| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2019-12-09 | |||
| 13:49:52 | sean-k-mooney | :) | |
| 13:49:54 | kashyap | melwitt: Hiya, will look | |
| 13:50:09 | huaqiang | thanks | |
| 13:50:19 | sean-k-mooney | huaqiang: ill finish the email im typeing and review it then | |
| 13:51:02 | huaqiang | I prepared the POC code, under topic: bp/mixed-cpu-instance-set4 | |
| 13:51:18 | huaqiang | not good enough but it works somehow | |
| 13:51:59 | sean-k-mooney | huaqiang: ideally the topic shoudl match the bluepinrt name so it should be bp/use-pcpu-vcpu-in-one-instance | |
| 13:52:37 | huaqiang | sean-k-mooney: I'll make the name be consistent in next update | |
| 13:53:54 | kashyap | melwitt: Ah, never mind, I see it's merged - the CPU comaprison check on AArch64 | |
| 13:53:54 | sean-k-mooney | ok i just helps when trying to find all the related patches if they have the same topic as the spec and it matchs the blueprint | |
| 13:55:30 | huaqiang | I saw that, appreciate that. | |
| 14:04:15 | aarents | hi, | |
| 14:04:26 | aarents | mriedem: I don't know if you got some news from Matt Booth about https://review.opendev.org/#/c/696084/ , he's probably off or busy? | |
| 14:05:52 | kashyap | aarents: He normally goes by 'mdbooth' in this channel | |
| 14:08:27 | mdbooth | kashyap: Did you look at the arguments involved https://review.opendev.org/#/c/696084/2/nova/virt/libvirt/imagebackend.py ? | |
| 14:08:37 | mdbooth | Is that guaranteed to flatten the qcow2? | |
| 14:09:01 | kashyap | mdbooth: Afraid, not yet; was addressing something else. /me goes to look... | |
| 14:12:04 | kashyap | mdbooth: To flatten a chain *offline* shouldn't one require `qemu-img commit`? (Online is 'block-commit') | |
| 14:12:58 | mdbooth | kashyap: Right. I haven't checked the exact args used there, or refreshed my memory on qcow2 flattening, but I seemed to recall there was more involved | |
| 14:13:41 | kashyap | mdbooth: Definitely commit is required. | |
| 14:13:51 | kashyap | So says my 2012 "handout" even :D -- https://kashyapc.fedorapeople.org/virt/lc-2012/snapshots-handout.html | |
| 14:14:07 | kashyap | If you have: [base] <-- [overlay1] | |
| 14:14:28 | kashyap | To "flatten" it, qemu-img commit sn2.qcow2 | |
| 14:14:35 | kashyap | [Err, bad copy/paste.] | |
| 14:14:47 | kashyap | To "flatten" it, `qemu-img commit overlay1.qcow2`. | |
| 14:15:19 | kashyap | [If you have more than two files in a chain U2014which is it in our case — | |
| 14:15:51 | kashyap | [If you have more than two files in a chain — which is it in our case — then you'd also have to update the backing file pointer.] | |
| 14:16:09 | lyarwood | I don't think you need to do that anymore kashyap, I've only ever seen just a normal qcow2 to qcow2 convert used to flatten qcow2 disks. | |
| 14:16:26 | kashyap | lyarwood: Sorry, what is not required anymore? | |
| 14:16:34 | kashyap | lyarwood: Ah, the 'commit'? | |
| 14:17:05 | kashyap | Right, 'qemu-img conver' is another way. Which begs the question which method is preferred over the other, and why | |
| 14:17:33 | lyarwood | yeah, http://paste.openstack.org/show/787313/ | |
| 14:19:32 | kashyap | lyarwood: I _think_ the advantage of using 'convert' is that it retains some qcow2 properties... | |
| 14:26:02 | lyarwood | kashyap: yeah I think so, I've commented on the change anyway, for qcow2 I think we need to rebase on the original cached image otherwise each unshelve is going to eat up a whole load of disk with each call to flatten. | |
| 14:26:19 | lyarwood | mdbooth: ^ not sure if you agree | |
| 14:26:53 | mriedem | gibi_off: i replied to your questions in https://review.opendev.org/#/c/637070/ - let me know if you need anything else | |
| 14:27:20 | mriedem | elod: do you know when gibi is back? | |
| 14:28:07 | kashyap | lyarwood: There are _three_ ways in total Including 'rebase' -- I was remembering it only changes the backing file pointer. | |
| 14:28:13 | elod | mriedem: wednesday, if I'm not mistaken | |
| 14:28:17 | kashyap | lyarwood: Will update the change with the three ways, and pros/cons | |
| 14:28:21 | mriedem | elod: ok thanks | |
| 14:30:45 | lyarwood | kashyap: ah I wasn't aware that you could rebase disks between backing files, that would be super useful here. | |
| 14:30:51 | lyarwood | kashyap: thanks! | |
| 14:31:37 | kashyap | lyarwood: Yes, indeed. By default 'qemu-img rebase' does a "real rebase" option. And 'qemu-img rebase -U' -- "unsafe" -- will only update the backing file | |
| 14:31:57 | kashyap | The man page explains it. /me recalls documenting that part of 'qemu-img' many moons ago :-) | |
| 14:32:14 | kashyap | (Read the Safe mode vs. Unsafe mode discussion) | |
| 14:36:11 | aarents | mdbooth: kashyap Hi, basicly by using convert, I reused stuff done when we extract a snapshot to glance (we need to flaten before upload) | |
| 14:42:43 | kashyap | aarents: Hi, there are a couple of trade-offs here, based on the method we're using | |
| 14:43:38 | kashyap | 'convert' has the advantage of also handling sparsification (as guessed earlier); but has the disadvantage of being relatively slow _and_ requires double the space | |
| 14:44:52 | kashyap | aarents: Slow because, 'convert' copies both base and overlay into a new image; and thus double the space. (Instead of copying only into base or the overlay, in case of 'commit' or 'rebase') | |
| 14:45:23 | kashyap | aarents: Writing a further comment in the change, once we have the options laid out, then we can make a decision | |
| 15:00:21 | aarents | kashyap: ok great | |
| 15:18:17 | openstackgerrit | Matt Riedemann proposed openstack/nova master: Follow up to I5b9d41ef34385689d8da9b3962a1eac759eddf6a https://review.opendev.org/698028 | |
| 15:18:36 | kashyap | mdbooth: aarents: lyarwood: Added comparison notes of the three possible ways we can take: https://review.opendev.org/#/c/696084/ | |
| 15:20:54 | efried | slaweq: catching up... | |
| 15:23:18 | efried | slaweq: Any time I see spurious grenade fails in the last couple months, I attribute it to oversubscribed CI nodes, per mriedem's "State of the Gate" thread, started here http://lists.openstack.org/pipermail/openstack-discuss/2019-October/thread.html#10484 and continued here http://lists.openstack.org/pipermail/openstack-discuss/2019-November/thread.html#10502 | |
| 15:24:48 | openstack | bug 1844929 in OpenStack Compute (nova) "grenade jobs failing due to "Timed out waiting for response from cell" in scheduler" [High,Confirmed] https://launchpad.net/bugs/1844929 | |
| 15:24:48 | efried | slaweq: and per the first note in that thread, yes, the bug you identified (bug 1844929) is the one we're "tracking" the issue with. | |
| 15:26:10 | efried | ...and the right solution is to get the CI providers to tweak their environments accordingly. We would rather have lower job throughput and lower failure rates. | |
| 15:26:18 | efried | But so far there has been no reaction from them. | |
| 15:26:45 | mriedem | note that for that particular bug the vast majority of fails are on ovh nodes | |
| 15:27:33 | mriedem | i don't know why it's mostly on grenade jobs | |
| 15:27:58 | mriedem | must have something to do with restarting mysql a few times, but idk | |
| 15:29:00 | mriedem | note that it also started with train, i don't know why though | |
| 15:29:12 | mriedem | tl;dr i don't really know much of anything | |
| 15:33:26 | sean-k-mooney | do we also restart rabbitmq during the grenade upgrade. i assume so but i have not checked | |
| 15:33:58 | sean-k-mooney | it is porably restarted at least once by devstack on the second stacking | |
| 15:35:11 | sean-k-mooney | i noticed that the subnode hand messaging timeouts in both the n-cpu and q-agt services | |
| 15:35:57 | sean-k-mooney | so it looked like we lost messages or something | |
| 15:45:14 | slaweq | mriedem: efried thx for confirmation that this is the bug which we hit most of the times now | |
| 15:45:18 | efried | dansmith: it would seem that the 422 event code is masked when there's only one event (as is the case with cyborg) | |
| 15:45:25 | slaweq | sean-k-mooney: and thx for info about rabbitmq too | |
| 15:45:36 | efried | slaweq: It would help to get some more voices complaining at the node providers. | |
| 15:45:54 | dansmith | efried: really? I didn't see that condition | |
| 15:45:58 | efried | dansmith: when all events are dropped, the handler turns the whole thing into a 404 | |
| 15:46:06 | efried | https://github.com/openstack/nova/blob/master/nova/api/openstack/compute/server_external_events.py#L146 | |
| 15:46:41 | efried | which is arguably a bug | |
| 15:46:44 | dansmith | efried: ah, yeah, that code isn't really right.. "no instances found for any event" isn't true, just no hosts.. | |
| 15:46:46 | dansmith | yeah | |
| 15:47:05 | efried | to "fix" the bug would be a microversion? | |
| 15:47:07 | dansmith | unless mriedem feels that's a microversion problem, I can change that | |
| 15:47:31 | dansmith | I dunno, it's correcting the error code, which I thought was allowable in some situations | |
| 15:47:51 | efried | what would you correct it to? 207? | |
| 15:48:26 | efried | Meaning there's actually no way the API ever returns an error. | |
| 15:48:28 | dansmith | that's what we return per-event in that case right? | |
| 15:48:32 | dansmith | no, there is | |
| 15:48:42 | dansmith | if the instance itself is not found | |
| 15:48:50 | efried | instances plural, right? | |
| 15:49:21 | dansmith | yeah | |
| 15:49:22 | efried | so if some of the instances are not found, that's a 207 with some 404s | |
| 15:49:32 | efried | but if none of the instances are found, that's an overall 404? | |
| 15:50:28 | efried | If some events succeed, 207 with the granular payload | |
| 15:50:28 | efried | If all events tank, 400 with the granular payload | |
| 15:50:28 | efried | or, | |
| 15:50:51 | efried | If all events tank for the same reason, make that the overall status code | |
| 15:50:51 | efried | or, | |
| 15:51:10 | efried | otherwise use 400 or 207 | |
| 15:51:33 | dansmith | it really should be that if everything was a 404, then you get overall 404. otherwise it's 207 with granular statuses right? | |
| 15:51:58 | dansmith | or maybe if everything is 200, you return 200 as well, I dunno | |
| 15:52:34 | efried | I'm not offended by the idea that you always return 207 | |