Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-10-02
20:00:45 dansmith edleafe: the top level one is schedule_and_build_instances, the other is build_and_run or something like that
20:00:56 edleafe dansmith: thanks
20:01:07 edleafe rewriting the alternate hosts spec
20:01:21 edleafe totally spaced that the call to compute is a cast
20:01:59 mriedem hmm, if we deprecate personality files from the API, we would presumably also deprecate showing limits on personality files in the API too - and if the alternative for personality files is user_data and config drive, we don't have any quota limits on those - so would we need to add some?
20:02:36 mriedem edleafe: superconductor doesn't cast to cell conductor
20:03:38 edleafe mriedem: I was talking about the cell conductor -> compute. That's the cast
20:03:43 mriedem superconductor (schedule_and_build_instances) -> scheduler (select_destinations) -> superconductor (schedule_and_build_instances) -> compute (build_and_run_instance) -> cell conductor (retry = build_instances)
20:03:56 mriedem yeah then it's build_and_run_instance both ways
20:04:02 sean-k-mooney mriedem: well the user-data is stored perinsance correct so im not sure we need limits on it if its a fixed lenght string in the db
20:04:30 cdent ‘night all
20:04:32 mriedem sean-k-mooney: yeah, wasn't sure if that's why we had quota on injected files or what
20:04:36 mriedem i guess it's rate limiting
20:05:27 mriedem ooo we have 3 options for file injection quota
20:05:47 sean-k-mooney if its sored as a text field we may want a quota
20:06:13 sean-k-mooney injected files, injected file content bytes and ?
20:06:19 mriedem quota_injected_file_path_length
20:07:05 mriedem melwitt: AbsoluteResource just means it's global per deployment? or per project?
20:07:07 sean-k-mooney is that litrally just a path lenght ? e.g. a file system constrait
20:07:23 mriedem yes, default is 255 for the file path length
20:07:41 sean-k-mooney that the fat32 limit right
20:07:56 mriedem yeah i think so
20:11:16 mriedem god i've already forgotten how all of this quota checking code works already
20:11:43 sean-k-mooney looking at the instances table it looks like user_data mediumtext field which means its got a 16 Mib limit on mysql
20:12:11 sean-k-mooney *user_data is a mediumtext
20:14:39 mriedem ok it looks like this quota limit is per project/user
20:15:32 mriedem so by default you can have 5 files, at 10mb each
20:15:45 mriedem wait, no
20:15:46 mriedem kb
20:17:10 sean-k-mooney mriedem: yes it is. userdata however is handeled seperatly from that limit today
20:18:17 mriedem right, there is no quota limit on user_data
20:18:20 mriedem it's just the limit in the db per instance
20:18:25 mriedem which is as you said 16MB
20:18:30 sean-k-mooney yes
20:19:04 sean-k-mooney we could change that in the future or mysql could be we dont have a contract in place with the end user on this today
20:20:32 sean-k-mooney having a qouta on this in the future would not nessicalily be a bad thing it is sotre on you db node after all so it could cause issues if allowed to grow too large
20:35:05 dansmith mriedem: are you aware of any cells-based short circuiting around InstanceMapping?
20:35:17 dansmith in the api_samples_tests
20:35:46 dansmith in looking at some of them, they create a server, which exists in the cell db, but we never set the InstanceMapping.cell_mapping for those
20:35:59 dansmith ISTR these were problematic with cells for some reason, but I can't really remember why
20:37:34 mriedem not really
20:38:13 dansmith ah, maybe it's SingleCellSimple that's getting me
20:38:25 mriedem i was going to say, i thought there was a fixture we used, but don't see where it's set
20:38:28 dansmith yasss
20:39:38 dansmith yeah, the wrinkle is that the api sample tests use singlecellsimple unlike the other functional ones
20:39:53 mriedem oh i'm looking at newton still, derp
20:40:27 mriedem If1138331f3a46f5aed87e898ce19879a787d435f
20:40:51 dansmith I added a new CellMappingList method and so I have to put that in the fixture or we always get back an empty list
20:47:48 mriedem we said at the ptg that if we deprecated personality files from the api, and you can specify personality files during rebuild, that we'd allow passing new user_data during rebuild, but i'm not sure if we should
20:47:56 mriedem i think we just allow passing personality files during rebuild b/c we don't persist them
20:48:05 mriedem so that was a hack workaround for the lack of persistence,
20:48:15 mriedem so i'm not sure we should allow passing user_data to rebuild just b/c personality is gone
20:48:24 mriedem sdague: remember that discussion? ^
20:48:54 mriedem cfriesen: ^ you might care since you seem to love rebuild
20:49:39 sdague mriedem: you are right
20:50:52 openstackgerrit Dan Smith proposed openstack/nova master: Make get_instance_objects_sorted() be smart about cells https://review.openstack.org/509003
20:50:52 openstackgerrit Dan Smith proposed openstack/nova master: Add CellMapping.get_by_project_id() query method https://review.openstack.org/509002
20:51:29 sdague mriedem: though, I think people were using that as an end run
20:51:37 sdague we can't update user_data on instances, right?
20:51:48 dansmith mriedem: ^ that's the last cellsv2 perf optimization I had in my head to consider it legit usable for queens
20:52:04 mriedem sdague: no you can't specify user_data during update
20:52:20 mriedem i don't really know if people are using personality files on rebuild as a workaround for missing user_data
20:52:26 mriedem i don't know how people use our apis really
20:52:36 sean-k-mooney jaypipes: after nested resouce providers lands would you have any objection to seperating https://github.com/openstack/nova/blob/master/nova/objects/resource_provider.py into a placement-lib repo so it could be imported by neutron/os-vif for the ovo neutron port binding work and bandwith based scheduling?
20:53:02 efried os-placement ftw
20:53:18 mriedem dansmith: are you going to run https://review.openstack.org/#/c/509003/ through your performance tests, compared to before/after that change?
20:53:28 mriedem but on top of https://review.openstack.org/#/c/505418/
20:53:38 dansmith mriedem: I'd have to have lots of cells to make any difference
20:53:52 dansmith so I wasn't really planning to, unless you just want to make sure it's not super painful or something
20:53:58 jaypipes sean-k-mooney: yes, I would. those objects are not intended for use by anything outside placement. That said, I think the /nova/scheduler/client/report.py code can be moved into a placement lib
20:53:59 mriedem dansmith: couldn't you measure the relative difference by just having 1000 instances in cell1, all ACTIVE?
20:54:06 mriedem that would exclude cell0 at least
20:54:46 dansmith mriedem: you're trading one one-row api db result for one empty cell0 result
20:54:47 dansmith so I doubt you'd be able to measure it
20:55:17 dansmith if you really want, I could just synthesize a bunch of cells with two tenants or something
20:55:21 sean-k-mooney jaypipes: ok the usecase was to be able to define a placement request and a placement allocation object that could be used between nova and neutron
20:55:58 jaypipes sean-k-mooney: yeah, that's report.py, not resource_provider.py :)
20:56:06 sean-k-mooney jaypipes: i originally taught of having them defined in os-vif but taught that felt a litle weired as they could be used in cinder
20:56:32 jaypipes sean-k-mooney: note that objects in nova/objects/resource_provider.py do NOT get sent over the wire, ever.
20:56:42 jaypipes sean-k-mooney: and we intend to keep it that way :)
20:57:42 sean-k-mooney but they do map to the structure recived from the rest api?
20:58:11 sean-k-mooney currently neutron is manipulating raw json responces which is a little annoying
20:59:58 sean-k-mooney actully that code look kind of like what neutron is doing but i was hoping we could have a singel place for it that both nova and neutron used so they dont get out of sync
21:01:54 sean-k-mooney jaypipes: https://github.com/openstack/neutron/blob/b1dd13abfb121ca2a3e1cebc6bc2cef4a056f401/neutron/services/segments/placement_client.py is basically a stripped down version
21:05:23 jaypipes sean-k-mooney: right, and I'm agreeing with you that it would be good to have a placement-lib repo that would store common code for callers of the placement API :)
21:05:44 jaypipes sean-k-mooney: I'm just saying that that common code is the scheduler/client/report.py code and not the nova/objects/resource_provider.py code :)
21:07:21 sean-k-mooney i am glad we are in violent agreement :) ya that make senese this is the same comment code that would be imported in osc
21:08:49 sean-k-mooney does it make sense to look at createing placement-lib after nested resouce providers or are there other api level changes to plaement in queens that we should also wait for
21:10:13 sean-k-mooney perhaps limit? that should be non invasive to add to the client code however so im less worried about that.
21:11:20 sean-k-mooney jaypipes: i just dont want the creation of placement-lib or os-placement to impact your other work, we can temporaily create the functions we need on the neutron side for now
21:12:06 jaypipes sean-k-mooney: sorry, what do you mean by limit?
21:13:05 sean-k-mooney cdent's spec for limits on the allocation candidates
21:14:58 sean-k-mooney https://github.com/openstack/nova/blob/master/nova/scheduler/client/report.py#L325 would need a minor update to be able to pass the limit to the api.
21:21:05 jaypipes sean-k-mooney: oh, I see. well that would be a microversion that would be passed by the calling client.
21:21:32 jaypipes sean-k-mooney: so yeah, a minor change to the report.py client module and the similar code in mlavalle's code
21:22:50 melwitt mriedem: I think most of the quota limits (except fixed_ips, networks, floating_ips) can be set per project or per project and user. so quota_injected_file_path_length would be global to a deployment if it hadn't been set via the quota-sets update API
21:23:32 melwitt and AbsoluteResource just means that the resource usage isn't counted based on values stored in the database
21:24:30 melwitt the injected_file_path_length is evaluated on-the-fly when a request is made to verify that the requested injected file length is less than or equal to the quota limit
21:29:44 mriedem melwitt: so it's not global to the deployment right? just per-request
21:29:47 mriedem so api rate limiting

Earlier   Later