Image Metadata Properties weigher
This blueprint proposes adding a new weigher that will pack/spread instances based on th image properties requested by the instance and the image properties of existing instances on individual hosts.
A typical usecase is for Windows guests. Given the bill would be huge if we spread all the Windows instances into multiple nodes, we usually say that the admin should use aggregates and the AggregateImageP
That said, the concern is that in general admins don't really know how many nodes they should be using for the Windows instances, and if they provide more than it needs, these nodes wouldn't be used.
That's why, instead of using this filter and asking operators to create aggregates, we would rather like to create a specific ImageMetadataPr
Blueprint information
- Status:
- Complete
- Approver:
- Sylvain Bauza
- Priority:
- Undefined
- Drafter:
- Sylvain Bauza
- Direction:
- Approved
- Assignee:
- Sylvain Bauza
- Definition:
- Approved
- Series goal:
- Accepted for 2025.1
- Implementation:
-
Implemented
- Milestone target:
- None
- Started by
- Sylvain Bauza
- Completed by
- Sylvain Bauza
Related branches
Related bugs
Sprints
Whiteboard
sean(2025-01-06): i would suggest that we implement this by looking at the set of all distinct image properties currently in use by instances on a host + the distinct requested image properties "total_prop" and then take the intersection of the request props and the current properties "common_props" and define the return value of the weiehr as "common_props / total_props"
total_props is just the the union of all image properties for all instance on the current host "existing_props" + the requested properties fo the current instnace form the request spec.
common_props is just the intersection of the the props form the request spec with existing_props.
that pretty simple to compute form a perfroamce point fo view as we should already have all that info in memory in the host_state object + request spec so i would expect no additional db look ups.
so while this wont be the cheapest weigher i don't expect it to be excessively expensive.
if we default the multiplier to 0.0 so that it has no effect by default it would have no upgrade impact and operators can use a negitive/positive multiplier to opt in
common_props divided by total naturally packs.
if we agree this ^ design in general i think we could proceed with this as a specless blueprint.
(gibi)[2025-01-07]: "total_props is just the the union of all image properties for all instance on the current host" Is this something that we need to now collect and put it into the host state?
[20250107 bauzas] Approved as specless during today's meeting
Gerrit topic: https:/
Addressed by: https:/
Add a new ImageProperties
Addressed by: https:/
per-property ImageMetaPropsW
Addressed by: https:/
Add fill_metadata() to InstanceList
