fix BestEffort for overused

pkg/scheduler/plugins/priority/priority.go

@@ -20,6 +20,8 @@ import (
 	"github.com/golang/glog"
 	"github.com/kubernetes-sigs/kube-batch/pkg/scheduler/api"
 	"github.com/kubernetes-sigs/kube-batch/pkg/scheduler/framework"
+
+	"k8s.io/api/core/v1"
 )
 
 type priorityPlugin struct {
@@ -37,6 +39,11 @@ func (pp *priorityPlugin) OnSessionOpen(ssn *framework.Session) {
 		lv := l.(*api.TaskInfo)
 		rv := r.(*api.TaskInfo)
 
+		// TODO quick fix, enhance it later by other option.
+		if res := compareQoS(lv.Pod, rv.Pod); res != 0 {
+			return res
+		}
+
 		glog.V(3).Infof("Priority TaskOrder: <%v/%v> prority is %v, <%v/%v> priority is %v",
 			lv.Namespace, lv.Name, lv.Priority, rv.Namespace, rv.Name, rv.Priority)
 
@@ -81,3 +88,28 @@ func (pp *priorityPlugin) OnSessionOpen(ssn *framework.Session) {
 }
 
 func (pp *priorityPlugin) OnSessionClose(ssn *framework.Session) {}
+
+// make BestEffort > Burstable/Guarantee
+func compareQoS(l, r *v1.Pod) int {
+	lq := l.Status.QOSClass
+	rq := r.Status.QOSClass
+
+	glog.V(3).Infof("Priority TaskOrder: <%v/%v> QoS is %v, <%v/%v> QoS is %v",
+		l.Namespace, l.Name, lq, r.Namespace, r.Name, rq)
+
+	if lq == rq {
+		return 0
+	}
+
+	// BestEffort > Burstable/Guarantee
+	if lq == v1.PodQOSBestEffort {
+		return -1
+	}
+
+	// Burstable/Guarantee < BestEffort
+	if rq == v1.PodQOSBestEffort {
+		return 1
+	}
+
+	return 0
+}

test/e2e/job.go

@@ -133,5 +133,18 @@ var _ = Describe("Job E2E Test", func() {
 		err = waitTasksReady(context, job3.Name, int(rep)/3)
 		Expect(err).NotTo(HaveOccurred())
 	})
+	It("Schedule BestEffort Job", func() {
+		context := initTestContext()
+		defer cleanupTestContext(context)
+
+		pgName := "test"
+
+		createPodGroup(context, "burstable", pgName, 2)
 
+		bestEffort := createJobWithoutPodGroup(context, "besteffort", 1, "busybox", nil, pgName)
+		burstable := createJobWithoutPodGroup(context, "burstable", 1, "busybox", oneCPU, pgName)
+
+		err := waitJobReadyWithPodGroup(context, pgName, bestEffort.Name, burstable.Name)
+		Expect(err).NotTo(HaveOccurred())
+	})
 })

test/e2e/util.go

@@ -620,3 +620,104 @@ func getPodOfJob(ctx *context, jobName string) []*v1.Pod {
 
 	return qjpod
 }
+
+func createPodGroup(
+	context *context,
+	jobName, name string,
+	min int32,
+) {
+	jns, _, _ := splictJobName(context, jobName)
+
+	pg := &arbv1.PodGroup{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      name,
+			Namespace: jns,
+		},
+		Spec: arbv1.PodGroupSpec{
+			MinMember: min,
+		},
+	}
+
+	pg, err := context.karclient.Scheduling().PodGroups(jns).Create(pg)
+	Expect(err).NotTo(HaveOccurred())
+}
+
+func createJobWithoutPodGroup(
+	context *context,
+	name string,
+	rep int32,
+	img string,
+	req v1.ResourceList,
+	pg string,
+) *batchv1.Job {
+	containers := createContainers(img, req, 0)
+	jns, jn, _ := splictJobName(context, name)
+
+	job := &batchv1.Job{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      jn,
+			Namespace: jns,
+		},
+		Spec: batchv1.JobSpec{
+			Parallelism: &rep,
+			Completions: &rep,
+			Template: v1.PodTemplateSpec{
+				ObjectMeta: metav1.ObjectMeta{
+					Annotations: map[string]string{arbv1.GroupNameAnnotationKey: pg},
+				},
+				Spec: v1.PodSpec{
+					SchedulerName: "kube-batch",
+					RestartPolicy: v1.RestartPolicyNever,
+					Containers:    containers,
+				},
+			},
+		},
+	}
+
+	job, err := context.kubeclient.BatchV1().Jobs(job.Namespace).Create(job)
+	Expect(err).NotTo(HaveOccurred())
+
+	return job
+}
+
+func waitJobReadyWithPodGroup(ctx *context, pgName string, jobNames ...string) error {
+	return wait.Poll(100*time.Millisecond, oneMinute, taskReadyWithPodGroup(ctx, -1, pgName, jobNames))
+}
+
+func taskReadyWithPodGroup(ctx *context, taskNum int, pgName string, jobNames []string) wait.ConditionFunc {
+
+	return func() (bool, error) {
+		pg := &arbv1.PodGroup{}
+
+		readyTaskNum := 0
+		for _, name := range jobNames {
+			jns, jn, _ := splictJobName(ctx, name)
+
+			queueJob, err := ctx.kubeclient.BatchV1().Jobs(jns).Get(jn, metav1.GetOptions{})
+			Expect(err).NotTo(HaveOccurred())
+
+			pods, err := ctx.kubeclient.CoreV1().Pods(jns).List(metav1.ListOptions{})
+			Expect(err).NotTo(HaveOccurred())
+
+			for _, pod := range pods.Items {
+				labelSelector := labels.SelectorFromSet(queueJob.Spec.Selector.MatchLabels)
+				if !labelSelector.Matches(labels.Set(pod.Labels)) ||
+					!metav1.IsControlledBy(&pod, queueJob) {
+					continue
+				}
+				if pod.Status.Phase == v1.PodRunning || pod.Status.Phase == v1.PodSucceeded {
+					readyTaskNum++
+				}
+			}
+
+			pg, err = ctx.karclient.Scheduling().PodGroups(jns).Get(pgName, metav1.GetOptions{})
+			Expect(err).NotTo(HaveOccurred())
+		}
+
+		if taskNum < 0 {
+			taskNum = int(pg.Spec.MinMember)
+		}
+
+		return taskNum <= readyTaskNum, nil
+	}
+}