Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2019-11-26
18:14:42 openstackgerrit Lee Yarwood proposed openstack/nova stable/train: compute: Take an instance.uuid lock when rebooting https://review.opendev.org/696151
18:15:52 openstackgerrit Lee Yarwood proposed openstack/nova stable/stein: compute: Take an instance.uuid lock when rebooting https://review.opendev.org/696152
18:16:18 openstackgerrit Lee Yarwood proposed openstack/nova stable/rocky: compute: Take an instance.uuid lock when rebooting https://review.opendev.org/696153
18:16:47 openstackgerrit Lee Yarwood proposed openstack/nova stable/queens: compute: Take an instance.uuid lock when rebooting https://review.opendev.org/696154
18:35:51 efried dansmith: I'm confused by some of the new wording, which might just be me, but pretty sure the test has a mistake https://review.opendev.org/#/c/695985/
18:37:36 dansmith efried: ah, frack, that was supposed to get dedented
18:37:43 efried :)
18:37:54 efried dansmith: but please add the self.fail() as well
18:38:18 dansmith efried: you mean a self.fail after the early exit to make sure we don't get there right?
18:38:23 efried correct
18:38:26 dansmith sure
18:38:32 efried self.fail('this code should have been unreachable!')
18:39:12 dansmith self.fail('I could totally take eric in a a street fight')
18:39:20 dansmith we'll never run that code, so we'll never know
18:40:31 efried hahaha
18:41:24 openstackgerrit Dan Smith proposed openstack/nova master: Add a way to exit early from a wait_for_instance_event() https://review.opendev.org/695985
18:51:08 gmann johnthetubaguy: these test change is due to base class assert on error message. these tests will be changed only where granularity is added which should not be many or we can remove the error message checks . - https://review.opendev.org/#/c/648480/20/nova/tests/unit/policies/test_services.py@63
18:54:44 openstackgerrit Merged openstack/nova stable/train: Replace time.sleep(10) with service forced_down in tests https://review.opendev.org/696088
19:05:42 openstackgerrit Ghanshyam Mann proposed openstack/nova master: Add new default rules and mapping in policy base class https://review.opendev.org/645452
19:05:58 openstackgerrit Ghanshyam Mann proposed openstack/nova master: Suppress policy deprecated warnings in tests https://review.opendev.org/676670
19:07:49 openstackgerrit Ghanshyam Mann proposed openstack/nova master: Add new default roles in os-services API policies https://review.opendev.org/648480
19:08:02 efried dansmith: +2, thanks.
19:08:02 efried Can I get you to look at the vTPM spec again please?
19:08:15 efried https://review.opendev.org/#/c/686804/
19:08:41 dansmith no
19:09:57 efried mriedem: if you have a minute, would you please take a look at https://review.opendev.org/695985 (lets us fix races to wait_for_instance_event())
19:12:33 dansmith efried: mriedem: just left a comment about _how_ we'll use this with cyborg to close that race: https://review.opendev.org/#/c/631244/46/nova/compute/manager.py@2627
19:14:32 efried dansmith: ++ makes perfect sense. Do you have a strong opinion about whether the bind should be kicked off from conductor instead of compute at this point?
19:14:53 dansmith efried: I think it'd be better if we did it from conductor for all the reasons I originally thought that
19:15:35 dansmith so unless there's some new revelation about why that's bad, I have the same opinion
19:16:01 efried I think it might have had something to do with how easy/hard it is to clean up. But I can't think why it would matter.
19:16:15 efried otherwise it's just a question of having to redo the code moar.
19:16:49 dansmith well, the major benefit is the overlap of work getting done to build the accelerator before we get down to the last minute
19:17:13 sean-k-mooney the only issue i was aware of was it would require use to pass the arq info as part of the spawn call to the compute
19:17:20 dansmith if you do it from conductor, you're letting cyborg know about things before you take the trip over rabbit to the compute, wait in the parallel build queue, set up volumes, networking, etc, etc
19:17:21 sean-k-mooney but we may already need to do that
19:17:49 dansmith the current code is not passing it, AFAICT, and building all of that into the compute,
19:17:55 dansmith which is another good reason not to do it on the compute
19:18:12 dansmith the more the compute implies about the request from the state of the system, the harder it is to effect an upgrade that changes those implications later
19:18:19 sean-k-mooney right but i think they choose to build it on the compute initally to avoid the rpc chagne to pass it
19:18:29 sean-k-mooney i prefer doing it in the conductor too by the way
19:18:39 sean-k-mooney i just think that is why the current code does not
19:18:39 dansmith right now, all the computes are examining the flavor directly and doing what they think they need to
19:19:16 dansmith if we change something about how that works, conductor gets upgraded first and can change the behavior, just passing the arq to the compute, which just needs to wait and hook it up
19:39:12 mriedem efried: the bind from compute/conductor thing was in the ML awhile back, probably worth dredging that up if we're changing our minds now
19:40:12 mriedem from what i remember, i think i also pushed for conductor but then after lots of wah wah it's easier from compute without a bunch of rpc interface changes i relented
19:40:26 dansmith mriedem: we had originally said it would be from conductor, then the spec changed
19:40:46 dansmith mriedem: the spec seemed to call out the eventing as the primary reason why it had to happen from compute
19:40:57 dansmith which is both not correct, and also laughably terrible in the implementation currently proposed
19:41:00 mriedem i don't remember the details, i just know it came up in the ML
19:41:22 sean-k-mooney it also came up in the denver and maybe dubling ptg
19:41:28 sean-k-mooney its been a while
19:41:55 sean-k-mooney i think long term we all agreed teh conductor was better.
19:42:00 mriedem ok, let me rephrase: the last time i remember there being thoughtful and documented takes on that was in the ML
19:42:11 mriedem rather than some undocumented stuff from dubling
19:42:20 sean-k-mooney i think at some point there was a request to avoid the rpc chagne but i dont really rememebr
19:42:33 mriedem "i don't really remember" is why i refer back to the ML
19:42:53 sean-k-mooney ya fair
19:42:57 dansmith I don't remember a thread on the ML, but there was discussion and documentation on the spec, which seemed to be the most fresh to me
19:43:54 dansmith https://review.opendev.org/#/c/603955/11/specs/train/approved/nova-cyborg-interaction.rst@250
19:44:01 dansmith this ^
19:44:23 mriedem wacka wacka http://lists.openstack.org/pipermail/openstack-discuss/2019-June/thread.html#6979
19:44:33 dansmith the discussion basically says either can do it, but it's hard to do it from conductor and then catch the event from compute
19:44:34 dansmith that is not a problem
19:46:01 dansmith mriedem: looks like that ended with you still opting for as much in conductor as possible
19:46:16 dansmith but no real summarizing agreement that it should be in compute
19:46:49 mriedem in http://lists.openstack.org/pipermail/openstack-discuss/2019-June/007013.html i'm saying please create arqs in conductor and bind in compute, like john was trying to do with ports at one time
19:47:09 mriedem bind generates the event if that's what is important
19:47:25 sean-k-mooney isnt that what we agreed to do
19:47:33 dansmith right, but bind is the thing that takes a long time,
19:47:41 dansmith so doing that in conductor is what gets us the most gain
19:47:55 dansmith the original email states the reason for moving is because of the eventing being hard, which is is not
19:48:03 mriedem having not really any skin in this game, it doesn't really matter to me anyone re this cyborg series
19:48:13 mriedem i agree the original email is mostly excuses
19:48:25 mriedem minimal change to get a thing to sort of work as a poc
19:48:42 mriedem to which i said essentially, yeah that's why we have a lot of turdy spaghetti code we've wanted to refactor for years
19:48:51 mriedem and why artom calls nova fugly
19:49:19 dansmith yeah, understand.. the reason for moving to compute based on the spec and that original email seemed to be excuses around eventing, which are not valid,
19:49:42 dansmith so I'd prefer we revert back to the original plan which had good reasoning, which is to overlap the programming with all the other things we can do in parallel,
19:49:47 dansmith which means bind at conductor
19:50:09 sean-k-mooney they baically need to do the equvalent of double check locking in the driver. check if its bound start waiting for event, check if its bound again incase you mised it and if so cancel waiting if not wait
19:50:17 mriedem does that include building a new interface into conductor so when the api receives the external event it can route it to conductor rather than compute?
19:50:40 dansmith sean-k-mooney: did you see the explanation I linked above about how to do the check and event without the race?
19:50:47 dansmith sean-k-mooney: they do not need to check twice
19:50:56 dansmith mriedem: no it does not
19:51:16 dansmith read this: https://review.opendev.org/#/c/631244/46/nova/compute/manager.py@2627
19:51:18 sean-k-mooney no but i trust you when you say there is a way to do that
19:51:24 sean-k-mooney but neighter woudl be hard to do
19:51:36 dansmith neighter?
19:51:53 sean-k-mooney neither
19:51:58 dansmith oh
19:52:07 sean-k-mooney ya that was more off then normal
19:52:12 dansmith the double check doesn't actually help, AFAICT, it would have to be N checks i.e. polling
19:52:47 sean-k-mooney well once you have register to wait for the event if you check the status and its not complete you can just wait
19:53:04 dansmith sean-k-mooney: right, that's one check and that's what I mention in the above comment
19:53:19 dansmith it's just that you need a new thing to be able to exit the wait after you have started if you know it is not necessary which is not easy today
19:53:41 dansmith you said "equivalent of double check" so, I see what you mean now
19:53:46 sean-k-mooney ah ok i was suggesting checking that api once ot skip that incase its already done but that is not needed
19:53:51 dansmith yes
19:53:58 dansmith skip is the thing we can't do without the patch I proposed

Earlier   Later