Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-02-03
21:33:57 mriedem and southern charm
21:34:02 mriedem / hospitality
21:34:03 cburgess LOL
21:35:41 superdan melwitt: they use their nova-cell.conf file, per cell yeah
21:36:03 superdan melwitt: we bypass the lookups now, but that's what we were doing before and it worked other than that the timing of noticing the replicated statuses was then off
21:36:27 superdan melwitt: is this on the step by step thing?
21:36:28 melwitt superdan: okay, so at this stage we do want to tell them to go ahead and create a cell v2 per cell v1 cell
21:36:41 superdan melwitt: yeah and that's what we do in devstack
21:36:52 superdan otherwise the work to convert later will be messy
21:36:57 melwitt superdan: no but I'm asking to know if I should add it. sounds like I do need to. I had been thinking we were doing just a cheat of a single cell
21:37:14 melwitt yeah. okay
21:37:22 superdan melwitt: you mean from the api cell's database?
21:37:40 superdan there's no reason to do that since they'll all be wrong when they want to go use them
21:37:51 melwitt superdan: I guess so. but that wouldn't make sense because there's no compute hosts in there
21:38:11 melwitt I hadn't thought it through
21:38:17 superdan yeah
21:39:17 melwitt okay, I'll add a new section for cells v1 people, I think that would be most clear
21:39:28 superdan melwitt: okay
21:39:34 melwitt thanks
21:39:48 superdan melwitt: maybe in another patch so we have a better chance of getting this in while we noodle over the cellsv1 stuff?
21:40:18 melwitt superdan: okay, if you think that's best
21:40:32 superdan whatever you want
21:41:17 melwitt I'm not opinionated about it. I hadn't thought of doing a follow up
21:41:39 superdan I'd just like that bit to be in rc2 and I wouldn't want to miss that if we're messing around with cellsv1 stuff that's all
21:41:57 melwitt yeah
21:42:05 superdan the mainline bit I mean
21:42:14 melwitt right
21:46:08 melwitt superdan: I'm trying to think of how the scheduling works with where we're at now. we're not yet pulling compute hosts from all cells or anything for choosing a host ... and cells v1 doesn't care about that since it chooses a cell first. so I guess that's how I had been thinking things could be gotten away with by setting up only one cell for cells v1
21:46:59 melwitt but doing that would be like, arbitrarily picking a cell to convert over and then later you'd have to do the rest, so not a great idea anyway
21:47:02 superdan melwitt: yeah, but I think we're okay because we choose a cell first, then go to a nova-scheduler that only knows about the lower levels
21:47:12 superdan no, I don't think we'll be able to do that unfortunatel
21:47:17 superdan not from v1 to v2 anyway
21:47:21 mriedem hmm, if you have a cell mapping per cell v1 cell, and the scheduler was checking all cell mappings, then it could hit something in another cell v1 cell couldn't it?
21:47:32 superdan no
21:47:35 mriedem if that were the case, i guess we'd just do if CONF.cells.enabled
21:47:36 superdan because scheduler is at the cell level
21:47:52 superdan and there's a cells scheduler that hasn't been touched in ages that knows nothing of this stuff
21:48:10 mriedem "scheduler is at the cell level"
21:48:20 mriedem cells v1? - there are multiple levels, can you clarify?
21:48:23 superdan if cells.enabled is false in the api cell, then it will start rpc'ing to the global schedule
21:48:27 melwitt superdan: don't think we'll be able to what? didn't catch what part you were referring to
21:48:28 mriedem i thought the cells scheduler was at the top level cell in cells v1?
21:48:35 superdan mriedem: yeah, in cellsv1 the scheduler is at the lower levels
21:48:43 superdan mriedem: no there's a special scheduler up there, regular scheduler down below
21:48:50 mriedem right
21:49:03 superdan melwitt: won't be able to have a cell-by-cell conversion from v1 to v2
21:49:18 superdan melwitt: but it'll be more of a "turning stuff off" kind of procedure I think, if we get the mappings in place ahead of time
21:49:50 superdan these are not friday afternoon questions
21:50:01 superdan ask me shit like "is it beer o clock yet" or something
21:50:12 melwitt superdan: heh, sorry. I blame dtp for asking me to think about stuff by asking about cells v1 step-by-step
21:50:18 mriedem scheduler is global in cells v2 world right?
21:50:22 superdan mriedem: yeah
21:50:33 mriedem so computes in multiple cells v2 cells are reporting into placement,
21:50:41 mriedem scheduler asks placement for an RP, gets one, sends that to conductor,
21:50:45 mriedem which is mapped to a cell via hostmappings,
21:50:53 mriedem and then conductor creates the instance in that cell and builds it there
21:51:10 superdan with pure v2 yes, but with v1,
21:51:15 superdan that all happens inside a cellv1 cell
21:51:31 melwitt superdan: I was thinking it doesn't matter if they set up all their cells now (technically) and that they could do them later. but I think that would make it more confusing
21:51:41 mriedem what's preventing the low-level filter scheduler in cells v1 not picking an RP from another v1 cell?
21:52:05 superdan mriedem: because they don't have visibility of each other? they'll all need their own placement, just like they need their own api
21:52:24 superdan mriedem: otherwise we have to teach placement about cellsv1, which would be ungood I think
21:52:24 mriedem ok i thought placement was still global in cells v1
21:52:28 mriedem right
21:52:36 mriedem we don't want cells v1 specific garbage in placement
21:53:00 mriedem maybe we want a note of that somewhere so when i ask it 5 more times in 12 months, you can point me
21:53:03 superdan placement is kinda part of the regular scheduler, which is replicated per v1 cell, which means placement is too
21:53:40 superdan melwitt: are you doing a general "what to make of the v1 mess" section? you could put that ^ in there if so
21:54:57 melwitt superdan: I'm doing a step-by-step for cells v1 peeps. what's this note about, just telling why not to worry about scheduling working right?
21:55:09 mriedem melwitt: that placement is per cells v1 cell
21:55:12 mriedem not global in cells v1
21:55:13 superdan that ^
21:55:29 melwitt oh ... hm
21:55:38 mriedem in other words, placement only has knowledge of the computes in the cells v1 cell that it's running in
21:55:50 superdan probably needs to go into some placement doc
21:55:56 mriedem yeah, which we have
21:56:10 melwitt yeah, I dunno anything about how to deploy placement for cells v1
21:56:17 mriedem just like nova-api
21:56:25 mriedem except in the child cell
21:56:31 superdan mriedem: I'll add it to placement.rst
21:56:43 mriedem ok http://docs.openstack.org/developer/nova/placement.html
21:56:47 melwitt so more like nova-scheduler then?
21:56:48 mriedem not sure what the best place is
21:56:54 mriedem melwitt: yeah
21:57:02 superdan melwitt: right, goes with scheduler
21:57:03 melwitt k, that makes sense
21:57:03 mriedem don't we have nova-scheduler in the top cell too?
21:57:07 mriedem or is that the cells service?
21:57:12 mriedem with the cell scheduler
21:57:27 melwitt it's the cells scheduler, I dunno if it's the same code as nova-scheduler or not
21:57:28 mriedem cells-region or whatever
21:57:29 superdan hmm, you know
21:57:54 superdan because placement is all service-catalog-y, I wonder if people will be able to do that
21:59:29 openstackgerrit Dan Smith proposed openstack/nova master: Clarify the deployment of placement for cellsv1 users https://review.openstack.org/429056
21:59:55 superdan mriedem: no it's different I think
22:00:15 superdan nova/cells/scheduler.py
22:01:40 melwitt ah, cool
22:15:03 superdan mriedem: your rc1 todo etherpad says we have until monday to do rc1 but the announce already went out right?
22:15:19 mriedem yeah, i'll fix
22:15:25 superdan okay

Earlier   Later