RFC: Add a new option --scheduler-name for ignoring pods which should be h… #170
Conversation
k8s-ci-robot
added
cncf-cla: yes
|
xref #164 |
pkg/batchd/cache/cache.go
Outdated
| @@ -88,6 +89,9 @@ func newSchedulerCache(config *rest.Config) *SchedulerCache { | |||
| FilterFunc: func(obj interface{}) bool { | |||
| switch t := obj.(type) { | |||
| case *v1.Pod: | |||
| if strings.Compare(obj.(*v1.Pod).Spec.SchedulerName, schedulerName) != 0 { | |||
There was a problem hiding this comment.
Only filter Pods with the same scheduler name may be not enough. It may cause kube-batchd schedule pods to a node overload.
Such as some pods from another scheduler are running on a node, however, the node is idle for kube-batchd due to kube-batchd doesn't get those running pods info. When kube-batchd try to schedule pods on that node, it may fail due to lack of resources.
There was a problem hiding this comment.
+1
We should watch 'pending & schedulerName' or 'running' pods.
pkg/batchd/cache/cache.go
Outdated
| @@ -88,6 +89,10 @@ func newSchedulerCache(config *rest.Config) *SchedulerCache { | |||
| FilterFunc: func(obj interface{}) bool { | |||
| switch t := obj.(type) { | |||
| case *v1.Pod: | |||
| pod := obj.(*v1.Pod) | |||
| if strings.Compare(pod.Spec.SchedulerName, schedulerName) == 0 && pod.Status.Phase == v1.PodPending { | |||
There was a problem hiding this comment.
it also adds pending pods from other scheduler(nonTerminatedPod() will return true)
There was a problem hiding this comment.
ah I see, will fix it soon, thanks
|
@jinzhejz updated the condition, could you take a look? |
|
@mitake there is build error, could you take a look And other looks good to me. |
|
@jinzhejz oops... really sorry for that, I'll fix it soon. |
…andled by the default scheduler This commit adds a new option --scheduler-name to kube-batchd. kube-batchd handles pods which have the name specified with the option in its v1.Pod.Spec.SchedulerName. The motivation is separating pods which should be controlled by kube-batchd from other pods which can be handled by the default scheduler.
k8s-ci-robot
added
size/M
|
@jinzhejz fixed the error, could you take a look? |
|
thanks, team; merged :) |
RFC: Add a new option --scheduler-name for ignoring pods which should be h…
RFC: Add a new option --scheduler-name for ignoring pods which should be h…
…andled by the default scheduler
This commit adds a new option --scheduler-name to
kube-batchd. kube-batchd handles pods which have the name specified
with the option in its v1.Pod.Spec.SchedulerName. The motivation is
separating pods which should be controlled by kube-batchd from other
pods which can be handled by the default scheduler.
I'm still not fully sure this change is valuable for kube-arbitrator's design. If I can have comments, I'm really glad.