Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2018-12-05
21:42:25 sean-k-mooney dansmith: well for the same reason as teh have the cpu allocation ratios vs inital cpu allocation ratios spec
21:42:28 efried agreed, all the inventory values ought to be owned by the virt driver, period. We start making exceptions, we end up with messes like allocation ratio... and reserved.
21:42:36 sean-k-mooney controling via api or config
21:42:57 efried but making that contract stick for this kind of workaround is going to be tricksy.
21:42:58 dansmith sean-k-mooney: that's not an example that helps your case I think :)
21:43:24 mriedem the alternative was a trait i think
21:43:32 mriedem and a pre-request placement filter in nova-scheduler
21:43:32 efried unless we do it with a handy-dandy provider config yaml file https://review.openstack.org/#/c/612497/
21:43:35 mriedem or something like that
21:43:51 mriedem this isn't blues clues
21:43:55 dansmith mriedem: yeah, that's also an option
21:44:28 mriedem so api sets a trait (or removes it), virt doesn't overwrite it, and scheduler filters on it (essentially it becomes the ComputeFilter)
21:44:31 efried omg, now you totally remind me of Steve from blues clues.
21:44:35 sean-k-mooney efried: provider config yaml for what?
21:44:43 mriedem efried: there were a couple of steves
21:44:48 mriedem or was it one and then another guy?
21:44:50 dansmith mriedem: not scheduler filters, request filter
21:44:51 efried no, there was only one steve.
21:44:55 sean-k-mooney how does that allw you to set it via the api and not have the compute overriede
21:44:56 efried the other guy was...
21:44:59 efried joe?
21:45:07 mriedem yes joe
21:45:17 mriedem http://www.gstatic.com/tv/thumb/persons/250022/250022_v9_ba.jpg
21:45:48 mriedem sean-k-mooney: b/c the compute/virt driver isn't supposed to overwrite externally set traits
21:46:01 mriedem and the api can set traits just like it mirrors aggregates
21:46:06 sean-k-mooney anyway if we say only the entity that created the RP may set the reserved value on the invetories im fine with that if we document it
21:46:58 dansmith I thought there was some other reason for doing the rpc cast on disable, but I can't remember what it was
21:47:46 dansmith maybe so ironic driver can do something?
21:47:52 sean-k-mooney on disableing a compute node. i think some dirver can take actions on disable and ironic might be one of them but i dont remember eiter
21:48:14 efried <do a right thing to disable the inventory>
21:48:14 efried if self.disabled:
21:48:14 efried we can ask each driver's upt to do
21:48:39 dansmith efried: we still have to do it as soon as the api is called though,
21:48:44 dansmith so waiting for the next periodic is not good enough
21:48:51 efried mm
21:49:08 dansmith oh yeah, so..
21:49:11 mriedem right i could disable an entire cell,
21:49:11 efried do we have a ComputeDriver hook that we call on disable?
21:49:15 dansmith disable is on the service not the node, right?
21:49:23 efried or could/should we implement one?
21:49:25 mriedem er half a cell, and then live migrate stuff within the cell
21:49:27 dansmith so if you disable the thing running on ironic, the nodes will still be schedule-able
21:49:30 mriedem dansmith: yes on the service
21:49:53 mriedem efried: we don't have one for os-services disable no
21:49:55 dansmith so you have to do the service->nodes translation, and set the trait on all the nodes
21:50:04 mriedem we had one for os-hosts update status, but that was something else that only xen implemented
21:50:09 dansmith but if you rpc to the compute, it gets to do that itself
21:50:10 mriedem dansmith: yup
21:50:11 efried um, or we could make upt that thing - i.e. call it immediately on disable
21:50:40 dansmith mriedem: that also gets wonky if you disable a service and then they rebalance, but you have set the trait on some nodes that rebalanced out
21:50:42 sean-k-mooney mriedem: so the trait way woudl be add status_down trait and change placment query to include status_down forbien trait?
21:50:46 dansmith I guess they'll eventually un-disable, but.. ick
21:50:52 dansmith well, not actualy
21:51:00 dansmith because we just said virt has to leave those alone if the api set them
21:51:18 mriedem sean-k-mooney: essentially yes
21:51:28 mriedem if ComputeFilter is enabled or something
21:51:30 dansmith efried: right point being we don't have an rpc call for disable right now, but I think we probably should
21:51:35 mriedem or some new boolean that replaces ComputeFilter
21:51:57 efried oh, because the disabling is being prompted not by the compute itself.
21:52:03 mriedem correct
21:52:04 dansmith it's an api call
21:52:09 dansmith in most cases
21:52:12 mriedem https://developer.openstack.org/api-ref/compute/#compute-services-os-services
21:52:14 dansmith however,
21:52:16 sean-k-mooney well the trait could elminate the host before the filster so we dont need the computeFilter anymore
21:52:24 mriedem if the hypervisor dies the libvirt driver auto-disables it's own compute service
21:52:28 dansmith mriedem: computes can self-disable, so you'd also have to have them be able to set/clear the trait
21:52:38 efried this would be a great use case for sean-k-mooney's reverse-required traits :P
21:52:39 mriedem i smell spec
21:52:52 dansmith if you just let the compute always manage this in response to an rpc call, the ownership of it is a lot more clear I think
21:52:59 efried ++
21:53:03 mriedem sure i'm fine with rpc call to compute to do the thing
21:53:18 sean-k-mooney efried: you mean the " you cant be schduled to this RP without asking for trait X"
21:53:23 mriedem btw this was the old thing from the os-hosts API https://github.com/openstack/nova/blob/master/nova/compute/api.py#L5026
21:53:25 efried sean-k-mooney: yeah, that one.
21:53:57 mriedem we could maybe re-use that
21:54:38 mriedem that was only used by this before https://developer.openstack.org/api-ref/compute/#update-host-status
21:54:41 mriedem which is deprecated
21:55:14 dansmith I wonder if jaypipes is okay with a status_enabled or status_disabled trait on compute nodes
21:55:35 sean-k-mooney mriedem was that repalce by the mark host down thing that we added for opnfv
21:55:46 mriedem sean-k-mooney: on
21:55:47 mriedem *no
21:56:02 dansmith m-h-d is not just for opnfv
21:56:08 mriedem sean-k-mooney: https://specs.openstack.org/openstack/nova-specs/specs/pike/implemented/deprecate-os-hosts.html
21:56:29 sean-k-mooney dansmith: ya i know but they where the group i had the most conversation with
21:57:46 sean-k-mooney mriedem the more you know. we have/had a lot of duplicate/legacy apis
22:30:56 cfriesen just saw mriedem's mail on placement and disabled computes...rather than mucking around with inventory values why not just explicitly add a separate field that says whether a given RP is enabled or not?
22:33:52 mriedem field?
22:34:00 mriedem omg have you talked to jay or chris ever?
22:34:11 mriedem it's a trait at best
22:34:14 mriedem but jay might not even like that
22:34:18 edleafe cfriesen: because that feels like having placement mirror nova's design. IMO it would be cleaner to set reserved=total
22:34:19 mriedem since it's not a capability
22:35:03 edleafe mriedem: although it would be fun watching Jay foam at the mouth
22:35:22 mriedem yes i was hoping dan pinging him would summon the beast
22:35:31 sean-k-mooney well the cappablity is its abiltiy to provide resouces.
22:35:49 sean-k-mooney there is proably a general usecase there but ya likely not a new field
22:36:09 mriedem https://assets.rbl.ms/14621189/980x.jpg
22:37:04 sean-k-mooney for a second i tought it was a crock and wasl like aw its a lab
22:37:41 cfriesen edleafe: I don't like modifying "reserved" when nova hasn't really reserved anything, and I'm concerned that the math makes no sense if we set "reserved" equal to total but have already-allocated resources as well

Earlier   Later