Earlier  
Posted Nick Remark
#openstack-nova - 2021-10-08
15:18:13 sean-k-mooney lyarwood: when we have done this clean up in the past we have paired it with other changes with the logic being a client tha want to use that new feature via a microversion need to be updated anyway so they can adapt to both change at the same time.
15:18:21 dansmith let me ask this way: what client behavior is going to be wrong if we return 200 that will be right if we return 202?
15:18:31 dansmith if there's something there, then it's worth changing
15:18:39 dansmith sean-k-mooney: right
15:19:37 lyarwood tbh with 200 to 202 it's not our clients that I'm worried about
15:19:56 lyarwood it's other callers outside of them and their understanding of the API
15:20:21 dansmith lyarwood: right, but ... what's the behavior or understanding that will be wrong?
15:20:22 lyarwood the CSI plugin in k8s for OpenStack for example has been misusing it for a while
15:20:29 sean-k-mooney your assumeing they are reading the reponce code but not the docs that say this is an async api
15:20:43 lyarwood dansmith: assuming it's sync and not polling
15:20:46 dansmith meaning, generally 202 means "if you query this thing right after the PUT/POST then you will not see it updated so don't expect it to be" .. is it that? or is it "this operation takes a while?"
15:21:17 dansmith lyarwood: but is the actual API operations sync and just not the attachment?
15:22:04 lyarwood we return after we cast to the compute to do the actual attachment
15:22:09 sean-k-mooney dansmith: i think k8s has been assuming that the attachment was completed when it got the 200
15:22:32 lyarwood so it takes a while, but the CSI driver would start polling Cinder instead of Nova and that has caused various issues
15:22:48 lyarwood because they assumed the attachment in Nova was done at that point
15:23:05 dansmith I'm not sure that
15:23:17 dansmith is reasonable behavior to assume from either 200 or 202 though
15:23:47 dansmith you're giving it attachment information, which it is storing, and that triggers some behavior right?
15:24:21 dansmith like, IIRC, 200 is for "I've stored your object, here's the ID you can use to get it again" and 202 is for "I will store that later, but I can't necessarily give you its identifier yet" right?
15:24:27 sean-k-mooney for a 202 they cannot assuem its complete and should expect to poll nova for the compleation
15:25:05 lyarwood dansmith: yeah but my understanding with 200 was that processing was complete with the return
15:25:08 sean-k-mooney for 200 generally you would assume it blocks until its completed but its valid to retrun before that
15:25:17 lyarwood dansmith: and as sean-k-mooney said, 202 requires additional processing
15:25:34 dansmith I think that's true of the REST object and not necessarily the actions implied from the thing you're storing
15:26:18 sean-k-mooney lyarwood: has the k8s csi plugin be fixed to pool nova by the way
15:26:21 dansmith but again, as sean-k-mooney said, this is really a result of them not reading the docs that say it's async and implying which part of the operation is synchronous from the 200
15:26:26 sean-k-mooney regardeless of if we cahnge the api
15:26:36 lyarwood sean-k-mooney: https://github.com/kubernetes/cloud-provider-openstack/issues/1645 not yet
15:27:28 sean-k-mooney ok but they are aware of it and it will get eventually fixed
15:27:33 lyarwood yup sure, just outlining my motivation to fix this
15:27:51 lyarwood ;)
15:27:58 sean-k-mooney :)
15:28:17 sean-k-mooney i assume we have a customer hitting this so that normally seed up fixing things
15:28:21 dansmith to be clear, making it 202 won't fix this, we're just assuming the human that wrote the code would have assumed they should do something different if the code was 202 right?
15:28:38 lyarwood yup that's the assumption, it wouldn't fix anything in CSI
15:29:01 lyarwood I got triggered by the TODO today and wanted to take a swing a killing it
15:29:06 sean-k-mooney if we wnated to fix CSI issue in nova we would have to block
15:29:11 dansmith so, I mean.. we're like changing something that could easily break other clients that did read the docs, based on the assumption that someone who didn't would have done the right thing if it was different
15:29:14 sean-k-mooney wich i dont think we want to do
15:29:18 dansmith pretty meh :)
15:30:26 bauzas ok, sorry guys, I need to bail out from now
15:31:48 lyarwood bauzas: \o if you want to chat on Monday ping me an email, I should have time to chat once I'm on London
15:31:55 lyarwood in*
15:32:15 bauzas lyarwood: don't worry, I'll use the power of community brainstorm to move on with your spec ;)
15:32:32 dansmith I think we've all put on a few pandemic pounds, but being "on london" would require a substantial increase in waist size, I think
15:33:15 lyarwood dansmith: it has been a rough 16 months, it wouldn't take much more
15:33:24 dansmith heh
22:23:34 opendevreview Julia Kreger proposed openstack/nova master: Ignore plug_vifs on the ironic driver https://review.opendev.org/c/openstack/nova/+/813263
#openstack-nova - 2021-10-09
03:35:26 opendevreview norman shen proposed openstack/nova master: Recreate mdev devices according to placement https://review.opendev.org/c/openstack/nova/+/810220
04:04:16 opendevreview norman shen proposed openstack/nova master: Recreate mdev devices according to placement https://review.opendev.org/c/openstack/nova/+/810220
13:08:56 opendevreview Julia Kreger proposed openstack/nova master: Ignore plug_vifs on the ironic driver https://review.opendev.org/c/openstack/nova/+/813263
#openstack-nova - 2021-10-10
14:19:47 opendevreview norman shen proposed openstack/nova master: Add a timeout for waiting volume snapshot ready https://review.opendev.org/c/openstack/nova/+/813297
#openstack-nova - 2021-10-11
07:45:43 bauzas good morning Nova
07:52:06 opendevreview Dmitriy Rabotyagov proposed openstack/nova master: Ensure MAC addresses characters are in the same case https://review.opendev.org/c/openstack/nova/+/811947
08:24:47 plibeau4 hello, I need your help about this proposal change: https://review.opendev.org/c/openstack/nova/+/803713.
08:50:25 gibi good morning
08:51:03 gibi this week feels to be as crazy as the last was for me. you can ping me but I cannot promise things :/
08:55:43 bauzas plibeau4: ack, will add your change for a review
08:56:09 bauzas thanks for reaching us
09:24:41 plibeau4 thx guys take your time it's just to have your feedback It's already in production on myside :)
09:25:17 plibeau4 and if you have time -> https://review.opendev.org/c/openstack/nova/+/803600
09:26:39 sean-k-mooney to me https://review.opendev.org/c/openstack/nova/+/803713 kind of feels like intoducing a race condetion
09:29:12 sean-k-mooney i guess its not a race as it does not quiesce the disk until after the dev rebase is complete deterministically
09:43:45 sean-k-mooney plibeau4: im not conviced moving the quiece is safe to have in production
09:44:09 sean-k-mooney plibeau4: i think it can lead to data currportion if we are not very careful
09:50:05 gibi sean-k-mooney: re: waiting for vif-plugged during hard reboot. What is the piece of information in the port that can be used to decide if the networking backend sends plug event at plug time?
09:53:35 sean-k-mooney gibi: https://specs.openstack.org/openstack/neutron-specs/specs/train/port-binding-extended-information.html
09:53:45 sean-k-mooney binding_details: {
09:53:47 sean-k-mooney ...
09:53:49 sean-k-mooney "connectivity": "l2",
09:53:51 sean-k-mooney "bound_drivers": {"0": "openvswtich"}
09:53:51 gibi thanks
09:53:53 sean-k-mooney }
09:54:00 sean-k-mooney gibi: we could use the bound_drivers to infer this
09:54:13 sean-k-mooney the spec orginaly had an events section that told use when it sent events
09:54:42 sean-k-mooney that was unfrotunetly removed because we could not agree on how to model that
09:55:12 sean-k-mooney so we would have to create a lookup table in nova and map the driver name to it
09:55:36 sean-k-mooney opendaylight was bind time, opendaylight_v2 is plug time
09:56:36 gibi ack. I might need to create a solution on Pike level downstream hence my continued digging in it
09:56:41 sean-k-mooney the conneectivy field is waht we need to use to know if we can safely allow unadress port too
09:56:45 gibi but if the infor is in train then I cannot use that in Pike
09:56:48 sean-k-mooney if its l2 it does not need an ip
09:57:43 sean-k-mooney gibi: i think a workaround config option to force_rebind_on_reboot is the way to go and backport that
09:57:55 gibi yeah I thinking about a config too
09:58:02 sean-k-mooney there are reasons to have that on master by the way
09:58:11 sean-k-mooney combidnind with unpluging on reboot
09:58:52 sean-k-mooney basically if you make reboot: power off-> unplug-> unbind->rebind->plug->spwan
09:59:21 sean-k-mooney then you can do things like change the firewall driver or maybe even ml2/backend
09:59:22 gibi why do you want to add bind to the picture?
09:59:28 gibi ahh I see
09:59:37 sean-k-mooney currently you have to live migrate
09:59:40 gibi to allow change thing that needs a binding
09:59:53 sean-k-mooney but we have some customer saying they want to do it inplace and are ok with stoping the vm
09:59:58 sean-k-mooney yes
10:00:55 sean-k-mooney basicaly some large customes have said to use we cant live migrate every vm we have but we can stop them cause we run everything in duplex mode and can go simplex for a period of time without issue
10:01:45 gibi OK so there is two parallel direction to pursue. 1) add a workaround flag for something like wait_for_plug_event_at_reboot and 2) force_rebind_on_reboot to allow in place update of binding related data
10:02:06 sean-k-mooney so they are fine with stop(tearing down all networking) and start(setting up all networking) to "migrate" form one firewall driver to another (iptables->ovs) or between mech drvers(ml2/ovs->,l2/ovn)
10:02:31 sean-k-mooney gibi: am yes i guess so
10:02:51 gibi the problem (most probably) on my side is that they do VM reboots in big batches due to upgrade and nova sends a lot of unplug - plug event to ODL which fails to finish setting up OVS before the VM starts requesting IP via DHCP

Earlier   Later