Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2017-09-20
17:29:54 tasker 3; no. it and all other instances were built to the same requirements.
17:30:33 tasker melwitt: does that have a memory? or is each migration request independent of the previous?
17:30:43 melwitt each request is independent
17:31:30 tasker the only thing different is that this instance was built while the cluster was at M. now my cluster is at N.
17:32:06 melwitt I would grep for the request-id of the failing live migration request in nova-conductor and nova-compute logs and see if there's anything about it
17:32:31 tasker I am live watching all the logs.
17:32:46 tasker conductor says "no valid host" because that's what scheduler says.
17:32:53 mriedem i'm wondering if the retry filter shouldn't be ignored here...
17:33:16 tasker if i send another instance to the same target, retryfilter passes the host.
17:33:42 melwitt RetryFilter removing a host implies that it tried to run the pre-migration on a host already
17:34:01 melwitt so something is weird here
17:34:21 tasker nothing logged at the compute at INFO or lower.
17:34:25 mriedem i'm wondering if the original instance request spec has retries set on it, and that's goofing up the retry filter during live migration
17:34:55 tasker something I could see in the database?
17:35:30 mriedem yeah, you should be able to pull the serialized json request spec out of the db
17:36:04 mriedem select spec from nova_api.request_specs where instance_uuid=xyz;
17:36:34 mriedem melwitt: this is where we get the request spec during live migration https://github.com/openstack/nova/blob/master/nova/conductor/tasks/live_migrate.py#L227
17:36:40 mriedem we don't set a retry attribute on the request spec
17:37:16 mriedem we pass it along from the api if we are able to look it up https://github.com/openstack/nova/blob/master/nova/compute/api.py#L3860
17:37:36 mriedem which would be the original request spec from when we built the instance, which would have retry stuff set on it
17:38:03 mriedem if that wasn't set, we'd ignore it here https://github.com/openstack/nova/blob/master/nova/scheduler/filters/retry_filter.py#L31
17:38:42 mriedem which, this is all kinds of weird because the conductor task is handling it's own retry logic https://github.com/openstack/nova/blob/master/nova/conductor/tasks/live_migrate.py#L344
17:38:53 mriedem which tells me, when doing a live migration, we don't want to even ask the RetryFilter
17:39:01 mriedem bauzas: are you around?
17:39:44 mriedem tasker: if you can get that request spec json blob out of the db, that should tell us if the retry attribute is set and what hosts and how many times it's been retried during the initial build
17:41:43 tasker mriedem: yes, "retry" is set and is populated with both hypervisors that I'm trying to send it to.
17:42:16 tasker would you like to see it?
17:42:43 mriedem sure, throw it in a paste
17:46:32 mriedem i think this would actually fix the problem http://paste.openstack.org/show/621556/
17:46:40 melwitt mriedem: so the first time you build an instance, if it fails host X, it will save that to the RequestSpec so host X will never be tried again ever? I didn't know it persisted something like that
17:46:47 mriedem oops one issue there
17:47:00 tasker http://paste.openstack.org/show/621557/
17:47:09 mriedem melwitt: yeah
17:48:54 mriedem melwitt: i think this https://github.com/openstack/nova/blob/master/nova/conductor/manager.py#L583
17:49:05 mriedem conductor build_instances is called from compute on a reschedule
17:49:14 melwitt well, I learned a thing. I had thought retries were a locally tracked thing per request
17:49:14 mriedem and will put the chosen host in the retry object's list of hosts it's tried
17:49:28 mriedem request spec is the persisted object that keeps on giving
17:49:32 mriedem even when you don't want the gift
17:49:59 mriedem so i think this is the fix http://paste.openstack.org/show/621558/
17:50:16 melwitt yeah. I wouldn't have chosen to track retries _forever_. I thought RequestSpec would only contain the original request requirements so they could be honored in the future
17:50:17 mriedem just like how we don't want the original forced host/node during live migration, we don't want the original retry hosts either probably
17:51:45 tasker mriedem: I'm going to try your patch out.
17:52:01 mriedem tasker: yeah so the request spec had 2 original attempts, on compute-3.openstack.local and compute-2.openstack.local
17:52:11 tasker both targets I'm trying to send it to.
17:52:16 mriedem ok
17:52:30 mriedem and that's why this fails https://github.com/openstack/nova/blob/master/nova/scheduler/filters/retry_filter.py#L44
17:52:36 tasker which part of nova would I apply that to? compute, conductor, or scheduler?
17:52:50 mriedem conductor, but note, that diff is against master branch code
17:53:20 tasker let me see if I can adapt to the branch that I'm using.
17:54:09 mriedem http://paste.openstack.org/show/621559/ is latest stable/newton
17:55:12 tasker cool, thanks. I'm going to step out for some "fresh air" and then apply this.
17:55:19 mriedem it does stand to reason that if this instance failed to build originally on those 2 hosts, that live migrating it there might fail too...but we don't know why it originally failed, could have been a resource claim issue at the time
17:55:36 tasker that's a fair observation.
17:57:56 melwitt yeah, often it's a failed claim. and also what if that compute host is eventually replaced over the lifetime of the cluster, making it a fresh candidate for several instances that might still avoid it because they once failed to build there back when it was a different machine
18:06:45 tasker live migration successful
18:08:17 tasker mriedem and melwitt -- thank you so much for helping to figure out what was wrong.
18:11:01 mriedem nice
18:11:07 mriedem tasker: want to open a bug? i have a fix with the test locally
18:11:40 tasker sure. let me collect my notes.
18:12:31 cdent dansmith: responded to some of your comments on https://review.openstack.org/#/c/500410/ . you semi-accidentally identified a separate problem. The nullable thing I’m not quite sure how to proceed, depending on what we want to do.
18:14:56 mriedem cdent: congratulations, you have the first complete blueprint in queens https://blueprints.launchpad.net/nova/+spec/placement-deregister-objects
18:15:35 dansmith cdent: replied, I don't think there's an issue.. just make those not nullable and (separately) always set them and I think we're good
18:24:25 cdent dansmith: except that they are only ever used on input, never output, so why bother reading them?
18:25:01 tasker mriedem: https://bugs.launchpad.net/nova/+bug/1718512
18:25:02 openstack Launchpad bug 1718512 in OpenStack Compute (nova) "migration fails if instance build failed on destination host" [Undecided,New]
18:25:07 mriedem thanks
18:25:19 dansmith cdent: because it's (effectively) free, it makes the object consistent
18:25:26 tasker I hope that summary / description is clear enough.
18:28:40 cdent dansmith: okay, I’m happy to do that. I’m not sure why consistency matters _now_ but if we’d like it as a general rule, that’s fine.
18:29:48 dansmith cdent: clearly it doesn't have a user now, but we're an abstract model on top of the data store and unless there is a reason not to, we should do that thing, IMHO
18:30:09 openstackgerrit Matt Riedemann proposed openstack/nova master: Ignore original retried hosts when live migrating https://review.openstack.org/505771
18:30:17 mriedem tasker: ^ i'm going to wip that and send something to the dev/ops lists for this, because if we change this for live migration, we also have to do it for cold migrate/evacuate and unshelve
18:30:23 mriedem melwitt: ^
18:30:43 melwitt ack
18:30:58 cdent dansmith: does YAGNI count as a reason not to?
18:31:42 dansmith cdent: I dunno what GNI is in there, but if you think there's a reason not to, then don't do it and we can just leave my -1 on there
18:32:52 cdent I’ve just always been brought up on the idea of not doing things unless there is a reason to do so, but as I said, I’m happy to do it and will do it.
18:34:24 mriedem if they are in the same table it seems easy peasy to just load them, it'd be one thing if we were lazy loading a big table join or something,
18:34:42 mriedem but if/when someone needs these, it's going to be weird debugging why they just aren't already in the object when it's read from the db
18:34:45 mriedem like every other object we have
18:34:48 dansmith mriedem: it's a join, but it's super tiny
18:34:52 mriedem oh
18:35:12 mriedem like the usages/consumers table join thing?
18:35:13 dansmith mriedem: like a 1:1 integer join
18:35:15 dansmith yes
18:35:16 dansmith consumers
18:35:39 mriedem i'm assuming at some point we actually intend on using these fields?
18:36:07 dansmith mriedem: if they're fields they should be loadable, and if we're never going to use them that way then maybe we shouldn't have them be fields
18:36:31 mriedem yeah i'd agree with that
18:36:36 mriedem i don't have context here though clearly
18:36:46 dansmith since consumer isn't a top level object in placement, getting them with allocations is the only way we can pull them out otherwise
18:36:51 dansmith unless we're querying for a consumer already
18:37:18 dansmith like if we needed to get all allocations against a resource provider
18:48:45 mriedem tasker: http://lists.openstack.org/pipermail/openstack-operators/2017-September/014233.html
18:49:08 openstackgerrit Chris Dent proposed openstack/nova master: WIP: [placement] manage cache headers https://review.openstack.org/495380
18:55:54 tasker mriedem: thanks! I'll keep an eye out on these.
18:57:45 tasker mriedem and melwitt: I want to thank you two for taking the time to work with me on my issues for the last few days. your support has been a positive influence in my cloudy days. with your help, I've been able to progress with my upgrade trials.
18:59:20 tasker If you two are ever in the Kansas City area, let me know. first round's on me.
19:05:32 mriedem tasker: cool, you're welcome. you hit some weird issues so i'm glad we flushed those out and have fixes

Earlier   Later