You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What would you like to be cleaned:
It is possible that we assign mode Preempt, even when preemption is never possible. E.g. imagine the case that the CQ never allows preemption, and yet the workload's requests are below nominal quota.
However, the fact that we chose Preempt and not NoFit is later used to reserve resources that belong to the CQ - see #4105.
Nevertheless, this is confusing. Something like the following categories may make more sense:
NeverFit - current NoFit, where workload is too big even if cluster is empty
NoFit - workload can fit if no workloads are running, but it has to wait, as preemption is not an option
Preemption
Fit
The fact that NeverFit workloads are handled as NoFit might be even considered a bug currently, because such a workload blocks indefinitely other "valid" workloads without a clear response saying "why".
What would you like to be cleaned:
It is possible that we assign mode
Preempt
, even when preemption is never possible. E.g. imagine the case that the CQ never allows preemption, and yet the workload's requests are below nominal quota.kueue/pkg/scheduler/flavorassigner/flavorassigner.go
Lines 622 to 631 in 14e2ca7
However, the fact that we chose
Preempt
and notNoFit
is later used to reserve resources that belong to the CQ - see #4105.Nevertheless, this is confusing. Something like the following categories may make more sense:
/cc @mimowo
The text was updated successfully, but these errors were encountered: