Earlier  
Posted Nick Remark
#openstack-nova - 2018-02-23
15:26:00 dansmith mriedem: ugh, that max_number_per_host thing is just silly.. that's never going to do something useful, IMHO
15:27:58 mriedem i can see the use,
15:28:03 mriedem it makes the soft policy less soft
15:28:08 mriedem jello-like if you will
15:28:19 dansmith what's the point in that?
15:28:41 mriedem there is some nfv use case in the spec
15:29:04 dansmith I think from a modeling point of view, it'd be hard to reason about that unless you knew "I'm going to start ten of the exact same thing for throughput and I want those spread across three nodes"
15:29:20 mriedem anti-affinity to a point, so they don't need all VMs in the group on separate hosts, but they can only tolerate up to a limit of VMs on the same host
15:30:14 mriedem otherwise today if you have 10 VMs in a soft-anti-affinity group, and only 2 available hosts, you could get like 1 on 1 host and 9 on the other
15:30:19 dansmith right, so a more useful thing would be to say "instances in this group should spread across N hosts before considering doubling up"
15:30:31 dansmith that's an easier to reason-about model from the outside I think
15:30:44 dansmith and it doesn't depend on the size of the group
15:31:02 mriedem i can't say if that's easier to reason about from the outside or not,
15:31:04 dansmith it focuses on the amount of redundancy you need regardless
15:31:10 mriedem since i've had this in my head the way it's proposed
15:31:21 mriedem dansmith: mention it in the spec?
15:31:36 mriedem either way i need to bring it up at the ptg b/c of some questions about how to model the thing in the api and db
15:32:21 dansmith I really hate the idea of adding extra_specs to anything, but especially groups
15:32:44 mriedem i said that on the spec
15:34:48 dansmith I guess their use-case actually specifically requires this as the model
15:37:12 mriedem is sgordon still a man about town?
15:37:18 mriedem isn't he in the UK/Ireland area?
15:37:40 sgordon mriedem, nah canadia
15:37:44 mriedem drats
15:37:52 mriedem sgordon: can you take a look at that spec regardless of your physical location?
15:38:09 sgordon mriedem, yeah i just got the email notification and opened it up
15:38:09 mriedem gibi might also have some input here since ericsson added the soft affinity stuff
15:39:14 sgordon mriedem, i have to admit i havent had anyone clamouring for this in a fair while, but that may just mean folks are working around it using host aggregates and higher level orchestration
15:43:44 bauwser mriedem: dansmith: we have a clear distinction between filters which hard-check possibilities and weighers which soft-check
15:44:15 bauwser having some middleground is something more than just for affinity
15:44:22 bauwser tbc, I'm -2 on the spexc
15:45:26 bauwser mriedem: dansmith: but johnthetubaguy had an idea to deprecate filters/weighers and have a different semantic for saying "here I want to hard-stop or fail-safe"
15:46:09 johnthetubaguy bauwser: got a pointer to the new one
15:46:22 johnthetubaguy might be better than my previous re-write the world plan
15:46:24 dansmith bauwser: this is harder than soft
15:46:38 bauwser soft is weighting
15:46:43 bauwser hard is filtering
15:46:54 bauwser here they want to have a soft-but-hard
15:46:57 dansmith bauwser: right, this spec specifies something hard, which is filtering,
15:47:08 bauwser bref
15:47:12 dansmith but it weakens the normally hard anti-affinity for some value of allowed overlap
15:47:16 dansmith I really don't like it
15:47:20 bauwser it's unclear and I don't think modifying filters or weighers is a good thing
15:47:26 bauwser yeah exactly that
15:47:45 bauwser instead of hijacking what we currently have, we should discuss on a better approach
15:47:55 bauwser which is what was johnthetubaguy exactly trying to do
15:48:39 bauwser johnthetubaguy: what is the new proposal, tbh I haven't gone thru all specs yet
15:49:00 mriedem i'm not following you
15:49:25 mriedem you can do hard anti affinity and fail if you can't place the VMs on separate hosts,
15:49:35 mriedem you can do soft anti affinity and get more VMs on a host than you'd really like
15:49:48 mriedem this is saying, i want soft but i can only tolerate up to a limit
15:49:56 mriedem or, soft is ok but to a limit
15:50:00 dansmith mriedem: I see it as a variant of hard not soft
15:50:08 mriedem sure, it's in between
15:50:18 dansmith mriedem: it's the same filter has hard but instead of ==0, it's <=$max
15:50:27 dansmith you can't do this with a weigher is my point
15:50:30 johnthetubaguy dansmith: +1
15:50:36 dansmith and that's what soft anti-affinity is
15:50:47 mriedem a weigher is not being proposed
15:50:52 dansmith I know
15:50:56 bauwser mriedem: soft anti-affinity should never hard-stop
15:51:19 bauwser hard anti-affinity is something we already have
15:51:21 mriedem umm
15:51:29 dansmith mriedem: I'm saying this is proposing changing our hard filter from ==0 to <$max
15:51:30 mriedem so flacid anti affinity as a new policy?
15:51:40 bauwser what the proposer asks is a way to have a mechanism that would describe a 'max" limit
15:51:50 mriedem yes i understand
15:51:53 bauwser dansmith: and I don't like that
15:52:01 bauwser because it would confuse a lot of people
15:52:04 dansmith bauwser: neither do I, I think that's a confusing model
15:52:31 dansmith I would rather a policy of min-redundancy, with a value of the minimum
15:52:32 mriedem so your issue is creating a group with the soft policy but then having this attribute which makes it more hard than soft,
15:52:33 bauwser my point is, instead of trying to hack our current model, rather try to describe a better model
15:52:41 mriedem so it's a conception issue
15:52:45 bauwser exactly
15:52:46 dansmith mriedem: but it wouldn't be that
15:52:46 mriedem so a new policy
15:52:55 dansmith mriedem: it would be hard with a grace limit
15:53:01 bauwser sometimes people want to hard-fail on a filter, but sometimes they prefer to soft-fail
15:53:17 dansmith bauwser: I'm not sure what that has to do here,
15:53:31 dansmith because this is still hard fail on this filter, it's just at a nonzero level of duplication
15:53:42 dansmith if you go over that limit, which is currently zero, you hard fail all the time
15:53:42 mriedem ok, so create a server group with the anti-affinity policy (hard) with some limit on overlap
15:53:45 dansmith with what is proposed here
15:54:00 dansmith mriedem: that's what is being proposed here
15:54:08 mriedem no it's not, it's the opposite
15:54:14 mriedem the spec is talking about the soft policy
15:54:18 mriedem you're saying soft policy is bad optics,
15:54:24 mriedem so flip it to be hard policy,
15:54:28 mriedem with the overlap limit
15:54:39 mriedem either way we get to the same thing
15:54:40 dansmith mriedem: but the spec is wrong in how things work
15:54:42 leakypipes did someone say flaccid affinity?
15:54:49 mriedem leakypipes: yes, word of the day
15:54:49 dansmith mriedem: they think they're adding a limit to soft, but you can't do that
15:55:01 dansmith you can increase the limit on hard though
15:55:04 mriedem dansmith: b/c soft is enforced via weigher
15:55:07 mriedem sure,
15:55:10 dansmith mriedem: exactly
15:55:11 mriedem so yeah let's make that point in the spec

Earlier   Later