TFJob with 1 replicas can't use gang-scheduling

[zionwu]

When I enabled gang-scheduling, I expect all the TF jobs to be scheduled by "kube-batch", so that all the jobs will have the same scheduling policy.

However, If I submit a TFJob with 1 replicas and specified the schedulerName to "kube-batch", The job stays pending. The cause is that TF-operator is not creating PDB if the replicas is less than 2 for the job:

func (jc *JobController) SyncPdb(job metav1.Object, minAvailableReplicas int32) (*v1beta1.PodDisruptionBudget, error) {
	labelJobName := jc.Controller.GetJobNameLabelKey()
	// Non-distributed training is not required gang scheduling
	if minAvailableReplicas < 2 {
		return nil, nil
	}
       .....

Can we remove this check to make the scheduling policy for all jobs consistent?

[gaocegege]

/cc @k82cn

I think we could remove it, while we are going to replace pdb with pod group. Maybe we could fix it after it.

[k82cn]

I think we could remove it, while we are going to replace pdb with pod group. Maybe we could fix it after it.

+1, kube-batch also support single pod in podgroup; it's safe to remove this check :)

[jlewi]

Anybody want to submit a fix?

[johnugeorge]

/assign