| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2017-07-24 | |||
| 19:23:58 | mriedem | we agree we need allocation requests and the host_state dict thingies i think | |
| 19:24:09 | mriedem | or maybe we don't agree | |
| 19:24:13 | dansmith | I think he's saying we need to pass the HostState mess, right mriedem ? | |
| 19:24:16 | mriedem | yes | |
| 19:24:25 | jaypipes | is even used.... | |
| 19:24:28 | mriedem | it is | |
| 19:24:29 | mriedem | in the claim code | |
| 19:24:32 | dansmith | oh it is | |
| 19:24:33 | dansmith | yeah | |
| 19:24:43 | mriedem | note that all of the docstrings say the limits are vcpus/ram/disk | |
| 19:24:45 | mriedem | not numa | |
| 19:24:48 | mriedem | so it's totally f'ing confusing | |
| 19:25:07 | mriedem | no desert eagle? | |
| 19:25:14 | mriedem | you want open casket? | |
| 19:25:15 | jaypipes | yeah, it is. :( | |
| 19:25:23 | dansmith | mriedem: it's what I have within reach | |
| 19:25:53 | jaypipes | dansmith, mriedem: k, so the returned value needs to be list of lists of (host_dict_with_limits_thing, alloc_request) | |
| 19:25:57 | jaypipes | edleafe: ^ | |
| 19:26:13 | dansmith | no | |
| 19:26:14 | dansmith | because | |
| 19:26:26 | dansmith | host_dict_with_limits is an unversioned structure and we're not adding a parameter with a LIST OF TUPLES OF THAT THING to one of our clean rpc calls | |
| 19:26:51 | mriedem | clean rpc calls? | |
| 19:27:03 | jaypipes | dansmith: don't we already pass the limits stuff down to build_instance()? | |
| 19:27:13 | mriedem | select_destinations today returns the unversioned host_dict_with_limits_thing | |
| 19:27:17 | dansmith | jaypipes: no we get it fromthe scheduler | |
| 19:27:25 | mriedem | jaypipes: the NUMATopologyFilter | |
| 19:27:52 | jaypipes | mriedem: no, I know that. I meant where does the limits get passed to instance_claim() on the compute host. | |
| 19:27:54 | mriedem | jaypipes: let me present exhibit Z98 https://review.openstack.org/#/c/484439/ | |
| 19:28:09 | mriedem | via the filter props | |
| 19:28:21 | mriedem | actually no, | |
| 19:28:25 | mriedem | build_and_run_instance has a limits kwarg | |
| 19:28:33 | mriedem | which is the limits from the scheduler | |
| 19:28:46 | mriedem | which as far as i know just contains the numa topology limits | |
| 19:28:47 | mriedem | EXCEPT | |
| 19:28:54 | mriedem | for out of tree scheduler drivers that stash crap in there too :) | |
| 19:28:58 | mriedem | *and filters | |
| 19:29:06 | dansmith | hmm | |
| 19:29:32 | dansmith | so maybe we talk to scheduler late enough in this process that we're only calling ugly rpc calls with limits as a param from here on out? | |
| 19:29:48 | mriedem | do we need a hangout? | |
| 19:29:55 | dansmith | I need a therapist. | |
| 19:30:07 | mriedem | laura's mom went to grad school for that i think, she'll talk your ear off | |
| 19:30:21 | mriedem | about vacuums that also polish wood floors | |
| 19:30:26 | mriedem | controlled via your iphone app | |
| 19:30:38 | jaypipes | I'm wondering if we can't just recreate the limits['numa_topology'] on the compute host.... | |
| 19:31:16 | mriedem | i guess i'm failing to see the issue with sending both back from select_destinations | |
| 19:31:25 | mriedem | we're already sending host_dict_with_limits_thing | |
| 19:31:35 | jaypipes | dansmith: ^ | |
| 19:31:37 | mriedem | we're just tacking allocation_request(s)? onto that | |
| 19:31:39 | mriedem | as a tuple | |
| 19:31:55 | mriedem | btw, are these allocation requests plural or singular? | |
| 19:32:05 | edleafe | mriedem: we're changing 2 things | |
| 19:32:05 | mriedem | one per host right? so singular | |
| 19:32:10 | jaypipes | mriedem: singular per host, yeah. | |
| 19:32:12 | edleafe | instead of a single host | |
| 19:32:13 | dansmith | yeah, like I said above, I think I was missing that we're already passing that grossness in the rpc calls downstream from where we get them | |
| 19:32:21 | edleafe | we're sending a list of hosts | |
| 19:32:34 | edleafe | and each of those has an associated allocation_candidate | |
| 19:32:35 | mriedem | edleafe: please provide context | |
| 19:32:42 | mriedem | "sending" from where to where? | |
| 19:32:51 | edleafe | to the cell conductor | |
| 19:32:57 | edleafe | from the super conductor | |
| 19:33:10 | dansmith | super conductor doesn't call cell conductor | |
| 19:33:19 | dansmith | super conductor calls the first compute node, which will call cell conductor on retry | |
| 19:33:35 | mriedem | edleafe: ok so here https://github.com/openstack/nova/blob/master/nova/conductor/manager.py#L1046 | |
| 19:33:44 | edleafe | I thought that was changing so that we supported alternates | |
| 19:34:17 | dansmith | no | |
| 19:34:35 | edleafe | So we're sending the compute node the big honking list of list of stuff? | |
| 19:34:42 | mriedem | api -> super conductor > scheduler > super conductor > compute > cell conductor (retry) > compute | |
| 19:34:52 | dansmith | right | |
| 19:34:54 | edleafe | And then it sends it to cell conductor on retry? | |
| 19:34:59 | mriedem | yes | |
| 19:35:05 | mriedem | the allocation requests are getting passed through | |
| 19:35:13 | mriedem | like barnacles | |
| 19:35:23 | mriedem | or a tube sock in mortier | |
| 19:35:26 | mriedem | *mortimer | |
| 19:35:27 | edleafe | more like kidney stones | |
| 19:35:45 | mriedem | 2nd question, | |
| 19:35:48 | melwitt | lol omg, mortimer | |
| 19:36:12 | mriedem | are we changing the build_and_run_instances compute rpc api to pass allocation requests, or shoving those into something else already being sent, like request spec or filter properties? | |
| 19:37:37 | dansmith | one of those | |
| 19:37:58 | mriedem | which would also impact the build_and_run_instance method in conductor rpc api | |
| 19:38:11 | mriedem | *build_instances | |
| 19:38:35 | dansmith | if we put it into something like reqspec, we won't be able to send them to older computes | |
| 19:38:41 | dansmith | which will break our upgrade process | |
| 19:38:46 | dansmith | because they'll kick the newer version back | |
| 19:39:00 | mriedem | personally i think it's cleaner as a new parameter on the rpc api | |
| 19:39:08 | dansmith | which I guess is the same for the new param approach | |
| 19:39:26 | dansmith | we just need to handle the case in the retry logic, | |
| 19:39:35 | dansmith | if we didn't get these new things, assume we can talk to the scheduler and do a reschedule | |
| 19:40:18 | mriedem | but if your cell conductor is blocked from up calls to the scheduler, how would that work? | |
| 19:40:31 | dansmith | if you have old computes, | |
| 19:40:39 | dansmith | then you don't have a multi-tier cellsv2 environment, | |
| 19:40:43 | dansmith | because we didn't support it before, | |
| 19:40:46 | openstackgerrit | Ed Leafe proposed openstack/nova master: Migrate Ironic Flavors https://review.openstack.org/484949 | |
| 19:40:46 | dansmith | thus it must be fine | |
| 19:41:11 | mriedem | idk, we don't really have anything doc'ed for this | |
| 19:41:21 | dansmith | eh? | |
| 19:41:24 | dansmith | it can't work right now | |
| 19:41:29 | dansmith | no need to doc it :) | |
| 19:41:40 | mriedem | what does old computes have to do with multi-tier cells v2? | |