Earlier  
Posted Nick Remark
#openstack-nova - 2020-09-11
07:42:26 gibi and good Friday morning to you too
07:42:58 bauzas gibi: and I -w'd the prelude because of the xenapi change still in the gate + this one that I think which needs to be in the prelude
07:43:30 gibi yeah, I kikked the xenapi series this morning with a recheck
07:43:35 gibi *kicked
07:43:59 bauzas I saw
08:23:52 sean-k-mooney o/
08:25:37 sean-k-mooney gibi: so i would still like to procee with the deprecation of the az filter and enableing the placment version by default in Victoria since its behavior is equivalent, but i think i am going to split out the Comptefilter changes and look to adress that in Wallaby when we have more time to think about it
08:27:08 sean-k-mooney i dont know if we want to reuse the COMPUTE_STATUS_DISABLED trait for force down or not and i dont really want ot rush figuring that out
08:28:15 sean-k-mooney i think having a sperate COMPUTE_STATUS_DOWN trait is proably more correct but that would need an os-traits release and i dont think this warrents a FFE for os-triats since its just tech debt removal
08:28:19 gibi sean-k-mooney: make sense, keep the ComputeFilter then for V and please open a bug to track the problem you detected with the pre-filter
08:29:17 sean-k-mooney yep i can do that that was actully going to be my question should i open a bug or blueprint. e.g. is this an oversight in the original filter that it ignore force down or a new feature
08:31:13 gibi i consider this as a bug
08:35:42 sean-k-mooney cool ill file one so then respin the patch to drop the computefitler changes
08:36:12 gibi thanks
08:38:46 bauzas sean-k-mooney: wait, what ?
08:38:56 bauzas sean-k-mooney: which filter are you going to deprecate ?
08:39:05 bauzas AZFilter or ComputeFilter ?
08:39:13 sean-k-mooney i was deprecating both
08:39:22 sean-k-mooney but the behavior of the ComputeFilter
08:39:31 sean-k-mooney is not identical to the placment prefilter for compute status
08:39:48 sean-k-mooney because the prefilter does not take accound of force_down
08:39:50 bauzas ok, now I better understand and yes
08:39:59 bauzas I was about to say this
08:40:32 bauzas but... are we considering the AZfilter deprecation for the Victoria cycle ?
08:40:35 sean-k-mooney so for now im just going to depercate teh az filter then file a bug for the gap in behavior and we can figure out the correct way to handel force_down in wallaby
08:40:38 bauzas I'd say it's a Wallaby thing
08:40:46 sean-k-mooney bauzas: yes i was
08:41:11 sean-k-mooney https://review.opendev.org/#/c/745605/
08:41:34 sean-k-mooney i have had that open for a while but only dug into the fun test fialure properly yesterday
08:41:48 bauzas ok, I feel we need to talk then
08:42:04 bauzas because I agree with the fact that the AZfilter is on par with the pre-filter
08:42:49 sean-k-mooney i could delay this if you want by the way but if you dont have an object ot the AZfilter then i would prefer to proceed with just that
08:42:50 bauzas but I know ops use filters by ordering them like they want
08:43:01 sean-k-mooney yes
08:43:10 sean-k-mooney but the AZ filter always goes first right
08:43:29 bauzas no
08:43:39 sean-k-mooney no?
08:43:41 bauzas by default, ComputeFilter goes first
08:43:54 bauzas and then AZ
08:44:08 bauzas if we deprecate the AZ, this will change
08:44:39 bauzas oh shit
08:44:41 bauzas I'm wrong
08:44:51 openstackgerrit Dat Le proposed openstack/nova stable/rocky: Fix unplugging VIF when migrate/resize/rebuild VM between hosts have difference VIF type https://review.opendev.org/751186
08:44:52 sean-k-mooney ya i was going to say https://review.opendev.org/#/c/745605/3/nova/conf/scheduler.py
08:44:57 bauzas that's... disturbing
08:45:13 sean-k-mooney it makes sense from an information gain point of view
08:45:14 bauzas I'm almost sure this wasn't the case years ago
08:45:27 bauzas but anyway
08:45:33 bauzas in this case, this sounds legit
08:45:48 bauzas sean-k-mooney: if so, I'll review your deprecation patch
08:46:06 sean-k-mooney the point is in anycase to move both operation eventually to the placment query
08:46:25 sean-k-mooney since in that case ordering does not matter
08:47:01 sean-k-mooney at least form teh point of view of the query construction
08:48:19 sean-k-mooney bauzas: ill ping you with the update patch with just the AZfilter deprecation in a few once its ready
08:48:34 openstackgerrit Balazs Gibizer proposed openstack/nova master: Spell out 3pp in vmware undeprecation reno https://review.opendev.org/751188
08:48:53 sean-k-mooney whats 3pp
08:49:06 sean-k-mooney oh hehe
08:49:13 sean-k-mooney third party
08:49:28 sean-k-mooney i see that was bauzas nit
08:51:39 bauzas sean-k-mooney: also, I left a comment on your deprecation patch, having the sheduler being underoptimized is not a performance issue
08:55:34 sean-k-mooney bauzas: thanks that said the orginal doc state that enabling the prefilter improves performance, so when i flip the docs to state its enabled by default now does it not make sense to now also state that reverting to the az filter woudl reduce performance
08:56:05 sean-k-mooney bauzas: i can remove that i just basically inverted the previous statement but if you dont think it adds value ill just drop it
08:56:25 sean-k-mooney bauzas: you were refering to your comment here right https://review.opendev.org/#/c/745605/3/doc/source/admin/availability-zones.rsthttps://review.opendev.org/#/c/745605/3/doc/source/admin/availability-zones.rst
08:57:54 sean-k-mooney the funny thing is that change was orginally just ment to be the az filter
08:58:39 sean-k-mooney i only added the compute filter when i was doing the doc update and was like waith there is a prefilter that "does the same thing" that can go too. i was wrong :(
09:13:31 openstackgerrit Federico Ressi proposed openstack/nova master: Log stack trace when fails launching a defined domain https://review.opendev.org/750652
09:13:47 openstackgerrit Federico Ressi proposed openstack/nova master: Log stack trace when fails launching a defined domain https://review.opendev.org/750652
09:14:58 bauzas sean-k-mooney: I know, we have a lot of docs that are wrong by saying "heh, prefilters are better for performance"
09:15:09 bauzas tbc: NO
09:15:21 bauzas a prefilter just allows you to get more candidates
09:15:53 sean-k-mooney bauzas: that is not true
09:16:13 bauzas filters perfomance is not an issue
09:16:15 sean-k-mooney bauzas: if you enable a prefilter and disable a schduler filter the db outperform python
09:16:25 bauzas sean-k-mooney: again, it's wrong
09:16:26 sean-k-mooney bauzas: in general yes
09:16:41 bauzas lots of folks (eg. CERN) looked at the scheduler performance
09:16:46 sean-k-mooney bauzas: we have see this form operators reporting that its faster
09:16:53 bauzas and none of them said that the filters were outperfomed
09:17:10 bauzas the problem was with the DB
09:17:13 sean-k-mooney i acutlly tought cern said it did
09:17:13 bauzas not the filters
09:17:20 sean-k-mooney i guess it woudl only help
09:17:35 sean-k-mooney as you said if we dont return the 1000 limit on allocitons
09:17:47 sean-k-mooney e.g. if it reduces the set below that
09:17:56 bauzas no, the problem is that NOBODY thought about the performance for getting more than 1000 candidates
09:18:04 bauzas I mean the API performance
09:18:17 bauzas so we said, "ok let's limit to 1000 candidates"
09:18:39 bauzas and then we said "oh but then shit, we won't then verify all the candidates"
09:18:53 bauzas and then that's why we started using prefilters
09:19:02 sean-k-mooney ya i know that i was there
09:19:14 sean-k-mooney and NOBODY is a strech
09:19:39 bauzas if the Placement API was able to provide *all* candidates, the performance in between checking all of them with a filter, or instead querying the Placement *before* wouldn't be differnet
09:19:40 sean-k-mooney we did discuss the subseting problem before we started using placment
09:20:02 sean-k-mooney bauzas: the different is the number of candiates pass to the schduler
09:20:17 bauzas meh, the ship has sailed either way
09:20:29 sean-k-mooney if we have 1000 nodes in 2 az of 500
09:20:36 bauzas again, not a problem
09:20:38 sean-k-mooney the az prefilter will limit the repsoce to at most 500
09:21:05 sean-k-mooney without it all 1000 nodes asumming all other thing are equall would pass to the scudler
09:21:28 sean-k-mooney and the az filter would have to filter the 500 nodes then

Earlier   Later