Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-10-04
14:00:33 dansmith metadata_items? isn't that actual storage in the db?
14:00:46 openstackgerrit Merged openstack/nova-specs master: Add image-precache-support spec https://review.opendev.org/686232
14:01:00 mriedem user metadata is just a foreign key pair in another table so i'm not sure why that would matter
14:01:09 mriedem i mean,
14:01:23 mriedem sure you don't want someone blasting 1K metadata keys per instance since that will likely slow queries down
14:01:27 dansmith because a billion metadata items cost actual space
14:01:33 dansmith seems different to me
14:01:34 dansmith anyway
14:01:41 dansmith we can haggle in the api patch
14:01:56 mriedem i know injected_files isn't in the db
14:01:59 dansmith at least a note in the ref about "if you specify a hundred things here, you'll likely timeout before we've asked glance about all of them"
14:02:24 dansmith no, but it is still resource-related
14:03:03 mriedem i probably wouldn't even mention anything in the ref unless/until someone reports it as a problem
14:03:14 mriedem so like i said, i'm not terribly worried about it
14:03:22 mriedem especially since it's admin-only by default
14:03:41 mriedem and yeah when you're admin listing 1K all_tenants instances we're going to ask neutron for the combined set of ports and security groups for all of them
14:03:53 dansmith I don't see what a note would hurt, given not everyone would piece together what is happening where, but... whatever, we can talk about it in the patch
14:04:26 mriedem i just have no idea what the limits would be, so wouldn't want to say "you'll likely timeout"
14:04:51 dansmith sure, it'd just be a warning about "too many"
14:05:10 dansmith whatever, I'll stop trying to address your concerns :)
14:05:12 mriedem "Since the number of images to cache is not rate limited use caution when sending a large set of images."
14:05:13 mriedem something like that
14:05:36 mriedem anyway
14:05:52 dansmith or I could push validation to the conductor
14:06:21 mriedem failing fast in the API is nice though...
14:06:23 mriedem for this kind of thing
14:06:30 dansmith so basically,
14:06:40 dansmith you will not be pleased with any outcome, is that right mr. midwest?
14:06:50 mriedem that's my general stance with anything in my life yes
14:06:56 dansmith right, got it :)
14:07:38 mriedem laura was all riled up about this impeachment stuff last night and i had to remind her to keep expectations very low and assume the worst in whatever outcome she wants - like i do!
14:07:40 mriedem it's very healthy
14:07:51 dansmith heh
14:09:12 mriedem so here is a thing to compare to in the api,
14:09:25 mriedem during server create i don't think we limit the number of bdms you can specify,
14:09:44 mriedem you can't include the same pre-existing volume more than once since we don't support multiattach that way,
14:10:04 mriedem but you could say, boot me 10 servers with 255 bdms each and they are all using the same image for the volume's disk,
14:10:23 mriedem the api will GET that image 2550 times at least
14:10:28 mriedem since we're too dumb to cache the results
14:11:08 mriedem that's in _validate_bdm called from a for loop in _provision_instances
14:13:06 dansmith yeah, you're saying we do that today and nobody is complaining?
14:13:23 dansmith obviously less likely to run afoul of that with 255 BDMs vs. images to pre-cache
14:13:29 dansmith but same deal
14:13:47 dansmith actually, that all happens in the conductor right?
14:13:49 dansmith used to be in the api
14:14:12 mriedem no it's in the api
14:14:22 mriedem and yeah no one is complaining about that right now
14:14:23 dansmith okay
14:14:46 mriedem i remember the power boyz really wanting a couple hundred volumes attached per server
14:14:58 mriedem so do that + multicreate
14:15:42 dansmith yeah
14:18:58 dansmith mriedem: well, we have the new imagecache group, maybe we just chuck a configurable limit in there.. no real reason not to I guess
14:20:10 mriedem config-driven api is the only reason not to...
14:20:21 dansmith yup
14:20:24 mriedem which is why i was thinking of quota config etc
14:26:00 openstackgerrit Matt Riedemann proposed openstack/nova master: Add boot from volume functional test with a huge request https://review.opendev.org/686734
14:32:34 openstackgerrit Eric Fried proposed openstack/nova-specs master: Add 'Core Liaison' spec process https://review.opendev.org/685857
14:33:27 efried updated to make dansmith his own liaison on his precaching spec, and took out the idioms gibi called me on ^
14:33:49 efried also clarified self-liaison wording can just be "None" as bauzas suggested ^
14:49:00 efried aspiers: update on SEV fix?
14:49:55 stephenfin mriedem: I'm reworking the NeutronFixture to mock out the security group-related APIs instead. It's going to be much larger but seems more correct
14:50:21 efried aspiers: I'm gonna say by... Tuesday we will merge the "document it's busted" patch if the fix isn't ready.
14:50:38 efried so that we can be sure it's through the gate by Thursday when we have to produce our final RC.
14:52:00 mriedem stephenfin: ugh....can we just do that in a follow up?
14:52:07 mriedem b/c that's going to slow things down quite a bit
14:52:43 stephenfin mriedem: We can. Want me to just fix that commit message comment and merge as-is so?
14:53:10 bauzas efried: honestly, I'm not opinionated, I just wanted to explain it would look a bit weirdo if the same name would be for both the owner and the liaison
14:53:28 bauzas efried: something either "None" or "Not needed" is cool with me
14:53:33 efried bauzas: I agree, I tried to explain it in the FAQ
14:53:40 mriedem stephenfin: if you're going to rev the commit message, i think you could also the thing i asked for in the stub which is only return a default security group for an instance that has a non-empty info_cache
14:53:44 mriedem stephenfin: i think that's as easy as:
14:53:48 efried bauzas: and yes, latest PS has "or None" in the template
14:53:57 bauzas ok, will look at it then
14:54:02 efried thx
14:54:16 mriedem if instance.get_network_info()
14:55:46 stephenfin I'm not sure if can rely on that since I think we mock a lot of it. See https://github.com/stephenfin/nova/commit/afb4698e8f7e39154e90dfe775d2d0e3e1f20dbc#diff-c731e6017d325067e73f0124f4ade91c
14:55:58 stephenfin Is that something else though?
14:56:13 openstackgerrit Matt Riedemann proposed openstack/nova master: Add boot from volume functional test with a huge request https://review.opendev.org/686734
14:56:18 mriedem dansmith: so ^ proves my point about the server create case, mayhap i should open a bug for that
14:56:28 dansmith mriedem: ack
15:00:50 openstackgerrit Matt Riedemann proposed openstack/nova master: Add boot from volume functional test with a huge request https://review.opendev.org/686734
15:02:54 mriedem stephenfin: ok so i guess for the current patch, how about just (1) fix the commit message and (2) leave a TODO in the new stub method that it should be smarter about instances that don't have networks/ports and thus won't have security groups
15:12:57 stephenfin mriedem: Yup, I can do that
15:13:00 stephenfin Will respin now
15:22:28 openstackgerrit Matt Riedemann proposed openstack/nova master: nova-net: Migrate 'test_quota_sets' functional tests https://review.opendev.org/684334
15:27:09 mriedem gibi: could i ask you to hit the bottom 2 FUPs in the cross-cell resize series https://review.opendev.org/#/c/676228/ since i think you're familiar with the changes they are...FUPing (aleady have a +2 on them)
15:27:58 gibi mriedem: sure, looking...
15:29:42 mriedem dansmith: what do you think about getting this CellTimeout logging traceback change in to debug the gate failure? https://review.opendev.org/#/c/684118/ - i'm unable to recreate on that change itself after many rechecks but it's one of the top gate failures: http://status.openstack.org/elastic-recheck/#1844929
15:30:59 dansmith mriedem: sure
15:33:09 gibi efried: so https://blueprints.launchpad.net/nova/+spec/support-move-ops-with-qos-ports is in Implemented state. So I think I need to open a new bp for U about the rest of the move operations (live migrate, evacuate, unshelve) and then propose a spec that is a slimmed down version of the original spec https://review.opendev.org/#/c/652608/ Do you agree?
15:33:25 mriedem i did get the mysqld logs published for these failing grenade jobs and it looks like we do start mysqld at least 3 times during the grenade run which i didn't expect
15:33:37 mriedem we can see things really start to go south about the time that we hit the cell timeout failures
15:33:38 mriedem https://zuul.opendev.org/t/openstack/build/4085120e390f4f1e971c6ff61304a596/log/logs/mysql/error.txt.gz#213
15:33:54 dansmith mriedem: yeah that seems weird
15:34:02 efried gibi: That would be fine. If you prefer I can un-implement the train bp and you can reuse it. I don't really care too much. But I think the old process hounds have a preference for creating a new bp (-ussuri suffix).
15:34:03 dansmith mriedem: are we actually upgrading mysql packages?
15:34:08 mriedem so i guess 2 starts would mean...old devstack and new devstack
15:34:11 openstackgerrit Stephen Finucane proposed openstack/nova master: tests: Correctly mock out security groups in NeutronFixture https://review.opendev.org/685927
15:34:15 stephenfin mriedem: ^
15:34:25 stephenfin think that's what you were looking for

Earlier   Later