Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-03-06
15:56:01 mriedem it was fubar'ed FAST
15:56:04 dansmith stephenfin: I was thinking it would be nice to merge another good chunk of those this cycle, but...
15:56:43 sean-k-mooney mriedem for what its work im hoping to get the os-vif gates down to that by droping the keystone and cinder apis test.
15:57:11 mtreinish keystone tests take like zero time
15:57:18 mtreinish dropping cinder will save a lot of time
15:57:45 mtreinish sean-k-mooney: if there is no integration between os-vif and cinder I was just disable cinder all together in those jobs
15:57:51 sean-k-mooney mtreinish: ya but os-vif is a lib that never talk to any other openstack service
15:57:53 mtreinish s/was/would/g
15:57:53 mriedem that's why i dropped the non-compute api tests from the nova-next job
15:58:08 sean-k-mooney mtreinish: yes im going ot
15:59:28 sean-k-mooney i think in the os-vif jobs i might keep the compute and networking api tests + the senario test for now and see what the job times look like
16:00:05 sean-k-mooney if they are under an hour ill be happy if i can get down to 45 mins or so even more so
16:02:11 mtreinish well if you disable cinder (not just blacklist the cinder api tests) it won't run a good chunk of the scenarios
16:02:37 mtreinish and those scenarios are individually among the slowest tests in a job
16:11:31 stephenfin dansmith: Done-zo https://review.openstack.org/#/c/624592/8
16:17:03 melwitt mriedem: ack, will fix. I think it's bc that patch changed things from counting instances => instance mappings, that test is probably only creating instance records and not mapping records
16:23:03 dansmith mriedem: probably need someone to make a call on breaking potentially n-net-only tests with this: https://review.openstack.org/#/c/624592
16:24:27 mriedem i know some of the converted privsep stuff is going untested because testing the privsep stuff is hard...
16:24:36 mriedem i remember wresting with trying to mock around that stuff at one point
16:24:38 mriedem i don't remember the details
16:24:49 mriedem *wrestling even
16:25:10 mriedem i'd rather not make that call today...
16:26:02 dansmith just seemed like some feature-y patches we could make independent progress on, and they were just refreshed, but.. I got stopped at the first one
16:27:00 sean-k-mooney dansmith: even in the nova-net case some of that code is likely dead code as os-vif will be used in most cases
16:28:05 dansmith sean-k-mooney: understand, but it's just hard to get on board with orphaning code with tests.. if it's dead code, then maybe we should remove it
16:28:10 mriedem in other fun news, when post live migration fails we don't set the instance to error status or reset the task_state to None apparently http://logs.openstack.org/43/635343/4/check/tempest-slow-py3/a2497ae/job-output.txt.gz#_2019-03-06_05_29_44_910801
16:28:18 sean-k-mooney the gateway code is likely still used
16:30:12 sean-k-mooney mriedem: ya i have seen that happen a few times
16:31:34 sean-k-mooney dansmith: by the way we were waiting mainly for cern to move to cellsv2 only before droping nova-networks
16:31:46 sean-k-mooney so can we drop it entirly in train?
16:31:59 dansmith sean-k-mooney: they're not fully off it, AFAIK
16:32:09 sean-k-mooney we are not going to froward port it to python3 im assuming
16:32:35 sean-k-mooney dansmith: ok we should check with them and see.
16:33:28 stephenfin There was a mail about this on openstack-discuss a few weeks back
16:33:52 stephenfin I recall they said they were planning to migrate *everything* to neutron, but I can't recall the timeline
16:34:13 dansmith yeah, that's been the plan for a while, I just haven't heard that they have
16:34:25 dansmith either way, removal is non-trivial so cern being off it isn't the only barrier :)
16:34:50 stephenfin Sounds like a challenge
16:35:24 stephenfin Does anyone know if config drives rely on the metadata service being up? I'm assuming 'nova-compute' handles stuff by itself but haven't dived too deep yet
16:35:35 melwitt mriedem: I have a osc-placement change up for a batch command for setting allocation ratios per aggregate in placement, if you're interested in reviewing https://review.openstack.org/640898
16:35:36 sean-k-mooney well i think we might be able to delegate some of the logic to os-vif in the interim
16:36:01 sean-k-mooney for example i want to pull the trusted vf stuff into a sriov plugin in os-vif
16:36:12 dansmith sean-k-mooney: if you think those gateway things are still used can you comment?
16:36:42 dansmith sean-k-mooney: I think there's also some default firewalling of nova services that still use a lot of the linux_net ip.* commands, right? like punching holes for nova-api, metadata, etc
16:42:50 sean-k-mooney dansmith: ya there is a bunch of firwalling stuff. i would prefer not to migrate that code to os-vif but we migth be able to clean it up abit
16:43:20 dansmith sean-k-mooney: yeah, not saying we should, I'm saying we shouldn't just be randomly dropping tests for stuff we convert to privsep because "it's hard" and "it might be only for n-net"
16:44:05 openstackgerrit melanie witt proposed openstack/nova master: Use instance mappings to count server group members https://review.openstack.org/638324
16:44:05 sean-k-mooney ill take a look and see if its used and ya.
16:44:57 sean-k-mooney i can take a look at some to the privisep changes and see if i can convert them or help if needed too
16:45:18 sean-k-mooney but that will be after ff
16:45:45 sean-k-mooney i am happy to work with stephen on paying donw that technical debt
16:46:50 dansmith I was just thinking that these feature-y patches might be a few easy wins to get in before FF, several were +2 already and just refreshed
16:46:55 dansmith but clearly I was wrong :)
16:47:49 sean-k-mooney actully they are more or less indepenate so maybe we can change the odering of some of the patches that dont drop tests/nova-net stuff
16:48:15 sean-k-mooney e.g. merge the stuff that is a driect port/obviously correct and deffer the rest
16:49:04 dansmith yep, I +W'd one above that first one
16:50:25 artom stephenfin, the config drive is independant of the metadata API
16:50:46 stephenfin artom: So put configuration in nova-compute?
16:50:59 stephenfin I'm thinking on the vendordata stuff, specifically
16:51:15 dansmith stephenfin: what does that mean?
16:51:29 artom Yeah, what do you want to put in nova-compute?
16:51:30 dansmith stephenfin: config for how to reach the vendordata server/
16:51:57 stephenfin dansmith: If I was using DynamicJSON, yeah
16:52:00 dansmith stephenfin: config_drive literally just imports the python module that generates the metadata response and uses it to dump things into a disk image
16:52:08 dansmith stephenfin: then yeah, nova-compute would need that config
16:52:16 sean-k-mooney artom: you can add force_config drive or something but you can also just enable config drive in the flavor/image
16:52:22 stephenfin dansmith++ ta
16:52:54 stephenfin I ended up down a rabbithole yesterday and ended up doing this https://review.openstack.org/#/c/640730/ and now I'm more confused than ever. Back to reviews
16:56:13 mriedem https://bugs.launchpad.net/nova/+bug/1818873
16:56:14 openstack Launchpad bug 1818873 in OpenStack Compute (nova) "When post_live_migration_at_destination fails the instance is not put into ERROR/None vm_state/task_state" [Medium,Triaged]
16:56:26 mriedem so if post_live_migration_at_destination fails early, the guest is on the dest but the instance db state is all f'ed up
16:56:46 mriedem mdbooth: don't you have a similar bug to this ^?
16:57:41 sean-k-mooney mriedem: yes
16:58:07 mdbooth The one I recall was the other way round. The issue is when post_live_migration fails on the *source*
16:58:18 artom mriedem, maybe https://bugs.launchpad.net/nova/+bug/1628606?
16:58:20 openstack Launchpad bug 1628606 in OpenStack Compute (nova) "live migration does not clean up at target node if a failure occurs during post migration" [Low,In progress] - Assigned to Artom Lifshitz (notartom)
16:58:25 sean-k-mooney i brought that up in https://review.openstack.org/#/c/603844/
16:59:00 artom mriedem, I've been meaning to refactor the flow to make sure we save the instance state correctly in all cases
16:59:16 artom I addressed the Cinder API fail case, there are more (like Neutron)
16:59:29 sean-k-mooney artom: didnt you partial fix this already. we just 1 addional try finally block effectily to make sure we always set the instance host
16:59:36 openstackgerrit Merged openstack/nova master: Support server create with ports having resource request https://review.openstack.org/636360
16:59:41 sean-k-mooney artom: you didnt address all the cinder issues
16:59:44 openstackgerrit Merged openstack/nova master: objects: Store InstancePCIRequest.numa_policy in DB https://review.openstack.org/624444
16:59:48 artom sean-k-mooney, yeah https://review.openstack.org/#/c/611093/
17:00:28 sean-k-mooney you missed on place where we call cinder that can raise an exception
17:00:50 artom sean-k-mooney, yeah, it's entirely possible
17:01:15 artom sean-k-mooney, it was meant as a bandaid until a more correct fix is done
17:02:29 mriedem this is clearly not...enterprise ready
17:02:38 mriedem one could spend days cleaning up this mess
17:03:10 sean-k-mooney artom: ya i know
17:03:12 sean-k-mooney https://github.com/openstack/nova/blob/38c96993fd7163afe1a7c5f380af0d57a1bd2903/nova/compute/manager.py#L6736
17:03:16 mdbooth It seems more intuitive to me to update instance.host to the destination *immediately* we're actually running there, i.e. from the source at the top of _post_live_migration, but IIRC there are weird gotchas with that.
17:03:22 sean-k-mooney that is that call im thinking of
17:03:24 mriedem another clunker is we pass the Migration object around via the MigrateData which is pretty brittle,
17:03:29 artom mriedem, what mdbooth said
17:03:30 mriedem i'd like to just partial that like what mdbooth did with source_bdms
17:04:01 sean-k-mooney it will hit the cinder attachments api and can raise an exception which means the vm is running on the dest but we never update the db
17:04:06 artom I think just making sure we always update instance state the moment it's irreversibly on the destination is a good first step
17:04:49 sean-k-mooney ya i wrote it down on my whiteboard after reviewing the other chage earlier in the week
17:06:41 mdbooth You know how sometimes you don't understand a thing, but you remember understanding a thing while somebody was explaining it? I vaguely recall dansmith explaining why instance.host needs to be updated later.

Earlier   Later