17:00:15 #startmeeting cinder-nova-api-changes 17:00:16 Meeting started Mon Jan 16 17:00:15 2017 UTC and is due to finish in 60 minutes. The chair is ildikov. Information about MeetBot at http://wiki.debian.org/MeetBot. 17:00:17 Useful Commands: #action #agreed #help #info #idea #link #topic #startvote. 17:00:20 The meeting name has been set to 'cinder_nova_api_changes' 17:00:23 scottda DuncanT ameade cFouts johnthetubaguy jaypipes takashin alaski e0ne jgriffith tbarron andrearosa hemna erlon mriedem gouthamr ebalduf patrickeast smcginnis diablo_rojo gsilvis xyang1 raj_singh lyarwood 17:00:56 o/ 17:01:22 o/ 17:01:37 hi :) 17:02:07 let's wait a bit hoping we will have more people around 17:03:12 ok, lets dive in 17:03:18 we have a few patches up for review 17:03:40 in the server side spec there's a comment from patrickeast regarding terminate_connection and ordering 17:03:44 #link https://review.openstack.org/#/c/414632/11/cinder/volume/manager.py 17:04:01 line 4542 17:04:21 mriedem: did we agree on anything on the Nova side about what the ordering of things is when we detach? 17:04:42 umm 17:04:48 as in today or in the future? 17:04:56 i thought we had like one detach call now? 17:04:59 in the future 17:05:17 not in yet and there is one attachment-delete call 17:05:36 from what i remember of johnthetubaguy's spec, we detach with cinder api and then get something back, and pass that to os-brick's disconnect_volume 17:05:47 although my understanding is that Nova will still call out to os-brick for instance 17:05:49 and the issue was if disconnect on the host fails 17:06:06 mornin 17:06:16 hemna: morning :) 17:06:19 since we don't have like a terminate_connection - disconnect - os-detach (ack) flow anymore 17:06:37 jgriffith: Are you around? 17:07:31 * diablo_rojo_phon sneaks in the back of the room 17:09:00 mriedem: ok, I need to re-read that part 17:09:27 * jgriffith shows up late 17:09:56 jgriffith: we're talking about patrickeast's comment and detach/disconnect in general 17:10:34 jgriffith: I mean the one in line 4542 here: https://review.openstack.org/#/c/414632/11/cinder/volume/manager.py 17:11:24 looking 17:11:51 oh, yeah; so I can change that 17:11:58 yah patrick is correct 17:12:01 but I have a fundamental problem with the FC stuff at this ponit 17:12:02 point 17:12:16 jgriffith: Haven't you always? ;) 17:12:23 Patrick may be correct, but the FZM stuff is obviously completely f'd up 17:12:23 the FCZM code will remove the zone if the target map is passed back in terminate_connection 17:12:48 it uses it's own interpretation of things and ignores every other driver in the code base 17:13:04 it's frankly super annoying 17:13:14 but anyway 17:13:17 fczm code doesn't care about non FC drivers, and only looks for the FC related information 17:13:29 hemna yes, but an interface is an interface! 17:13:38 we are changing the terminate_connection behavior, so we need to adjust things accordingly 17:13:56 You can't just choose to inherit from a base class and change it however you feel in this case 17:14:19 hemna I'm not aruging this here, I'll work on addressign the FZM stuff 17:14:21 some FC drivers don't inherit from the FC base class 17:14:23 nuf said 17:14:36 I've tried many times to get folks to adopt it 17:14:54 but there are several FC drivers that are combined with iSCSI drivers 17:15:10 instead of having 2 drivers iSCSI and FC, they combine them both into 1 driver 17:15:11 :( 17:15:22 so there is that too 17:15:33 hemna yeah and frankly that's a good thing, again I don't want to air this out in this meeting 17:15:41 not the appropriate time or place 17:15:49 ok, just sharing information. 17:15:53 since it's relevant 17:17:41 jgriffith: is the plan for attachment_delete to return a True/False to indicate whether there's a shared connection or not? 17:17:54 ildikov that *was* the plan yes 17:18:11 ildikov that had been the discussed proposal for a while now, but that's not going to work with the FZM stuff 17:18:20 at least, not the way it is currently 17:18:47 do we have alternatives? 17:18:54 not really 17:19:08 I mean, no real alternatives other than fixing the mess 17:19:27 and making it work for both. Ideally making things consistent again 17:19:39 wait 17:19:40 so 17:19:41 so the current plan is to fix the mess and stick to the above plan? 17:19:49 the decorator is on terminate_connection in the driver 17:20:18 that code has already been executed by the time _connection_terminate returns 17:20:43 jgriffith: I kinda like the idea of either changing the fczm code to be smarter with the wrapper function return value or a new driver API that can just like list shared connections or something 17:20:47 the drivers are only supposed to return the target map if there are no more connections from the backend to the nova host 17:21:06 so I'm not sure anything is wrong here 17:21:30 the zoning has already been done by the time that return call is made 17:21:45 hemna: yep, the fczm can just return a book correctly based on the map it got 17:21:50 and drivers are already tracking if there are connections left or not. 17:21:57 Bool* 17:22:19 patrickeast I liked "book" better :) 17:22:29 Lol 17:22:53 the zoning code just returns what it got from the driver 17:22:57 so I don't see a problem here 17:24:46 I don't see a conflict 17:24:50 or I'm just slow 17:27:21 It's backwards is all, the fczm needs to return true when it didn't get a request to tear down zones since the volume connection is still in use on that initiator.... I think 17:27:59 So like empty map means return true, map with stuff to clean means return false 17:28:31 Assuming my sleepy logic checks out 17:28:49 patrickeast yeah, so the "absence" of anything currently means "nothing shared, nothing to do" 17:28:56 Yea 17:29:01 patrickeast from the new codes perspective that is 17:29:23 the fczm just returns what the driver does 17:29:43 so, the driver can simply return it, because it already knows there are no more connections or not. 17:29:52 the FCZM's job is to add or remove zones only. 17:29:58 hemna but you have a return in your decorator as well 17:30:13 which is a pass through from the driver 17:30:42 Right, my proposal is that we make it a little more than a pass though 17:30:56 So that it complies with the new api 17:30:57 to do what? 17:31:20 To return true or false 17:31:37 Like the new terminate connection should 17:31:44 patrickeast +1 17:31:56 So iscsi drivers will return true/false but FC drivers will return a dict? 17:32:19 we are changing every driver to return something different now in terminate_connection ? 17:32:21 ?! 17:32:23 smcginnis: yea, they need to know that the fczm monkeys with things 17:32:47 patrickeast: Not a big fan of it, but if we can get that to work, I guess I'm fine. 17:32:59 hemna the idea is to have a consistent interface, which would be kinda nice IMO 17:33:08 jgriffith: +1 17:33:12 There's already a presumed assumption that FC driver developers know to use the decorator and what to return for the decorator. 17:33:29 jgriffith: +1 for consistent interface. 17:33:37 part of why all of this work and things like multi-attach have been difficult is lack of consistency in interfaces 17:33:44 everybody did "their own thing" 17:33:46 They could all just return maps of initiators similar to the fczm 17:33:56 patrickeast and that's fine as well IMO 17:34:16 patrickeast just so long as it's consistent and we know what it means 17:34:22 patrickeast: I kind of like that just because then it is also consistent from the driver implementation side. 17:34:42 Ehh kinda 17:34:57 the point for me is I would like to get rid of all the snow-flake implementations of attach/detach 17:35:09 that was the whole point of reworking these to begin with 17:35:38 I don't care what you do in the driver of course, but the interfaces and what they are perceived to do should be consistent 17:35:47 so, we could also change the way drivers interact with the fczm 17:35:56 and have the drivers call the fczm remove code manually 17:35:58 Everyone has their own idea of hosts and initiators, it might be similar for FC where the spec sorta forces it, but on iscsi and other protocols it might get more weird 17:36:00 instead of as a decorator 17:36:16 hemna: +1 17:36:17 that would make the driver return from terminate_connection consistent 17:36:27 My preference is always return an initiator map, or have an explicit call for the FC management. 17:36:41 personally I would like to see the decorator 86'd 17:36:47 It is almost universal in new driver reviews that we have to point out FC drivers need a decorator and what they need to return for it 17:36:52 * jgriffith is not a fan 17:37:12 I learn towards removing the decorator if we aren't going to modify it transparently 17:37:21 this is one of the reasons why I wanted FC drivers to extend the FC base class 17:37:32 some FC drivers don't use the zone manager 17:37:32 then we could modify the base FC class and everyone gets the changes for free 17:37:46 xyang1, yah and those FC drivers don't work :) 17:37:52 in a real openstack deployment 17:38:16 hemna: they work using their own zone mgmt 17:38:29 Well... we actually use open zoning quite a bit since the array handles isolating the target and initiator. 17:38:47 So fczm just adds complexity. 17:38:49 smcginnis +1 same here 17:38:52 ok do you want me to look at a follow up patch? 17:38:59 but I'm still resisting actually supporting it anyway 17:39:04 :) 17:39:10 jgriffith: :) 17:39:22 well some arrays don't have fabric management 17:39:35 hemna honestly I'd prefer we have an answer and include it in the proposed patch 17:39:39 hemna: Right, some I think absolutely need it. 17:39:51 there have been WAY too many cases where we say "follow up" and it never happens, or doesn't work 17:39:54 I don't think it would be hard to adjust things 17:40:00 leading to more and more 1/2 in 1/2 out impls 17:40:09 ok do what you want 17:40:09 SO what are our options? All return initiator map or explicit FCZM calls in the driver? 17:40:12 I was just offering help 17:40:40 I'd prefer explicit fczm calls in initiialize_connection and terminate_connection. 17:41:53 hemna: I kind of lean that way too. Then we can have a cleaner interface and more explicit code in the drivers. 17:42:07 patrickeast: Was that what you were saying too? 17:42:42 Yep 17:43:16 Thats my preference if we are going to make changes to all the FC drivers 17:43:30 jgriffith: Work for you? I just wonder how we're going to get all the drivers updated, but I guess we will need to either way. 17:43:36 sure 17:45:47 IMO just fiddling with decorator return values isn't so bad for an interim solution though, wouldn't require additional driver patches to work with the new APIs 17:46:16 But if the concern is that it's too confusing I'm OK with not doing it 17:46:26 I just wanted to ask what's the roadmap for this what you all seem to agree on? 17:47:01 just to clarify this for less knowledgable people on the driver side like me :) 17:47:18 To get all of them updated? 17:47:44 to get to the next step with the new attach/detach API 17:48:10 if everything needs to be updated before anything happens with that than yes 17:49:10 sorry for killing the joy with administrational questions :/ 17:49:22 Lol 17:49:54 patrickeast I'm fine with fiddling with the decorator to make everything consistent 17:50:14 patrickeast and if later we want to dump the decorator and force drivers to impl the method that's fine 17:50:28 patrickeast my only requirement is consistency 17:50:39 don't care how that's done right now 17:50:55 jgriffith: sounds good to me 17:51:08 I was just confused on what if any initiators *needed* that return info 17:51:20 if they don't need it anyway, was confused on why it was there :) 17:51:20 so any updates on getting nova to use cinder v3 in the last <10 min? 17:51:55 mriedem: did you see the latest patch scottda uploaded last week? 17:52:34 mriedem: that solves the endpoint not found error 17:53:16 mriedem: although I'm not sure what's the preferred way of setting the api micorversion for the cinderclient for instance 17:53:29 ildikov: do you mean this? https://review.openstack.org/#/c/420201/ 17:54:09 mriedem: yeap, sorry didn't have the link handy 17:54:11 re: the microversion to use, i commented on that here https://review.openstack.org/#/c/385682/3/nova/volume/cinder.py@100 17:54:19 nova definitely doesn't want to use the latest available 17:54:32 nova needs to say, i want 3.x, is that available? 17:54:50 if nova needs 3.24 and it's not available, then we can't do whatever we needed 3.24 for 17:55:12 the client has to opt into whatever version it's going to use so it can handle backward compat issues 17:55:43 mriedem: did you see the comment about the config option here: https://review.openstack.org/#/c/420201/2/nova/volume/cinder.py ? 17:55:53 line 98 17:56:41 we don't want a config option 17:57:02 that was my thought too 17:57:25 but wasn't sure how it is supposed to work in general 17:57:43 nova is going to hardcode the microversion it needs 17:57:52 if cinder can provide it, we use it, else we fallback to the old flow 17:58:31 ok, that sounds good to me 17:59:30 btw http://logs.openstack.org/01/420201/2/check/gate-tempest-dsvm-neutron-full-ubuntu-xenial/e66fd9a/logs/screen-n-api.txt.gz?level=TRACE 17:59:45 so i guess he's around the endpoint issue which is good 18:00:16 I think the get_version patch is not merged yet in the client 18:00:29 we're at time btw 18:00:49 mriedem: https://review.openstack.org/#/c/420119/ 18:01:03 mriedem: yep, we are, tnx 18:01:16 anything urgent for this meeting? 18:02:09 ok thanks everyone, let's catch up on the Cinder channel if there's anything 18:02:12 ildikov: Thanks. 18:02:28 otherwise chat next week! 18:02:33 #endmeeting