself, which is less tech debt
11:03:02 sean-k-mooney bauzas: it wont require it to be set manually
11:03:09 sean-k-mooney if that is what you mean
11:03:13 bauzas sean-k-mooney: that's why I think we said to avoid as much as possible to expose the traits in flavors
11:03:23 sean-k-mooney partly
11:03:44 bauzas sean-k-mooney: what I mean is that I would strongly disapprove the qemu spec if I hear people using the trait directly in a flavor
11:03:49 sean-k-mooney i dont like askign for things twice (once with a traint and again with an extra spec)
11:03:58 bauzas but I meh it, if we have a prefilter that hides that trait
11:04:26 bauzas operators using that trait directly for scheduling would then do something unsupported
11:04:27 sean-k-mooney right so in general a prefilter or and extention to ResouceRequest.from_request_sepc
11:04:40 sean-k-mooney is the preferd way to use reqirued traits for standard traits
11:05:10 sean-k-mooney ResouceRequest.from_request_spec is what you shoudl extend if its going to be in both the flavor and image
11:05:13 bauzas correct, standard traits should never be used directly in flavors, and I think we said that while ago
11:05:27 sean-k-mooney cool with me
11:05:41 bauzas only custom traits can be used in flavors
11:05:45 sean-k-mooney well
11:05:53 bauzas but I think we never documented that
11:05:55 sean-k-mooney no there are some standard taits that can be in the flavors
11:06:11 sean-k-mooney specificaly the one for cpu fetature flags
11:06:14 bauzas I know, and I'd want us to take a stand
11:06:24 sean-k-mooney i.e. ones that are not related to any flavor extra spec
11:07:13 sean-k-mooney so the nounce is standard traits shodl be avoided if they can be requested via an extra spec instead
11:07:35 bauzas that's sad, PTG is just behind us but I realized we never went to a conclusion on whether we should expose our traits in flavors and the consequences it has
11:08:20 sean-k-mooney well you knwo i generally dont like exposing resouces: or traits in the flavor/image when we can have an indriction via something else
11:08:37 bauzas ditto here, but again, we haven't documented that tenet
11:09:05 sean-k-mooney right but we cant say (you shoudl never request a standard triat in the flvor/image)
11:09:06 bauzas and operators could be trying to test fancy things by exposing those traits in a flavor
11:09:37 stephenfin zigo: nothing I can do about that, I'm afraid. I've pushed patches for Masakari and repeatedly trumpted the fact that this needs to be addressed in other projects
11:09:38 bauzas I'd rather do an explicit list of traits we allow to specify
11:09:39 sean-k-mooney specific these ones are genrally valid to request https://github.com/openstack/os-traits/tree/master/os_traits/hw/cpu
11:09:43 bauzas sean-k-mooneylike a public API
11:09:56 bauzas we have a stand on the contractual API we support
11:09:59 sean-k-mooney bauzas: we could always enchnae the flavor validtor
11:10:13 stephenfin zigo: Those will break when we uncap SQLAlchemy. They'll have to fix it then 🤷�♂�
11:10:13 stephenfin zigo: Those will break when we uncap SQLAlchemy. They'll have to fix it then 🤷‍♂️
11:10:35 opendevreview Steven Relf proposed openstack/nova master: Adding basic auth to dynamic vendordata api calls https://review.opendev.org/c/openstack/nova/+/900252
11:10:44 bauzas sean-k-mooney: I don't know what to do, that's the point
11:11:02 bauzas what's easier is when you have a prefilter or the requestspec translator
11:11:17 bauzas then we can say 'sorry, no, not intended to be used this way'
11:11:27 sean-k-mooney yep so when adding new extraspecs/image proeprts and a trait
11:11:32 zigo stephenfin: Thanks for letting me know, I'll see if we can invest time in at least cloudkitty that we use in production.
11:11:35 sean-k-mooney i think we shoudl alwasy prvied the prefilter/translator
11:11:54 bauzas sean-k-mooney: sounds a direction
11:12:19 sean-k-mooney and we could enhance the flavor validation api to also reject the use of the "resrved traits" in the future
11:12:41 sean-k-mooney bauzas: https://github.com/openstack/nova/blob/master/nova/api/validation/extra_specs/traits.py
11:12:45 bauzas I mean, for the short future, we need to agree between us as cores
11:13:15 sean-k-mooney yep
11:13:33 sean-k-mooney we shoudl likely update the contibutors guide or add a new refence doc
11:13:34 bauzas I'll capture that in a doc patch
11:13:42 sean-k-mooney i.e. how to design with triats and resouce classes
11:13:44 bauzas that's a start
11:14:14 sean-k-mooney i woudl add a new doc in https://github.com/openstack/nova/tree/master/doc/source/reference
11:14:27 sean-k-mooney or here https://github.com/openstack/nova/tree/master/doc/source/contributor
11:14:58 sean-k-mooney bauzas: on a related note i really dont like usign resouces: in genreal in the flavor either
11:15:00 bauzas yeah the latter was the one I thought
11:15:10 sean-k-mooney i kind of whish we had vgpu: instead
11:15:23 bauzas sean-k-mooney: I don't disagree
11:15:35 sean-k-mooney the main reason i say this is i dont like the way grouping works
11:15:46 bauzas particularly if a VGPU resource could be a mdev or a VF in the next future
11:15:51 sean-k-mooney i think that is leaking too much about how we modle things in placment
11:15:57 bauzas yeah
11:16:06 bauzas and it creates us some binding
11:16:10 sean-k-mooney ya so i woudl prefer vgpu: to work kind fo liek the pci alias
11:16:27 bauzas does it sound another action item for vgpus ?
11:16:33 sean-k-mooney so we have a layer of indirection
11:16:43 bauzas man, I really should invent the cloning machine first
11:16:48 sean-k-mooney am im not sure based on the lifespan of vgpus
11:16:59 sean-k-mooney but doign this as mdev: i could get behind
11:17:35 sean-k-mooney im expecgint that we will just use the pci_alias goign forward for sriov vgpus implemenations
11:17:55 bauzas actually, having mdevs: flavor resources sounds better for operators if we start support vfio-pci for GPUs
11:17:57 sean-k-mooney having this indrection for mdevs in general i thikn would be nice but not required
11:18:03 bauzas yeah
11:18:34 sean-k-mooney so effectivly we woudl extend the devices section or one of the dynmic section with an alias option
11:18:37 bauzas when I read my code last week, I thought I spoiled too many vgpu acronyms in the code while this can be mdev agnostic
11:18:49 stephenfin zigo: nw. The topic bauzas mentioned will bring up most patches, and I suspect there are solutions to most potential upgrade issues there
11:19:00 stephenfin topic:sqlalchemy-20
11:19:10 dvo-plv sean-k-mooney, bauzas I a little bit confused after this conversation. you talked regarding prefilter, as far as I understand about that https://review.opendev.org/c/openstack/nova/+/876075/25/nova/scheduler/request_filter.py#275
11:19:28 dvo-plv But i did not get if i need to change approach
11:19:31 zigo stephenfin: Yeah, I managed to upload Heat 21.0.0 with all the 12 patches already ! :)
11:19:37 zigo Currently working on Glance.
11:20:03 bauzas dvo-plv: no, you don't need to change the approach if sean-k-mooney and some other core agree on your updated spec
11:20:21 sean-k-mooney dvo-plv: that is fine in my view, no change required
11:20:32 sean-k-mooney dvo-plv: we just have 2 places where you can do this
11:20:44 sean-k-mooney dvo-plv: either as a trival prefilter there
11:20:50 bauzas based on the discussion I just had with sean at the last min, I could even approve the spec now, provided we draw the line
11:21:03 sean-k-mooney or in the schduler utils
11:21:51 sean-k-mooney dvo-plv: like this https://github.com/openstack/nova/blob/master/nova/scheduler/utils.py#L288-L304
11:22:13 sean-k-mooney dvo-plv: it has the same effect
11:22:37 sean-k-mooney dvo-plv: the only reason to prefer a prefiler over a translate function like that is if you want ot make it configurable
11:23:02 bauzas right
11:23:06 sean-k-mooney dvo-plv: in your case we dont need or want it to be configurable (its not in your patch)
11:23:08 bauzas you can opt-out
11:23:23 sean-k-mooney but we have mandaotry pre-filters too
11:23:28 bauzas (or opt-in, depending on the default value :) )
11:23:29 sean-k-mooney so either approch works
11:23:56 bauzas actually

Earlier   Later