Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-06-27
13:21:16 mriedem artom: was it intentional to not restrict block device tags to be unique for multiple bdms on the same instance?
13:21:21 mriedem we have no unique constraint for that
13:23:17 artom mriedem, yeah
13:23:40 artom mriedem, I don't see a use case for it either, but there might be a reason for it
13:23:59 artom For instance, two disks that have the same purpose get the same tag
13:24:03 edleafe jaypipes: I'll have to look more at that part. I was focusing on the changes to HostState
13:24:26 edleafe jaypipes: I am concerned how to select a host without first making a successful claim
13:24:38 artom mriedem, there was no real argument for or against, so by default it's allowed
13:24:50 artom Same for nics
13:24:55 artom And there it makes even less sense
13:25:07 artom Unless you're going to bond two nics and want them tagged the same? I dunno
13:25:12 jaypipes edleafe: we only make the claim after the scheduler filters and weighers have all run against the host states
13:26:06 mriedem artom: pl
13:26:08 mriedem *ok
13:26:19 bhagyashri_s johnthetubaguy: Hi, waiting for some inputs.
13:26:54 artom mriedem, plus I think semantically most people wouldn't expect a tag to be unique
13:27:02 artom When you think of tag clouds and such
13:27:16 artom You're not surprised when many elements share one or more tags
13:33:53 alex_xu artom: I guess you forget to set microversion on req obj https://review.openstack.org/#/c/376021/72/nova/tests/unit/api/openstack/compute/test_attach_interfaces.py@480
13:34:46 mriedem Vek: do we expect DeprecatedAction to always issue a warning even if you don't specify the arg? https://review.openstack.org/#/c/477212/
13:35:38 artom alex_xu, did I?
13:36:01 alex_xu artom: yea, those test should against on the 2.49
13:36:02 artom alex_xu, I ran those tests locally and they all passed... because 'tag' isn't a valid parameter in anything < 2.49
13:36:06 artom Doh
13:36:26 alex_xu artom: https://review.openstack.org/#/c/376021/72/nova/tests/unit/api/openstack/compute/test_attach_interfaces.py@516 this should use self.req
13:36:30 artom alex_xu, can you leave a comment and I'll fix
13:36:37 alex_xu artom: ok
13:37:20 artom alex_xu, thanks :)
13:37:47 edleafe jaypipes: I read ahead to the other PSs in the series to understand. It appears that select_destinations will return num_instances hosts to the manager - without claiming first. Right?
13:37:54 edleafe jaypipes: That means that there aren't any backup choices to select. The big gain we were hoping to get is when there is a race and the claim fails, the scheduler can simply select the next host in the weighed list. Having the manager make the claim means we lose that.
13:38:10 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Use ``assert_has_calls`` to check function calls https://review.openstack.org/477469
13:39:29 alex_xu artom: done
13:45:32 mriedem artom: couple easy things to do in https://review.openstack.org/#/c/435082/
13:45:34 mriedem and then i'm +2
13:45:44 artom mriedem, sweet, thanks :)
13:46:14 artom alex_xu, I don't think we can remove those mocks though
13:46:47 artom Wait
13:46:59 artom Dammit, that's twice now I say a thing and change my mind immediately after
13:47:03 artom Sorry :(
13:47:51 alex_xu artom: no worries :)
13:59:13 openstackgerrit Chason Chan proposed openstack/nova master: switch to openstackdocstheme https://review.openstack.org/477751
14:01:38 openstackgerrit Rodolfo Alonso Hernandez proposed openstack/os-vif master: Add abstract OVSDB API https://review.openstack.org/476612
14:04:56 artom mriedem, in the interest of reducing latency, is https://review.openstack.org/#/c/435082/50/nova/tests/unit/compute/test_compute_mgr.py@2537 satisfactory to you, or do you want to remove the tag and the call assertion entirely?
14:15:25 mriedem artom: replied
14:15:32 mriedem but now i'm shutting it down
14:16:15 artom mriedem, posthumous thanks :)
14:20:53 jaypipes edleafe: I thought we needed to return the alternate selections to the conductor anyway?
14:26:29 edleafe jaypipes: we need to pick a good host, so that we know what cell the alternates must be in
14:27:20 jaypipes edleafe: right, and those alternatives can be used to perform claim retries in the manager, no?
14:28:43 edleafe jaypipes: sure, but now you've kind of defeated the whole purpose of alternatives. If the setting is for 3 retires, and you burn 2 of them before you even get to the build...
14:30:05 jaypipes edleafe: it sounds like you're pretty gung-ho to just have a giant _schedule() method just do everything.
14:30:21 jaypipes I'm trying to add some separation of concerns into the codebase.
14:31:03 edleafe jaypipes: No, I'm "gung-ho" about having the _schedule() method return a good host
14:31:41 jaypipes edleafe: my patch series doesn't change _schedule()./
14:31:54 edleafe jaypipes: your separation is functionally the same as claims in the conductor.
14:32:05 jaypipes edleafe: what?
14:32:31 edleafe jaypipes: you're making a claim after the HostState objects are gone
14:33:11 jaypipes edleafe: what do you mean by "after the HostState objects are gone"?
14:33:48 edleafe jaypipes: in _schedule() we have a list of filtered, weighed HostState objects.
14:34:05 edleafe When you return they are no longer accessible
14:34:27 jaypipes edleafe: what do we need the host state objects for after the scheduler driver has filtered and weighed them?
14:34:38 edleafe So whether you're in the scheduler manager or the conductor manager, you're limited to the objects returned from _schedule()
14:34:39 jaypipes edleafe: we just need the resource provider UUIDs.
14:34:58 jaypipes edleafe: exactly! have the scheduler driver do one thing and do it well.
14:34:59 edleafe jaypipes: in a race-free world, sure
14:35:26 edleafe jaypipes: we had this discussion weeks ago
14:35:29 jaypipes edleafe: alternately, let's change the select_destinations() *driver* method to return the HostState list instead of a list of dicts.
14:35:40 openstackgerrit Artom Lifshitz proposed openstack/nova master: Delete disk metadata when detaching volume https://review.openstack.org/435082
14:35:41 openstackgerrit Artom Lifshitz proposed openstack/nova master: API support for tagged device attachment https://review.openstack.org/376021
14:35:56 edleafe dansmith was arguing for the conductor doing claiming, and you and I were arguing for the scheduler doing the claiming, using the exact thing I'm talking about now
14:36:25 jaypipes edleafe: I'm not doing claiming in the conductor. I'm simply doing it in the scheduler manager not the scheduler driver.
14:36:35 edleafe jaypipes: yes, I know that
14:36:52 edleafe jaypipes: which is why I wrote "functionally" equivalent
14:36:58 jaypipes but it isn't.
14:37:10 edleafe jaypipes: because you no longer have access to the weighed host list
14:37:27 jaypipes edleafe: so have the select_destinations() return a weighed host list, then.
14:37:35 edleafe jaypipes: and if you want to return the full list, well, there goes your separation goal
14:37:36 dansmith edleafe: claiming in scheduler manager is functionally equivalent to claiming in conductor?
14:37:56 edleafe dansmith: as far as having access to the weighed list of hosts, yes
14:38:09 dansmith o.O
14:38:11 jaypipes edleafe: if we return the weighted list to the manager, how does that mean "there goes my separation goal"?
14:38:58 edleafe jaypipes: because HostState is a driver concept
14:39:07 edleafe jaypipes: they don't exist in the manager
14:39:24 jaypipes I'm not saying return the list of host state objects. I'm saying return a weighted list of hosts.
14:39:45 jaypipes in other words, what my patch is doing -- minus the returning alternatives part.
14:40:24 dansmith even still, HostState is not a driver concept, IMHO
14:40:28 dansmith it's defined in host_manager even
14:40:32 edleafe jaypipes: http://p.anticdent.org/3900
14:40:46 dansmith I see no reason to keep that hidden from scheduler manager if it makes things more convenient
14:40:51 jaypipes or even if we did return a list of host state objects, I actually wouldn't care, as long as the manager isn't modifying those objects.
14:41:09 jaypipes or even if we did return a list of host state objects, I actually wouldn't care, as long as the manager isn't modifying those objects.
14:41:20 jaypipes edleafe: ^
14:43:36 edleafe jaypipes: one of your stated concerns was the chance scheduler. Chance scheduler doesn't use HostState. It's specific to filter scheduler
14:44:05 edleafe jaypipes: so I'm not sure how this is separating things more cleanly
14:44:08 jaypipes edleafe: right, which is why I changed the returned value from select_destinations() to be a list of dicts that included the UUID.
14:45:05 edleafe 09:40 < jaypipes>| or even if we did return a list of host state objects, I actually wouldn't care, as long as the manager isn't modifying those objects.
14:45:10 jaypipes edleafe: because the manager will loop through the allocation requests and attempt to claim resources on the destinations returned by select_destinations(). If select_destinations() doesn't return something with that UUID, the manager doesn't have the ability to match an allocation request to a host.
14:46:19 edleafe jaypipes: well, if keeping claiming out of the filter scheduler is worth it, then go for it
14:46:21 jaypipes edleafe: the alternative is to have the _schedule() method of the filter scheduler driver do everything under the sun: filter, weigh, get allocation requests, and claim.
14:47:10 edleafe jaypipes: and IYO it's *so* much cleaner to have the _schedule method do filter, weigh, and get allocation requests

Earlier   Later