| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2019-12-11 | |||
| 21:59:42 | efried | I mean, we enforce that for placement | |
| 21:59:45 | dansmith | you mean like an RPM or DEB dependency? | |
| 22:00:01 | efried | no, I mean like cyborg agent refuses to start up if compute isn't at version xyz | |
| 22:00:07 | dansmith | efried: how would it know? | |
| 22:00:30 | Sundar | dansmith: An old Nova would not query CYborg for device profiles or create/bind ARQs | |
| 22:00:39 | dansmith | Sundar: a new nova control plane would | |
| 22:00:50 | dansmith | Sundar: nova supports backlevel nova-compute services, which is the scenario I'm talking about | |
| 22:01:28 | Sundar | dansmith: You mean new n-api, n-super-cond, n-sch but old n-cpu? | |
| 22:01:42 | efried | ...do we seriously not have a way to query the nova version on a compute? | |
| 22:01:50 | dansmith | Sundar: yes, a very important scenario we have supported for a long time that people depend on | |
| 22:01:53 | efried | I mean, obv the RPC API does, which is what you're leading up to. | |
| 22:02:08 | dansmith | efried: we don't and shouldn't | |
| 22:03:43 | dansmith | I guess I better go ahead and drop that in the review somewhere before I go poof for the year and ya'll try to merge this :) | |
| 22:03:58 | Sundar | dansmith: I may be missing something. We had the create/bind and the wait all in n-cpu, which would have avoided this issue (at the expense of less concurrency), right? | |
| 22:04:08 | efried | Sundar: no | |
| 22:04:18 | dansmith | Sundar: no, you still have the same problem with that arrangement | |
| 22:04:21 | efried | So, ugh, what we actually need is something like: | |
| 22:04:21 | efried | compute advertises a COMPUTE_CAN_DO_CYBORG_SHIT trait so that | |
| 22:04:21 | efried | cyborg (which already queries that RP to know where to hang the accel RPs) can know whether it's even allowed to do that. | |
| 22:04:43 | dansmith | efried: or do it with service version but you might as well do an RPC version in that case | |
| 22:04:47 | efried | Sundar: The create/bind happens at deploy time, when you've already picked a host. We're talking about the bootstrapping process where cyborg decides to advertise accel resources in the first place. | |
| 22:04:55 | dansmith | efried: the trait is more expensive unless you've already got them | |
| 22:05:18 | efried | dansmith: expensive how? Because an extra placement query, once, on startup? Not worried about that. | |
| 22:05:23 | efried | or were you talking about something else? | |
| 22:05:37 | dansmith | something else, and I misunderstood what you mean so let me 'splain: | |
| 22:06:08 | dansmith | efried: you could either try to avoid cyborg hanging the inventory off the compute at bootstrap time, or you could avoid letting the conductor/scheduler start down this path for a compute that can't handle it | |
| 22:06:27 | efried | the former gets you the latter for free | |
| 22:06:31 | dansmith | IMHO, the latter is the job of nova anyway, | |
| 22:06:32 | dansmith | and would fit with a "this compute isn't new enough to do that thing you want" sort of check | |
| 22:06:50 | dansmith | efried: it depends on another service for upgrade correctness though | |
| 22:06:57 | efried | if you expose the inventory but have something else that prevents you from using it, that's a waste, isn't it? | |
| 22:07:10 | dansmith | i.e. it depends on cyborg being well-behaved, and/or not being modified to just always do it | |
| 22:07:29 | dansmith | efried: well, it means that cyborg can be upgraded and working before you've upgraded all your computes | |
| 22:07:37 | efried | heck, on that theory we could hang the inventory off a *really* old compute and it mushroom clouds because it can't handle nested providers at all. | |
| 22:07:38 | dansmith | so if you upgrade service by service, you don't have to loop back around | |
| 22:07:49 | dansmith | which was always annoying with things like ironic that had kinda circular deps with nova | |
| 22:08:00 | efried | cyborg is polling though, isn't it Sundar? | |
| 22:08:05 | dansmith | efried: well, that might be a good reason to do both then I guess, I dunno | |
| 22:08:42 | Sundar | efried: Sorry.Cyborg is polling for what? | |
| 22:08:42 | efried | Heh, mushroom cloud. That should be a thing. | |
| 22:09:04 | Sundar | Polling for new devices? Yes | |
| 22:09:45 | dansmith | efried: I understand that we avoid asking for the new thing in a roundabout way by not exposing a trait, cyborg not exposing inventory, placement not returning any candidates, and thus us not actually running our control plane code, | |
| 22:09:46 | efried | Sundar: If cyborg agent is shiny and new on a host, but the nova-compute on the host is downlevel | |
| 22:09:47 | efried | and then you upgrade the nova-compute | |
| 22:09:47 | efried | will cyborg eventually figure that out and realize it should now start reporting accel inventory on that host? | |
| 22:10:03 | dansmith | efried: but it would be a lot more in line with our regular checks to just be explicit about it, even if we should never hit it because of the other | |
| 22:10:21 | efried | dansmith: okay, so what is it that you're suggesting exactly? I didn't pick up on that. | |
| 22:10:50 | efried | An RPC version check from conductor to compute? | |
| 22:10:51 | openstackgerrit | Mykola Yakovliev proposed openstack/nova master: Fix boot_roles in InstanceSystemMetadata https://review.opendev.org/698040 | |
| 22:11:04 | dansmith | efried: if you pass the ARQs down in the spawn call, you need an RPC version, so you get a service version, which is something you can just check for easily in conductor, alternately just bump the service version without the rpc and check for that in conductor | |
| 22:11:23 | dansmith | efried: or expose it as a trait and have conductor/scheduler filter or check for that trait before agreeing to use that compute | |
| 22:11:55 | dansmith | efried: perhaps the easiest thing would be just "if we have accel resources, also add CAN_DO_CYBORG_SHIT trait into the requirements" before we call placement during scheduling? | |
| 22:12:05 | dansmith | you'll need to do it again for attach when we add that | |
| 22:12:14 | dansmith | so CAN_DO_CYBORG_SHIT_V2alpha | |
| 22:13:02 | efried | dansmith: I'm not offended by the thought of those capability traits, which seem on par with what we have for e.g. "can do multiattach" or whatever | |
| 22:13:09 | efried | so IMO we should do those regardless | |
| 22:13:12 | dansmith | ack | |
| 22:13:17 | dansmith | so cyborg can look for that, | |
| 22:13:21 | efried | those are simply flags on that virt driver dict | |
| 22:13:25 | dansmith | and we throw that into the scheduling request also if we're doing accel stuff? | |
| 22:13:49 | efried | Well, I don't think we need it in the sched request, if cyborg is using it to decide whether to present the inventory in the first place. | |
| 22:14:09 | efried | but if you think it should be there for form's sake, I wouldn't object. | |
| 22:14:11 | dansmith | again, that puts the onus on cyborg to be responsible, which I don't like | |
| 22:14:19 | efried | where's the trust?? | |
| 22:14:29 | Sundar | So, the use case for all this is that, the operator added accelerator to a compute node but did not upgrade n-cpu on that node? | |
| 22:14:31 | dansmith | right, thank you :) | |
| 22:14:35 | dansmith | efried: let me check my butt, hang on | |
| 22:14:57 | efried | That's not spelled "onus" dansmith | |
| 22:15:06 | dansmith | hehe | |
| 22:15:08 | efried | Sundar: yes, exactly. | |
| 22:15:37 | efried | Sundar: we're trying to make nova account for poorly-behaved a) operators, b) service code from other services. | |
| 22:15:57 | efried | and not-work in a predictable rather than unpredictable (and potentially unrecoverable) way. | |
| 22:16:11 | dansmith | just to be clear, | |
| 22:16:33 | dansmith | I definitely think that the rpc method change is the right thing to do, instead of just coding in new assumptions on both sides | |
| 22:16:57 | dansmith | I understand the hesitation to that, so I won't hold strictly to it, despite it being the safer option, IMHO | |
| 22:17:28 | efried | dansmith: if we did that, we would have to have a post-placement filter to remove hosts based on RPC version, right? | |
| 22:17:35 | efried | perhaps we already have that. | |
| 22:17:47 | Sundar | What's wrong with the option where Nova places a CAN_DO_CYBORG trait on some compute node RPs, and factors that into Placement query? | |
| 22:18:14 | efried | Sundar: nothing's wrong with that, and we should do it, but we're talking about doing another thing as well. | |
| 22:18:22 | dansmith | efried: no, you'd still want to pick a capable host during scheduling, and scheduling based on rpc version is not a thing | |
| 22:18:39 | efried | dansmith: so what's the RPC version for again? Just passing the arq uuids? | |
| 22:19:37 | dansmith | efried: yes, and to crystalize our request to the compute node so that (a) the RPC layer can tell if we're (for whatever reason) asking the compute node to do something it can't handle, and so that compute doesn't bake the assumption of the instance having a magical key in its flavor into meaning that it should call to cyborg to do all this stuff | |
| 22:20:13 | dansmith | efried: because sometime in the future, the mere presence of that flavor key will imply a slightly different thing, and we will have an upgrade concern to deal with because we have a bunch of old computes that will still be doing the assume-it thing | |
| 22:20:41 | dansmith | I made reference to this in a previous discussion about all of this | |
| 22:22:16 | efried | This is getting into details I probably don't care about rn, but do we get (a) for free (does the call bounce if the compute is too old) or do we have to do an explicit check? | |
| 22:22:37 | dansmith | summarized the bits you do care about here: https://review.opendev.org/#/c/631244/51/nova/conductor/manager.py | |
| 22:23:30 | efried | ack | |
| 22:23:55 | dansmith | efried: the rpc layer knows what the minimum supported version is across the cluster, so when it is converting the request, it can explode if doing so would drop an arqs parameter | |
| 22:24:14 | dansmith | efried: it gets you the circuit breaker, not necessarily the "graceful" part -- that should still be scheduler-based | |
| 22:24:31 | efried | cool | |
| 22:26:28 | dansmith | Sundar: while you're here, what's the plan for getting some small example of this running with an actual pci device in the intel pci ci system (or something) ? | |
| 22:27:53 | Sundar | dansmith: We have the server and the FPGA in a lab, and one person is working on that. We expect it by Jan, if not this year. | |
| 22:28:52 | Sundar | FWIW, I check each patch set with real FPGAs, with different device profiles, and also the fake driver. | |
| 22:29:49 | efried | We should have a 3rd party CI called "Intel Cyborg Sundar Manual CI" | |
| 22:30:08 | dansmith | Sundar: that would be good except that I totally don't trust you as much as a computer | |
| 22:30:10 | dansmith | :) | |
| 22:30:10 | Sundar | We already have that hehe | |
| 22:30:35 | dansmith | Sundar: so one server and one fpga and one person.. what is the plan? to have that be a manual "recheck use fpgas" command or something? | |
| 22:30:41 | Sundar | dansmith: Too bad, I am more reliable than other past Intel CIs ;) | |
| 22:30:56 | dansmith | Sundar: that's true, although doesn't really change my feelings about it :) | |