| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2017-12-01 | |||
| 14:55:35 | mriedem | if those instances were in a server group policy, we are likely destroying those | |
| 14:57:32 | giblet | without the late check we break the group policy, with the late check even nova doesn't reschedule we at least fail the evac instead of breaking the policy | |
| 14:57:43 | giblet | that is already better | |
| 14:58:05 | giblet | of course it would be nice to reschedule | |
| 14:59:32 | mriedem | https://bugs.launchpad.net/mos/+bug/1735407/comments/3 | |
| 14:59:34 | openstack | Launchpad bug 1735407 in Mirantis OpenStack 9.x "[Nova] Evacuation doesn't respect anti-affinity rules" [High,New] - Assigned to Oleksiy Molchanov (omolchanov) | |
| 14:59:37 | mriedem | my analysis | |
| 14:59:50 | mriedem | i don't really want to add a reschedule loop for evacuate | |
| 15:00:20 | mriedem | long-term we model affinity in placement and use that, but i'm not entirely sure that will fix this case either | |
| 15:00:31 | leakypipes | mriedem: ++ | |
| 15:00:48 | giblet | mriedem: +1 for your analysis | |
| 15:01:01 | giblet | mriedem: can we just add the post check without reschedule? | |
| 15:02:01 | mriedem | well i guess the affinity stuff in placement would mean we somehow 'claim' that the instance we're scheduling / evacuating is going to be on some target host, | |
| 15:02:10 | mriedem | so a concurrent request gets that information when asking placement where it should go | |
| 15:02:18 | mriedem | not sure how that is all going to work out | |
| 15:02:23 | mriedem | since we don't want to model server group policies in placement | |
| 15:05:30 | mriedem | giblet: https://bugs.launchpad.net/mos/+bug/1735407/comments/4 | |
| 15:05:31 | openstack | Launchpad bug 1735407 in Mirantis OpenStack 9.x "[Nova] Evacuation doesn't respect anti-affinity rules" [High,New] - Assigned to Oleksiy Molchanov (omolchanov) | |
| 15:05:49 | mriedem | giblet: doing the late check might be reasonable | |
| 15:05:56 | mriedem | fail the evacuate, so the user can retry | |
| 15:06:55 | giblet | I vote for option 1) now and maybe consider prioritizing affinity impl with Placement in the future | |
| 15:07:52 | giblet | mriedem: I have to run soon but I have already looked into adding such late check before so I can take this bug if it can wait until Monday | |
| 15:08:03 | mriedem | sure | |
| 15:08:25 | mriedem | assigned to you | |
| 15:08:45 | giblet | cool, thanks | |
| 15:09:51 | mriedem | superdan: here is belmiro's build request cellsv1 bug https://bugs.launchpad.net/nova/+bug/1735353 | |
| 15:09:52 | openstack | Launchpad bug 1735353 in OpenStack Compute (nova) "build_request not deleted when using cellsV1 and local nova_api DB" [Undecided,New] | |
| 15:16:43 | mriedem | cdent: figleaf: rest api error code question, | |
| 15:16:50 | mriedem | https://bugs.launchpad.net/nova/+bug/1735329 500 is surely not correct in this case, | |
| 15:16:51 | openstack | Launchpad bug 1735329 in OpenStack Compute (nova) "console-log return 500 if n-cpu is shutdown" [Undecided,New] - Assigned to hongbin (hongbin034) | |
| 15:16:54 | mriedem | i'm thinking 409? | |
| 15:17:38 | cdent | s | |
| 15:18:38 | fried_rice | mriedem cdent 503 "Service Unavailable" ? | |
| 15:19:09 | cdent | 503 doesn’t really mean that, we’ve tried to discourage it | |
| 15:19:22 | mriedem | was going to say, i've been told not to use 501, 503, etc | |
| 15:19:23 | cdent | mriedem: I reckon 409 is probably right | |
| 15:19:33 | mriedem | this api already returns 409 so that's a good fit | |
| 15:19:58 | mriedem | hongbin: ^ | |
| 15:20:08 | fried_rice | "conflict" seems weird for this, but whatevs. | |
| 15:20:11 | cdent | 503 is more for when a proxy is involved | |
| 15:20:12 | hongbin | o/ | |
| 15:20:30 | cdent | fried_rice: conflict is that the state of the resource is not aligned with what you need it to be | |
| 15:20:42 | fried_rice | That's a stretch. | |
| 15:20:49 | fried_rice | IMO | |
| 15:20:55 | fried_rice | But I'm no expert. | |
| 15:21:11 | cdent | that’s what conflict really means | |
| 15:21:34 | openstackgerrit | Dan Smith proposed openstack/nova master: Unmap compute nodes when deleting host mapping https://review.openstack.org/524641 | |
| 15:21:41 | mriedem | ooo | |
| 15:25:50 | mriedem | not sure how to find out our default http response timeout | |
| 15:26:03 | superdan | mriedem: ^ | |
| 15:26:12 | mriedem | superdan: i know, already -1ed :) | |
| 15:26:56 | mriedem | or is the http response timeout configurable in uwsgi | |
| 15:26:58 | mriedem | ? | |
| 15:27:21 | superdan | mriedem: https://github.com/openstack/nova/blob/master/nova/db/sqlalchemy/api.py#L707 | |
| 15:27:36 | mriedem | superdan: that's not what you want | |
| 15:27:39 | mriedem | that's the nodename | |
| 15:27:44 | mriedem | ComputeNode.host is the hostname | |
| 15:28:02 | mriedem | it's totally confusing | |
| 15:28:45 | superdan | um | |
| 15:30:08 | superdan | I thought get_all_by_host was by service hostname | |
| 15:30:17 | superdan | why else would be it a multi-return? | |
| 15:30:25 | superdan | I feel like I've made this mistake like eight times | |
| 15:30:59 | mriedem | let me see what the actual HostMapping.host field is set from | |
| 15:31:07 | openstackgerrit | Jay Pipes proposed openstack/nova master: placement: allow filter providers in tree https://review.openstack.org/377215 | |
| 15:31:07 | openstackgerrit | Jay Pipes proposed openstack/nova master: placement: adds REST API for nested providers https://review.openstack.org/384807 | |
| 15:31:08 | openstackgerrit | Jay Pipes proposed openstack/nova master: placement: update client to set parent provider https://review.openstack.org/385693 | |
| 15:31:20 | mriedem | host_mapping = HostMapping( | |
| 15:31:20 | mriedem | ctxt, host=compute.host, | |
| 15:31:20 | mriedem | cell_mapping=cm) | |
| 15:31:30 | mriedem | in _check_and_create_host_mappings | |
| 15:36:52 | superdan | oh hypervisor_hostnme is actually the node uuid for ironic isn't it | |
| 15:36:58 | superdan | gawd why can't we call things what they are | |
| 15:38:06 | cdent | superdan++ | |
| 15:41:14 | superdan | well, regardless, hostmapping uses cn.host | |
| 15:41:21 | superdan | so that's what I want anyway | |
| 15:43:44 | mriedem | superdan: ? so you're going to use get_by_hypervisor? | |
| 15:43:54 | superdan | no I mean the .host field | |
| 15:44:02 | superdan | I really thought .host was the service host | |
| 15:44:11 | superdan | because we have get_by_host_and_node | |
| 15:44:48 | superdan | I'm doing this all whilst on a call, which I really shouldn't do | |
| 15:44:49 | mriedem | what gets passed to delete_host is the computenode.host, so you have to use ComputeNodeList.get_all_by_host, yes? | |
| 15:44:52 | mriedem | ok | |
| 15:45:06 | mriedem | the only time this would return more than 1 node is ironic | |
| 15:45:26 | openstackgerrit | Merged openstack/nova stable/pike: Fix ValueError when loading old pci device record https://review.openstack.org/524157 | |
| 15:45:35 | superdan | that's what I mean, what we pass is what is in hostmapping.host, which is populated from computenode.host | |
| 15:46:16 | superdan | which I guess is the service hostname | |
| 15:46:55 | mriedem | yeah, it's what comes from CONF.host | |
| 15:47:12 | superdan | I really didn't think that's what was in hostmapping | |
| 15:53:35 | openstackgerrit | Dan Smith proposed openstack/nova master: Unmap compute nodes when deleting host mapping https://review.openstack.org/524641 | |
| 15:53:36 | superdan | mriedem: try that | |
| 15:57:14 | bauzas | superdan: mriedem: yeah, so the compute service hostname is the host, and the Ironic node UUID is the hypervisor_hostname | |
| 15:57:54 | bauzas | so, when we get all the nodes for a host, we should get all the ironic nodes related to the specific nova compute | |
| 15:58:07 | superdan | yeah, that's not really the confusion, | |
| 15:58:12 | superdan | it's what was in hostmapping.host for me, | |
| 15:58:18 | superdan | and I thought it was the node | |
| 15:58:25 | bauzas | ah no, unfortunatel | |
| 15:58:28 | bauzas | indeed | |
| 15:59:43 | bauzas | and yeah, the main reason why we don't modify the field to be "node" instead of "hypervisor_hostname" is because it would be a huge change | |
| 15:59:44 | mriedem | superdan: want to fix the alignment in the test? | |
| 16:00:04 | mriedem | putting comments next to the fields in the code would probably help | |