13:00:04 #startmeeting nova api 13:00:04 Meeting started Wed Nov 23 13:00:04 2016 UTC and is due to finish in 60 minutes. The chair is alex_xu. Information about MeetBot at http://wiki.debian.org/MeetBot. 13:00:05 Useful Commands: #action #agreed #help #info #idea #link #topic #startvote. 13:00:07 The meeting name has been set to 'nova_api' 13:00:11 who is here today? 13:01:18 o/ 13:01:45 cdent: good to see you get well 13:02:17 Thanks. I'm not 100% yet but bored, so getting on irc is a thing to do :) 13:03:32 emmm....holiday weeks 13:03:49 indeed 13:04:49 actually only thing want to talk about is about spec https://review.openstack.org/393205, people get different opionion, probably i will reach johnthetubaguy later to ensure what we should now 13:05:21 I tend to agree with jay: remove the vectors for bad 13:05:32 * johnthetubaguy wonders in late 13:06:03 johnthetubaguy: ah, just talk about https://review.openstack.org/#/c/393205/ 13:06:22 cdent also tend to jay 13:06:25 I think my problem is that "bad" is relative, in small deployments those queries are quite quick, and folks might be counting on them 13:06:56 yea, i'm more afraid to break the api, so tend to johnthetubaguy 13:07:03 I wonder if we just make it safe by default, i.e. policy stops them by default? 13:07:18 but folks that rely on them can restore the API, if needed 13:07:26 johnthetubaguy: after we have capabilities discovery? 13:07:29 although, that feels a bit like over-engineering 13:07:37 i was just gonna say that 13:07:48 alex_xu: I am thinking before we have it, just afterwards it could be discoverable 13:08:29 we got direct feedback on this at the summit, from operators saying don't break my searching 13:08:57 thats what really worry, feels like we are just ignoring them 13:09:34 ... at the same time, we should protect our users from cutting their legs off 13:10:12 is it worth an email to get board discussion? 13:10:20 I can't remember we were debating the user query or just the admin query 13:10:33 it seems that on the admin side, perhaps enough rope to hang yourself is okay? 13:10:51 thats where I my thinking about policy comes in really 13:11:11 I am OK restricting the user query by default (and I think the operators where too) 13:11:44 I am tempted to just restrict everyone by default, but let you relax it, if you want to take that risk 13:11:54 at least for the admins 13:12:06 but again, feels overcomplicated, so I don't feel happy with that direction as such 13:12:47 but not breaking users can really suck sometimes, maybe this is just one of those time 13:12:48 s 13:14:02 hmm, we got network delays again, or is everyone just very quiet today? 13:14:24 i'm jus thinking 13:14:51 thats allowed I guess :) 13:15:01 so....those query parameters already being here for very long time, people should cut their leg many times.... 13:15:14 cdent: do you think if its clear we protect people by default, that spec seems more reasonable? 13:15:17 * cdent has slow brain 13:16:02 Yeah, if the default is protection, but it can be turned off, that seems likely to make folk happy without totally rejecting the input of operators 13:16:41 turned off by policy? 13:16:56 yeah, off by default in the policy 13:17:21 and only off the admin part? 13:17:21 it does mean if we missed some things, we can tweak the groups of queries linked to each policy 13:19:10 filters won't break the user, it only ignore those filters. it just didn't return expected sliently 13:19:15 I duno, I was thinking off by default for everyone 13:19:30 sort will break user, the 400 returned 13:20:00 hmm, true, I guess this should be ignore using policy 13:20:05 based on the overall plan 13:20:15 a new microversion could add in the errors to help folks later 13:20:42 how about we ignore for invalid sort key? 13:21:01 at least it won't break the user. 13:22:03 yeah, I think we should ignore the not allowed ones, to be consistent 13:22:06 filters and sorts 13:23:14 or return 403 is enough, most of client should process the case when the request return 403 13:24:23 s/403/401 13:25:57 unsure 13:26:05 that feels inconsistent 13:26:28 in the future, we can have a microversion that does what we want, and rejects all input thats not acepted/acceptable 13:26:44 johnthetubaguy: so break them vs. ignore silently 13:27:10 johnthetubaguy: +1 microversion in the future to rejects all and return 400 13:28:03 yeah, for now, I think we are choosing silently ignore invalid input, unless its an allowed key that has an invalid value 13:28:22 to mirror the old behaviour, as close as we can, in a general way 13:30:42 the twists and turns we take "to mirror the old behaviour, as close as we can, in a general way" throughout all this stuff are both admirable and depressing 13:30:54 it's a nice quotable phrase too 13:32:53 I'm a little tent to the 401 returned, the 401 should be processed by client in general. I'm just afraid we return the result the client didn't expected, then just return a fault result siliently 13:32:58 yeah, its hard to describe how technical debt mounts up, thats sure a big cause of it 13:35:10 you mean 403? 13:35:27 emm....what we return when policy failed? 13:35:35 I think 403 13:35:51 401 suggests you need a new token I guess? 13:36:40 authenticated vs authorised I am assuming? 13:37:19 so what do we want to do next? 13:37:32 draft a new version of the spec, and see how that looks? 13:37:37 yea 13:37:46 yeah 401 is auth req, 403 is you got auth but it ain't no good 13:37:52 choice in the main, and write others in the alternative section 13:38:17 s/choice/choose one/ 13:38:36 yeah 13:38:50 thats fair, feels like a discussion in the alternatives section 13:39:24 #action alex_xu draft a new version of https://review.openstack.org/#/c/393205 13:39:47 oh, one clarification... 13:40:13 I think we agreed that all joined fields and anything that isn't using he same format as returned in the API, can be rejected without policy? 13:40:35 I think yes 13:40:53 joined fields means 'system_metadata', 'info_cache'...etc 13:41:04 the tags isn't in the list 13:41:04 yeah, stuff not in the instance table directly 13:41:15 tags are indexed OK, so that should be fine 13:41:29 well, tags are intentionally there, put it that way 13:41:44 what about metadata one, it works now, and metadata is the normal user faced data. 13:42:08 yeah, that one feels like we could stop it via policy for now, as its not very efficient 13:42:19 ok, got it 13:43:18 (1) stuff we block that may have worked or caused 500 (silently) (2) stuff we block with policy (silently) (3) stuff we allow for admins only by default (4) stuff we allow for users by default 13:43:28 I think thats the 4 camps they need splitting into 13:43:46 I guess there is (5) stuff not on any of those lists that is ignored silently 13:44:31 man this we managed to make this really complicated by taking so long to clear up this mess 13:45:13 emm...we block with policy and sliently ignore the invalid parameters. actually the policy equal to a config option 13:45:33 policy means it will return 403 13:46:01 it doesn't have to be, in this case its not going to trigger 403, it can in the new microversion 13:46:19 now in theory, all the stuff that causes 500s, could trigger 403, but that seems like an even worse mess 13:47:19 yea 13:47:28 what's an even worse mess? 13:47:53 having lists of all the things we should reject differently because of what they return today 13:48:02 oslo.db blocks a bunch of stuff that triggers 500s 13:48:26 its a recent-ish change we didn't notice happen 13:48:54 * cdent nods 13:49:16 It sort of read like you were saying 403s were messier than 500s, so I was confused 13:51:24 so...we have plan now 13:51:28 update the spec 13:51:45 cdent: ah, oops, true 13:52:12 alex_xu: sounds good, thank you! 13:52:25 johnthetubaguy: np 13:52:54 anything more? otherwise let us close the meeting 13:53:34 * cdent nods 13:54:08 ok, i think we back to work now 13:54:11 thanks all! 13:54:17 cdent: have a good rest 13:54:20 #endmeeting