| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2017-07-17 | |||
| 15:55:04 | mriedem | HostManager.get_host_states_by_uuids calls _get_host_states which returns a tuple | |
| 15:56:35 | mriedem | I think get_filtered_hosts is what converts to an iterator | |
| 15:56:41 | mriedem | hosts = six.itervalues(name_to_cls_map) | |
| 15:56:43 | bauzas | mriedem: the whole filters/weighers accepts an iterator | |
| 15:56:45 | bauzas | AFAIR | |
| 15:56:48 | mriedem | and passes that to filter_handler.get_filtered_objects | |
| 15:56:53 | bauzas | mriedem: yup | |
| 15:57:01 | mriedem | and ^ calls the filters | |
| 15:57:12 | bauzas | mriedem: but the thing is that we build a list once we get the weighed hosts | |
| 15:58:07 | mriedem | doesn't get_weighed_hosts return a sorted list? | |
| 15:59:36 | mriedem | ew now i see that WeighedObject is a wrapper around a HostState - wonder why someone used a wrapper rather than subclass HostState with a weight field? | |
| 15:59:45 | mriedem | WeighedHostState(HostState) | |
| 15:59:54 | mriedem | that's why we have that ugly host.obj.foo all over | |
| 16:00:37 | mriedem | anyway, i'm a bit lost in here wrt the iterator | |
| 16:01:13 | bauzas | yeah, seems like it's uglier than I thought | |
| 16:01:26 | mriedem | bauzas: the new _get_sorted_hosts method is still calling host_manager.get_filtered_hosts which still returns an iterator | |
| 16:01:44 | bauzas | mriedem: I was remembering we were using iterators, but looks like our filtering/weighing helpers work with lists... | |
| 16:01:46 | mriedem | and that iterator is still passed to host_manager.get_weighed_hosts | |
| 16:01:59 | mriedem | and weighed_hosts comes back as a sorted list | |
| 16:02:01 | mriedem | not an iterator | |
| 16:02:03 | mriedem | from what i can tell | |
| 16:02:07 | bauzas | yup | |
| 16:02:33 | mriedem | and by the time you get to chosen_host = hosts[0] it's really no different from before | |
| 16:02:38 | mriedem | so seems like something that doesn't matter to this change | |
| 16:02:42 | bauzas | I was just concerned by the new code taking more memory than previously, but it's just honestly not the case | |
| 16:02:48 | bauzas | yup | |
| 16:03:13 | bauzas | because we did work with iterators but once you enter filtering/weighting you're back to full lists of hosts anyway | |
| 16:03:42 | bauzas | anyway, I'm now looking at why that patch failed | |
| 16:05:23 | bauzas | mmmmpf http://logs.openstack.org/64/483564/2/gate/gate-nova-tox-functional-ubuntu-xenial/0f4534d/console.html#_2017-07-17_15_20_35_769898 | |
| 16:09:25 | mriedem | shouldn't that also be running through other filters? | |
| 16:09:35 | jaypipes | mriedem, bauzas: that whole comment about it being an iterator is nonsense. As soon as it enters the filter host manager, it converts it all to a list: https://github.com/openstack/nova/blob/master/nova/filters.py#L69 | |
| 16:09:39 | mriedem | like ServerGroupAntiAffinityFilter | |
| 16:09:58 | mordred | mriedem, sdague: if a client sends the legacy microversion header to a new nova, that's not an error, right? and if it sends both, that's also acceptable? | |
| 16:09:59 | jaypipes | mriedem: it was always trying to be too clever for its own good, IMHO. | |
| 16:10:07 | bauzas | jaypipes: yup, just saw that, see above the IRC discussion with matt | |
| 16:10:28 | bauzas | jaypipes: honestly, I'm not surprised | |
| 16:10:30 | sdague | mordred: correct, either are accepted | |
| 16:10:38 | jaypipes | mriedem: btw, I was trying to get rid of the host.obj foo nonsense. | |
| 16:10:48 | sdague | as are both, I think when both we take the standard one as gospel | |
| 16:10:58 | mordred | sdague: awesome. I believe I have a general solution to our earlier problem | |
| 16:11:06 | bauzas | jaypipes: someone probably did care of passing iterators, but someone else probably messed it up by just using lists afterwards | |
| 16:11:08 | sdague | cdent: I think you wrote the acceptor code, right? | |
| 16:11:16 | mriedem | jaypipes: yeah i noted that in the commit message | |
| 16:11:22 | cdent | sdague: yeah, it’s happy with both, prefers the newer one | |
| 16:11:22 | mriedem | less turducken is good | |
| 16:11:29 | mordred | cdent: sweet | |
| 16:11:40 | bauzas | jaypipes: anyway, it's not super critical, but something is : your patch is getting -1 because http://logs.openstack.org/64/483564/2/gate/gate-nova-tox-functional-ubuntu-xenial/0f4534d/console.html#_2017-07-17_15_20_35_769898 | |
| 16:11:44 | cdent | it always sends both back in the response too | |
| 16:12:35 | jaypipes | bauzas: I can't reproduce that locally :( | |
| 16:12:57 | openstackgerrit | zhang zhenzhong proposed openstack/nova master: fix rp registration failed while placement endpoint created after compute start https://review.openstack.org/483460 | |
| 16:13:15 | bauzas | jaypipes: I was just about to run it :) | |
| 16:13:38 | mriedem | http://logstash.openstack.org/#dashboard/file/logstash.json?query=message%3A%5C%22%2C%20in%20test_boot_servers_with_soft_anti_affinity_one_available_host%5C%22%20AND%20tags%3A%5C%22console%5C%22&from=7d | |
| 16:13:54 | bauzas | my child at home is killing my work because of netflix eating all the poor internet bandwidth FWIW | |
| 16:14:05 | bauzas | plex FTW | |
| 16:14:45 | bauzas | git pull is taking more than 1 min | |
| 16:14:53 | mriedem | fwiw, that test also failed in https://review.openstack.org/#/c/482288/ | |
| 16:14:57 | mriedem | which would be totally unrelated | |
| 16:15:43 | bauzas | jaypipes: did you run the whole servergroup functional test group, or just the one failing? | |
| 16:15:51 | jaypipes | bauzas: both | |
| 16:15:53 | bauzas | k | |
| 16:16:14 | bauzas | last time I had such problem, it was because of the fixture | |
| 16:16:30 | bauzas | but I was able to reproduce it | |
| 16:19:39 | mriedem | so the test creates a soft-anti-affinity server group and kills the 2nd compute service so there should only be 1 | |
| 16:19:51 | mriedem | since the 'soft' means it will go to the same host if there is only 1 available | |
| 16:19:56 | sdague | mriedem: you have an opinion on this - https://review.openstack.org/#/c/215385/21/nova/virt/ironic/client_wrapper.py - I thought we wanted N-1 compat | |
| 16:21:09 | mriedem | sdague: that's basically how it's always worked with nova + ironic as far as i know; yes it's not what we do with anything else | |
| 16:21:28 | sdague | mriedem: ok, well we definitely need upgrade stanza in the release notes then | |
| 16:22:06 | mriedem | i raised a similar concern in another nova ironic change, | |
| 16:22:36 | mriedem | https://review.openstack.org/#/c/419975/11/releasenotes/notes/add-interface-attach-detach-support-in-ironic-cb2bf11f3875350a.yaml@4 | |
| 16:23:01 | mriedem | basically they implemented something in a driver and then started claiming support for hot plugging interfaces in nova | |
| 16:23:25 | mriedem | my argument was, what if my ironic is from ocata and i upgrade nova, and was told, that doesn't happen because everyone knows to upgrade ironic first | |
| 16:23:56 | sdague | mriedem: so... I'm extremely suspicious of "everybody knows" | |
| 16:24:10 | sdague | I'm fine with it being policy, but we still need upgrade notes | |
| 16:24:41 | mriedem | yeah they added one later here https://review.openstack.org/#/c/419975/17/releasenotes/notes/add-interface-attach-detach-support-in-ironic-cb2bf11f3875350a.yaml | |
| 16:25:16 | sdague | I'd be more happy with an N-1 policy, but we do need to communicate it every time it would trip people up | |
| 16:26:55 | mriedem | i don't disagree | |
| 16:29:56 | mriedem | jaypipes: bauzas: i have an idea on that failure | |
| 16:29:59 | mriedem | the functional test failure that is | |
| 16:30:07 | jaypipes | yeah? | |
| 16:30:09 | bauzas | ? | |
| 16:30:18 | mriedem | the test setup creates 2 compute services/nodes/hosts | |
| 16:30:25 | mriedem | you can see the 2nd host CN created here http://logs.openstack.org/64/483564/2/gate/gate-nova-tox-functional-ubuntu-xenial/0f4534d/console.html#_2017-07-17_15_20_35_768000 | |
| 16:30:31 | mriedem | with uuid a945d18d-e6b6-4fd5-8b3b-e50c973a1ff9 | |
| 16:30:35 | mriedem | and it's posted into placement | |
| 16:30:57 | mriedem | now the purpose of the test is to restrict to 1 host | |
| 16:30:58 | mriedem | so | |
| 16:30:59 | mriedem | self.compute2.kill() | |
| 16:31:02 | mriedem | HOWEVER | |
| 16:31:13 | mriedem | filter scheduler is getting it's hosties from placement now | |
| 16:31:19 | mriedem | so that RP is still floating around in placement | |
| 16:31:55 | mriedem | you can see the scheduler do it's thing here | |
| 16:31:55 | mriedem | GET /allocation_candidates?resources=DISK_GB%3A20%2CMEMORY_MB%3A2048%2CVCPU%3A1 HTTP/1.1" status: 200 len: 992 time: 0.0521650 | |
| 16:32:02 | mriedem | and then | |
| 16:32:02 | mriedem | No compute service record found for host host2 | |
| 16:32:10 | mriedem | because the test killed it | |
| 16:32:15 | mriedem | Filter RetryFilter returned 0 hosts | |
| 16:32:29 | mriedem | what i don't understand is why it doesn't try the other compute host | |
| 16:34:38 | mriedem | so one easy hack is deleting the host2 RP from test_boot_servers_with_soft_anti_affinity_one_available_host - but that's pretty hacky, it seems odd that the filtering wouldn't check the other host | |
| 16:35:34 | jaypipes | mriedem: it's also odd that this test fails randomly. | |
| 16:39:11 | mriedem | yeah so we get this https://github.com/openstack/nova/blob/master/nova/scheduler/host_manager.py#L674 | |