Fix back off when scheduling cycle is delayed

Signed-off-by: Aldo Culquicondor <acondor@google.com>

pkg/scheduler/internal/queue/pod_backoff.go

@@ -21,12 +21,14 @@ import (
 	"time"
 
 	ktypes "k8s.io/apimachinery/pkg/types"
+	"k8s.io/kubernetes/pkg/scheduler/util"
 )
 
 // PodBackoffMap is a structure that stores backoff related information for pods
 type PodBackoffMap struct {
 	// lock for performing actions on this PodBackoffMap
-	lock sync.RWMutex
+	lock  sync.RWMutex
+	clock util.Clock
 	// initial backoff duration
 	initialDuration time.Duration
 	// maximal backoff duration
@@ -38,8 +40,9 @@ type PodBackoffMap struct {
 }
 
 // NewPodBackoffMap creates a PodBackoffMap with initial duration and max duration.
-func NewPodBackoffMap(initialDuration, maxDuration time.Duration) *PodBackoffMap {
+func NewPodBackoffMap(initialDuration, maxDuration time.Duration, clock util.Clock) *PodBackoffMap {
 	return &PodBackoffMap{
+		clock:             clock,
 		initialDuration:   initialDuration,
 		maxDuration:       maxDuration,
 		podAttempts:       make(map[ktypes.NamespacedName]int),
@@ -91,12 +94,16 @@ func (pbm *PodBackoffMap) ClearPodBackoff(nsPod ktypes.NamespacedName) {
 
 // CleanupPodsCompletesBackingoff execute garbage collection on the pod backoff,
 // i.e, it will remove a pod from the PodBackoffMap if
-// lastUpdateTime + maxBackoffDuration is before the current timestamp
+// lastUpdateTime + maxDuration >> timestamp
+// We should wait longer than the maxDuration so that the pod gets a chance to
+// (1) move to the active queue and (2) get an schedule attempt.
 func (pbm *PodBackoffMap) CleanupPodsCompletesBackingoff() {
 	pbm.lock.Lock()
 	defer pbm.lock.Unlock()
 	for pod, value := range pbm.podLastUpdateTime {
-		if value.Add(pbm.maxDuration).Before(time.Now()) {
+		// Here we assume that maxDuration should be enough for a pod to move up the
+		// active queue and get an schedule attempt.
+		if value.Add(2 * pbm.maxDuration).Before(pbm.clock.Now()) {
 			pbm.clearPodBackoff(pod)
 		}
 	}
@@ -106,7 +113,7 @@ func (pbm *PodBackoffMap) CleanupPodsCompletesBackingoff() {
 // and increases its numberOfAttempts by 1
 func (pbm *PodBackoffMap) BackoffPod(nsPod ktypes.NamespacedName) {
 	pbm.lock.Lock()
-	pbm.podLastUpdateTime[nsPod] = time.Now()
+	pbm.podLastUpdateTime[nsPod] = pbm.clock.Now()
 	pbm.podAttempts[nsPod]++
 	pbm.lock.Unlock()
 }

pkg/scheduler/internal/queue/pod_backoff_test.go

@@ -17,15 +17,17 @@ limitations under the License.
 package queue
 
 import (
+	"fmt"
 	"testing"
 	"time"
 
 	ktypes "k8s.io/apimachinery/pkg/types"
+	"k8s.io/apimachinery/pkg/util/clock"
 )
 
 func TestBackoffPod(t *testing.T) {
-	bpm := NewPodBackoffMap(1*time.Second, 10*time.Second)
+	timestamp := time.Now()
-
+	bpm := NewPodBackoffMap(1*time.Second, 10*time.Second, clock.NewFakeClock(timestamp))
 	tests := []struct {
 		podID            ktypes.NamespacedName
 		expectedDuration time.Duration
@@ -61,20 +63,23 @@ func TestBackoffPod(t *testing.T) {
 		},
 	}
 
-	for _, test := range tests {
+	for i, test := range tests {
-		// Backoff the pod
+		t.Run(fmt.Sprintf("step %d", i), func(t *testing.T) {
-		bpm.BackoffPod(test.podID)
+			bpm.BackoffPod(test.podID)
-		// Get backoff duration for the pod
+			backoff, ok := bpm.GetBackoffTime(test.podID)
-		duration := bpm.calculateBackoffDuration(test.podID)
+			if !ok {
-
+				t.Errorf("%v should be backed off", test.podID)
-		if duration != test.expectedDuration {
+			}
-			t.Errorf("expected: %s, got %s for pod %s", test.expectedDuration.String(), duration.String(), test.podID)
+			duration := backoff.Sub(timestamp)
-		}
+			if duration != test.expectedDuration {
+				t.Errorf("expected: %s, got %s for pod %s", test.expectedDuration.String(), duration.String(), test.podID)
+			}
+		})
 	}
 }
 
 func TestClearPodBackoff(t *testing.T) {
-	bpm := NewPodBackoffMap(1*time.Second, 60*time.Second)
+	bpm := NewPodBackoffMap(1*time.Second, 60*time.Second, clock.NewFakeClock(time.Now()))
 	// Clear backoff on an not existed pod
 	bpm.clearPodBackoff(ktypes.NamespacedName{Namespace: "ns", Name: "not-existed"})
 	// Backoff twice for pod foo

pkg/scheduler/internal/queue/scheduling_queue.go

@@ -214,7 +214,7 @@ func NewPriorityQueue(
 	pq := &PriorityQueue{
 		clock:            options.clock,
 		stop:             make(chan struct{}),
-		podBackoff:       NewPodBackoffMap(options.podInitialBackoffDuration, options.podMaxBackoffDuration),
+		podBackoff:       NewPodBackoffMap(options.podInitialBackoffDuration, options.podMaxBackoffDuration, options.clock),
 		activeQ:          heap.NewWithRecorder(podInfoKeyFunc, comp, metrics.NewActivePodsRecorder()),
 		unschedulableQ:   newUnschedulablePodsMap(metrics.NewUnschedulablePodsRecorder()),
 		nominatedPods:    newNominatedPodMap(),

pkg/scheduler/internal/queue/scheduling_queue_test.go

@@ -1515,3 +1515,85 @@ func createAndRunPriorityQueue(fwk framework.Framework, opts ...Option) *Priorit
 	q.Run()
 	return q
 }
+
+func TestBackOffFlow(t *testing.T) {
+	cl := clock.NewFakeClock(time.Now())
+	q := NewPriorityQueue(newDefaultFramework(), WithClock(cl))
+	steps := []struct {
+		wantBackoff time.Duration
+	}{
+		{wantBackoff: time.Second},
+		{wantBackoff: 2 * time.Second},
+		{wantBackoff: 4 * time.Second},
+		{wantBackoff: 8 * time.Second},
+		{wantBackoff: 10 * time.Second},
+		{wantBackoff: 10 * time.Second},
+		{wantBackoff: 10 * time.Second},
+	}
+	pod := &v1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "test-pod",
+			Namespace: "test-ns",
+			UID:       "test-uid",
+		},
+	}
+	podID := nsNameForPod(pod)
+	if err := q.Add(pod); err != nil {
+		t.Fatal(err)
+	}
+
+	for i, step := range steps {
+		t.Run(fmt.Sprintf("step %d", i), func(t *testing.T) {
+			timestamp := cl.Now()
+			// Simulate schedule attempt.
+			podInfo, err := q.Pop()
+			if err != nil {
+				t.Fatal(err)
+			}
+			if podInfo.Attempts != i+1 {
+				t.Errorf("got attempts %d, want %d", podInfo.Attempts, i+1)
+			}
+			if err := q.AddUnschedulableIfNotPresent(podInfo, int64(i)); err != nil {
+				t.Fatal(err)
+			}
+
+			// An event happens.
+			q.MoveAllToActiveOrBackoffQueue("deleted pod")
+
+			if _, ok, _ := q.podBackoffQ.Get(podInfo); !ok {
+				t.Errorf("pod %v is not in the backoff queue", podID)
+			}
+
+			// Check backoff duration.
+			deadline, ok := q.podBackoff.GetBackoffTime(podID)
+			if !ok {
+				t.Errorf("didn't get backoff for pod %s", podID)
+			}
+			backoff := deadline.Sub(timestamp)
+			if backoff != step.wantBackoff {
+				t.Errorf("got backoff %s, want %s", backoff, step.wantBackoff)
+			}
+
+			// Simulate routine that continuously flushes the backoff queue.
+			cl.Step(time.Millisecond)
+			q.flushBackoffQCompleted()
+			// Still in backoff queue after an early flush.
+			if _, ok, _ := q.podBackoffQ.Get(podInfo); !ok {
+				t.Errorf("pod %v is not in the backoff queue", podID)
+			}
+			// Moved out of the backoff queue after timeout.
+			cl.Step(backoff)
+			q.flushBackoffQCompleted()
+			if _, ok, _ := q.podBackoffQ.Get(podInfo); ok {
+				t.Errorf("pod %v is still in the backoff queue", podID)
+			}
+		})
+	}
+	// After some time, backoff information is cleared.
+	cl.Step(time.Hour)
+	q.podBackoff.CleanupPodsCompletesBackingoff()
+	_, ok := q.podBackoff.GetBackoffTime(podID)
+	if ok {
+		t.Errorf("backoff information for pod %s was not cleared", podID)
+	}
+}