-
Notifications
You must be signed in to change notification settings - Fork 228
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Validate resources that share the same flavors in a ClusterQueue #308
Comments
/kind feature |
Why? This is too strict. I think this is similar to cluster autoscaler, when scaling up, we may have some kinds of nodes as the first choice, but we may still have fallback choices in case of insufficient of the preference ones. They maybe general nodes(compared to high-performance nodes), pay-as-you-go ones or preemptible instances. This kind of nodes can be shared by different node pools. In terms of flavors, it's same, we may have some kinds of nodes for general usage which referenced by a special flavor and meets the requirements of different resources. Some immature ideas, is it possible to add |
I think the idea wasn't clear. With the proposal, this should be possible (removing the quotas just to show the point): resources:
- name: "cpu"
flavors: ["spot", "ondemand"]
- name: "memory"
flavors: ["spot", "ondemand"]
- name: "example.com/gpu"
flavors: ["a100", "t4", "k80"] In this case, CPU and memory are coupled together. When a workload is assigned to this ClusterQueue, it would be given the same flavor for CPU and memory. The next workload could get a different flavor, but it needs to be the same for CPU and memory as well. Achieving this coupling would be difficult if we allowed something like this: - name: "cpu"
flavors: ["spot", "ondemand"]
- name: "memory"
flavors: ["spot", "default"] Note that the limitation is within one ClusterQueue. If there is another ClusterQueue in which CPU and memory don't need to be coupled, they can use different flavors.
See the motivation for ResourceFlavor here: #59 Basically putting the node labels in a separate object prevents multiple ClusterQueues from definining the same flavor but with different labels. |
Still think we should allow this configurations.
This is actually a fallback strategy, common in use. pseudo-code below:
|
That particular example doesn't make sense to me: "spot" and "on-demand" are the opposite, and thus they cover all the possibilities. Also, why wouldn't "cpu": ["spot", "default", "on-demand"] not make sense? |
/assign |
/priority important-soon |
This is a spinoff from #167 (comment), and a pre-requisite for #296
We need to validate that, in a ClusterQueue two resources either have completely different flavors or they must share all flavors, in the same order
The text was updated successfully, but these errors were encountered: