Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-07-19
14:57:08 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Add Open vSwitch patch port https://review.openstack.org/485228
15:02:03 sdague sfinucan: confused on https://review.openstack.org/#/c/428241 - because I thought we were going the other direction
15:03:24 stephenfin sdague: Oh, I wasn't aware of that
15:03:26 stephenfin That changes things
15:03:46 sdague see efried's work there
15:03:54 stephenfin will do
15:12:17 mriedem bauzas: ok https://review.openstack.org/#/c/484828/
15:12:27 mriedem aller!
15:12:48 bauzas mriedem: so, I was thinking about accepting custom drivers if they use the placement API
15:13:09 bauzas mriedem: I mean, if all our drivers are given a list of hosts by the placement API
15:13:35 bauzas mriedem: the problem with the FilterScheduler driver is that it will call filters N times when you have N hosts
15:14:00 bauzas mriedem: so I think some operators would like to just call once a 3rd-party system
15:14:37 bauzas mriedem: if we do that, I think it would not be an interop problem given we'll be sure that any destination would be accepted by placement API
15:14:40 openstackgerrit Feodor Tersin proposed openstack/nova master: libvirt: Straighten resize condition in Image.cache https://review.openstack.org/485236
15:15:14 bauzas of course, like I said, it means that all the drivers would be calling placement for both getting hosts to verify, and would also claim the allocation too
15:15:31 bauzas thoughts about that?
15:15:35 mriedem bauzas: i'm not sure what that has to do with the bug fix at hand
15:15:47 mriedem which needs to go back to ocata where the custom scheduler driver loading was broken
15:15:58 bauzas mriedem: so, about that
15:16:01 bauzas mriedem: for the bug
15:16:14 bauzas mriedem: just adding the new choice would be okay I think
15:16:28 bauzas mriedem: but not saying it would be deprecated by Queens IMHO
15:16:32 sean-k-mooney bauzas: it would not be transparent on upgrade
15:16:49 bauzas sean-k-mooney: not sure I understand why
15:17:14 bauzas I wouldn't deprecate a choice
15:17:18 bauzas just adding a new one
15:17:24 sean-k-mooney bauzas: there was nothing that requried them to call the custom_driver custom_driver in the setup.cfg entrypoint before
15:17:30 bauzas the problem with upgrades is when you stop to use an opt
15:17:45 mriedem bauzas: adding choices is not ok
15:17:49 mriedem it's not backward compatible
15:17:55 mriedem when upgrading *to* ocata
15:18:24 mriedem we have to drop 'choices',
15:18:31 bauzas mriedem: not only then
15:18:42 mriedem and when loading up the entry point in the scheduler manager, if it's not in our whitelist of 'choices' that we support today, then we log a warning
15:19:02 bauzas mriedem: because if you upgrade from Newton, then you need to modify your nova.conf because the value wouldn't be the same
15:19:17 bauzas mriedem: in Newton, you needed to pass a python path
15:19:39 sean-k-mooney bauzas: you could not use the alis form stevador?
15:19:43 bauzas mriedem: in Ocata, you need to pass just a name for an existing entrypoint
15:21:17 sean-k-mooney bauzas: i guess the point still stands if we keep the choices field we would need a config change on upgrade and a change to the python module that exports the driver to export it as custom_driver
15:21:29 mriedem bauzas: i don't think that's right
15:21:55 mriedem in mitaka we deprecated the ability to classload and removed that in ocata
15:22:00 mriedem and replaced with stevedore entrypoint aliases
15:22:14 mriedem but still allows you to add your own stevedore entrypoint to load from
15:22:22 mriedem that was until we added that choices kwarg and broke everything
15:22:31 mriedem i can do some git history diving
15:23:02 mriedem https://github.com/openstack/nova/commit/fe3d6dba3d1db8f97dab4baffcc11dda56368096
15:23:17 sean-k-mooney stevedore also allows you to skip using the alais and use the class path it points to transparently so if the code was not moved the old config options would have worked
15:23:26 bauzas mmm
15:24:01 mriedem that merged on sept 22
15:24:14 mriedem https://review.openstack.org/#/c/349666/ introduced the choices kwarg which broke it
15:24:22 mriedem that merged on sept 30
15:24:30 bauzas mriedem: so, to make it clear, if you were newton
15:24:41 bauzas mriedem: and you didn't use the default value
15:24:56 bauzas mriedem: you were passing a python path for knowing which module to use, right?
15:25:36 mriedem not necessarily
15:25:42 mriedem you could have already been using a stevedore entry
15:25:53 mriedem https://review.openstack.org/#/c/339760/8/nova/scheduler/manager.py
15:26:01 TheJulia mriedem: apologies for the minor interrupt, but I and dtantsur replied to https://review.openstack.org/#/c/215385/ which hopefully will clear up the confusion.
15:26:08 mriedem ^ before that change, it would attempt to load from stevedore and failing that, try to load from classpath
15:27:41 bauzas mriedem: ah-ha, nevermind my comment because https://github.com/openstack/nova/blob/fe3d6dba3d1db8f97dab4baffcc11dda56368096/releasenotes/notes/sched_remove_classpath_import-5d0f48eb388e6948.yaml
15:28:01 bauzas mriedem: which is Newton
15:28:11 bauzas mriedem: so I was wrong and you were right
15:28:23 bauzas mriedem: okay, I'll then just remove the choices kwarg
15:28:45 bauzas mriedem: the big question being whether we want to deprecate the possibility to use a custom scheduler by queens
15:28:47 sdague bauzas / cdent - there a race with setting up placement/scheduler fixture it seems - http://logs.openstack.org/27/479027/1/gate/gate-nova-tox-functional-py35-ubuntu-xenial/0a995a0/console.html#_2017-07-19_14_33_46_324724
15:28:55 sdague which gives us novalid host on some functional tests
15:29:46 bauzas arf, I have a meeting to run
15:30:18 cdent sdague: there’s been quite a bit of kerfuffle with the compute api and the placement api fixtures not playing well together in the same tests. for a while there was some thought that it was eventlet related, but I’m not sure where that landed. mriedem may have more info?
15:30:55 mriedem bauzas: https://github.com/openstack/nova/blob/fe3d6dba3d1db8f97dab4baffcc11dda56368096/releasenotes/notes/sched_remove_classpath_import-5d0f48eb388e6948.yaml is not newton
15:31:01 mriedem https://review.openstack.org/#/c/339760/
15:31:07 mriedem Branches master, stable/ocata Tags 15.0.0, 15.0.0.0b1, 15.0.0.0b2, 15.0.0.0b3, 15.0.0.0rc1, 15.0.0.0rc2, 15.0.1, 15.0.2, 15.0.3, 15.0.4, 15.0.5, 15.0.6, 16.0.0.0b1, 16.0.0.0b2
15:31:12 sdague cdent: for things like that we could probably make super fake placement that says "thumbs up", given that we're running 1 compute
15:32:07 mriedem yes i've seen the timeout exception holding the lock in the compute update_available_resource
15:32:07 cdent yeah, that’s probably a good idea
15:32:09 bauzas mriedem: it's ocata, so anyway you have to change that
15:32:25 mriedem the PlacementFixture was turned on relatively global for functional tests recently
15:32:34 mriedem so it's running in a lot more functional tests, whether we need it to or not
15:32:57 mriedem sdague: semi related we also did this globally https://review.openstack.org/#/c/483972/
15:33:25 mriedem we didn't have a fingerprint for it, but it seems to have helped a little bit from anecdotal evidence, meaning me not having to recheck on weird wsgi errors as much
15:33:55 sdague mriedem: yeh, I think I'd stop doing placement globally, that seems like a lot of complexity when you don't need it
15:35:52 mriedem TheJulia: does a 'vendor driver' in ironic mean an out of tree driver?
15:36:34 mriedem if the limitation is on fibrechannel support, then why not just say, iscsi is currently the only supported volume type
15:37:45 ftersin mdbooth: hi. could you please review that (https://review.openstack.org/#/c/407440/) again? i addressed your comments and wait for new ones.
15:39:15 openstackgerrit John Haan proposed openstack/nova-specs master: Support volume_type with BDM paramter https://review.openstack.org/466595
15:40:46 TheJulia mriedem: anything produced by a vendor, so it could be in or out of tree, in the terms I wrote, I meant in-tree vendor drivers.
15:41:15 mriedem ok, not sure why 'vendor' is even a distinction then
15:41:19 TheJulia mriedem: I'd prefer to just drop the tag completely due to confusion at this point
15:41:19 mriedem what's a non-vendor driver in ironic?
15:41:36 mriedem the ironic coop driver :)
15:41:43 TheJulia mriedem: agent_ipmitool, pxe_ipmitool, or the ipmi, redfish hardware types
15:41:47 mriedem bm traded for hemp
15:41:50 mriedem ok
15:41:52 TheJulia lol
15:42:31 mriedem ok, so if the iscsi volume type is the only one that's currently supported, that's what i'd clarify in the support matrix and release note
15:42:37 jaypipes jangutter: 483921 approved.
15:42:50 TheJulia mriedem: those are the drivers we're able to test in the gate since we can emulate the BMC of baremetal hardware in those cases.
15:43:41 TheJulia mriedem: okay, I just want to make sure or at least hopefully make sure nobody freaks out when we ask to amend the support matrix later without putting code in nova :)
15:44:07 jangutter jaypipes, sean-k-mooney: the people in the office just looked at me really oddly. I don't think they expected that kind of sound emitting from human vocal cords.
15:44:17 TheJulia mriedem: that being, when we have that support in place in our reference drivers
15:44:21 jaypipes jangutter: :)

Earlier   Later