Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-08-03
18:46:00 mriedem not to make your contribution to the change seem small or otherwise unimportant, of course :)
18:47:57 dansmith jaypipes: just unit tests I need to look at right? functional all work?
18:48:24 jaypipes dansmith: nope, all units should work fine. it's one functional test in tests/functional/compute/test_resource_tracker.py.
18:48:32 dansmith oh okay
18:48:55 jaypipes dansmith: it's expecting instance_claim() to have written allocation records for an ironic instance but it's not writing those allocations for some reason.
18:49:06 dansmith okay
18:49:14 jaypipes dansmith: and if I had to guess, maybe something to do with how we're pulling resources now with extra specs?
18:49:36 dansmith yeah found it
18:49:40 dansmith the test I mean
18:50:36 prashkre mriedem: Hi. Have you created a LP bug for issue with patch in https://review.openstack.org/#/c/391113/?
18:52:12 mriedem prashkre: i did but invalidated it once i understood how the policy rules were being linked
18:54:28 openstackgerrit Matt Riedemann proposed openstack/nova master: Add track_instance_changes note in disable_group_policy_check_upcall https://review.openstack.org/490627
18:56:31 prashkre mriedem: I got confused with your query, now I got it why we had made it like that. due to compatablitiy with people who already started using that rule, it was made like that..
18:58:28 mriedem melwitt: L35 in here about cell caches https://etherpad.openstack.org/p/nova-pike-cells-v2-todos - i think we have that covered in the FAQs page
18:58:35 mriedem but let me know if you were thinking of something else
19:08:07 openstackgerrit Merged openstack/nova master: hardware offload support for openvswitch https://review.openstack.org/398265
19:10:30 openstackgerrit Ed Leafe proposed openstack/nova master: Handle addition of new nodes/instances in ironic flavor migration https://review.openstack.org/487954
19:15:53 mriedem dansmith: ok i traced this ServerGroupAntiAffinityFilter test which is doing min_count=2 so multi-create, and the filter scheduler correctly puts the instances on different hosts
19:16:19 mriedem because of this https://github.com/openstack/nova/blob/master/nova/scheduler/filter_scheduler.py#L229
19:16:29 dansmith mriedem: because one scheduling request for N instances maintains the state, but two parallel requests do not, right?
19:16:33 mriedem the first host chosen for the first instance is put in the request spec hosts list
19:16:46 mriedem yeah this is single request,
19:16:51 mriedem parallel would require the upcall safety check
19:16:56 dansmith right
19:16:59 mriedem single request, 2 instances, seems ok
19:17:02 dansmith I thought that's what I was tryin gto say
19:17:26 mriedem ah ok
19:17:49 mriedem so if you care about affinity/anti-affinity, the user has to do multi-create or stagger their create requests
19:18:03 mriedem i'd think most are doing multi-create in this kind of use case anyway...
19:18:08 dansmith well, if you are about it, presumably you want it to actually work, not just work in some circumstances
19:18:15 dansmith I dunno, maybe
19:18:36 dansmith if you're booting two nodes to be mirrors, then maybe
19:18:39 mriedem yeah if you don't care that much, you create the server group using the soft affinity/anti-affinity policy
19:19:22 mriedem if i'm creating 2 db VMs and want them on different hosts for protection, then i want it to work
19:19:59 dansmith right, which might be a num_instances=2 thing
19:20:08 dansmith but if you're bursting,
19:20:09 cfriesen mriedem: what about heat?
19:20:25 dansmith two things might be adding more webserver processes in parallel in response to load
19:20:26 cfriesen mriedem: we've had a bunch of issues due to heat issuing requests in quick succession
19:20:30 dansmith and those wouldn't be the single request
19:20:38 mriedem dansmith: yeah
19:21:14 mriedem sucks for the end user since the way this is going to behave is not up to them, it's dependent on the configuration of the cloude,
19:21:15 mriedem *cloud,
19:21:20 mriedem but in the beforetimes,
19:21:23 mriedem this could fail due to load anyway
19:21:26 mriedem or dropped messages
19:22:14 dansmith well, the late check will fail and error the instance if it can't verify
19:22:39 cfriesen dansmith: weren't we dropping the late check with cellsv2? or did that get added back?
19:22:50 dansmith cfriesen: that's what we're discussing
19:23:02 mriedem sweet my test passed http://logs.openstack.org/54/489754/6/experimental/gate-tempest-dsvm-neutron-multinode-full-ubuntu-xenial-nv/a4fbbb3/console.html#_2017-08-03_18_04_43_399430
19:23:29 cfriesen dansmith: for what it's worth, we're missing an equivalent late check in the live-migration path. we found a race there.
19:23:59 mriedem we need an equivalent to dan's "HA and Pray" slogan for live migration
19:24:40 dansmith live migration is always prayer-powered
19:25:39 dfisher i shouldn't have laughed that hard
19:28:12 cfriesen seems like to close these sched races we need something like "make the choice, store the choice in the DB, verify the choice was valid and retry if not".
19:28:16 dansmith jaypipes: you're gone right?
19:28:27 dansmith cfriesen: that's placement
19:32:45 dansmith mriedem: I think I've fixed jaypipes' thing
19:32:50 openstackgerrit Dan Smith proposed openstack/nova master: remove provider allocs in confirm/revert resize https://review.openstack.org/488510
19:33:53 cfriesen dansmith: so are we leaving races in there (due to parallel schedulers for example)? As I understand it we originally needed the re-verification because we didn't store the scheduler decision to the DB until after the request hit the compute node.
19:34:05 mriedem sdague: these old images actually seem pretty useful https://review.openstack.org/#/c/478470/
19:34:10 mriedem i wonder why they were no longer used
19:34:42 dansmith cfriesen: when we track affinity in placement we won't need any of this because we'll be making the writing to the db as part of making the decison
19:34:43 mriedem although the might be inaccurate now
19:35:22 cfriesen dansmith: ah, okay.
19:36:16 dansmith gdi there's some gabbit failure now
19:36:20 dansmith which I suck at
19:38:40 sdague mriedem yeh, sfinucan and I went back and forth on that patch. I pushed for wanting to keep them, but they aren't used today, and it is more important that we get things merged
19:39:32 mriedem sdague: we also have https://docs.openstack.org/nova/latest/reference/vm-states.html
19:39:38 mriedem which is probably due for an audit also
19:39:50 sdague mriedem: many things are due for an audit
19:39:51 mriedem i'm trying to find the ironic doc that has all of their state machinery
19:40:14 sdague while I was fixing up the all misrendered block quotes I got a survery through what's there
19:40:33 mriedem why don't we have a search toolbar with the new docs theme?
19:40:34 sdague we need a pretty concrete discussion on information architecture at the PTG
19:40:41 sdague mriedem: I don't know
19:41:03 mriedem oh i found it
19:41:04 mriedem way at the bottom
19:42:37 sdague that's not search in our docs though, what is that?
19:43:52 openstackgerrit Dan Smith proposed openstack/nova master: remove provider allocs in confirm/revert resize https://review.openstack.org/488510
19:43:53 openstackgerrit Dan Smith proposed openstack/nova master: Add resource utilities to scheduler utils https://review.openstack.org/490514
19:55:51 mriedem efried: what am i supposed to call powervm again? so as to not confuse it with powervm support via libvirt (phyp?)
19:56:33 efried mriedem PowerVM. That other thing doesn't exist. thorst_afk ?
19:57:09 mriedem PowerVMv2: the re-powering
19:58:35 efried mriedem Is there some context in which that other thing is actually present an causing confusion?
19:58:40 efried and*
20:00:06 thorst_afk It was my understanding that the libvirt code for powervm is not fully functional (I'll have to investigate a bit). I know there was some work on it several years ago, but I don't know that anyone has been maintaining or updating it for years.
20:00:34 mriedem no i was just going to add it to the list here https://review.openstack.org/#/c/480074/7
20:00:40 mriedem and wasn't sure if there was a more specific name
20:00:44 mriedem that wasn't NEO or something
20:00:59 mriedem but we use PowerVM in the feature support matrix, so that makes sense to use here also
20:01:15 thorst_afk PowerVM is what we want to go with
20:01:42 efried In the words of Michael Bolton, "Why should I change? He's the one who sucks."
20:05:55 mriedem great, now i'm thinking about all of michael bolton's classic smokey throated wonder hits
20:06:09 dansmith mriedem: powervm != powerkvm if that's what you mean
20:06:23 mriedem dansmith: heh, i know
20:06:29 mriedem powervm != powervc
20:06:29 dansmith I don't think powervm via libvirt works from nova
20:06:35 openstackgerrit Matt Riedemann proposed openstack/nova master: doc: Start using oslo_config.sphinxext https://review.openstack.org/482961
20:06:35 openstackgerrit Matt Riedemann proposed openstack/nova master: doc: Import installation guide https://review.openstack.org/477488

Earlier   Later