fix(api): allow PodGroup minMember of 0 (#1989)

Backport of 62c594b to v0.9, adapted
for non-pointer MinMember and the SubGroups-based scheduler API.

Signed-off-by: lin121291 <4jp33f9e@gmail.com>

Signed-off-by: gshaibi <gshaibi@nvidia.com>

Author: gshaibi

.changes/unreleased/fixed-20260728-212512.yaml

@@ -0,0 +1,3 @@
+kind: Fixed
+body: |-
+  Allow PodGroup minMember of 0 for workloads with no gang requirement

deployments/kai-scheduler/crds/scheduling.run.ai_podgroups.yaml

@@ -61,6 +61,7 @@ spec:
                   MinMember defines the minimal number of members/tasks to run the pod group;
                   if there's not enough resources to start all tasks, the scheduler
                   will not start anyone.
+                  A value of 0 means no gang requirement: all pods are scheduled elastically (e.g. scale-to-zero workloads).
                 format: int32
                 type: integer
               parallelism:

pkg/apis/scheduling/v2alpha2/podgroup_types.go

@@ -34,6 +34,7 @@ type PodGroupSpec struct {
 	// MinMember defines the minimal number of members/tasks to run the pod group;
 	// if there's not enough resources to start all tasks, the scheduler
 	// will not start anyone.
+	// A value of 0 means no gang requirement: all pods are scheduled elastically (e.g. scale-to-zero workloads).
 	MinMember int32 `json:"minMember,omitempty" protobuf:"bytes,1,opt,name=minMember"`
 
 	// Queue defines the queue to allocate resource for PodGroup; if queue does not exist,

pkg/podgrouper/podgrouper/plugins/knative/knative_test.go

@@ -4,7 +4,6 @@
 package knative
 
 import (
-	"strconv"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -113,6 +112,17 @@ func TestGetPodGroupMetadata(t *testing.T) {
 }
 
 func TestGetPodGroupMetadata_MinScale(t *testing.T) {
+	for minScale, expectedMinAvailable := range map[string]int32{
+		"3": 3,
+		"0": 0, // scale-to-zero: no gang requirement
+	} {
+		t.Run("min-scale="+minScale, func(t *testing.T) {
+			testGetPodGroupMetadataMinScale(t, minScale, expectedMinAvailable)
+		})
+	}
+}
+
+func testGetPodGroupMetadataMinScale(t *testing.T, minScale string, expectedMinAvailable int32) {
 	service := &unstructured.Unstructured{
 		Object: map[string]interface{}{
 			"kind":       "Service",
@@ -156,7 +166,7 @@ func TestGetPodGroupMetadata_MinScale(t *testing.T) {
 			Namespace: "test_namespace",
 			UID:       "2",
 			Annotations: map[string]string{
-				"autoscaling.knative.dev/min-scale": "3",
+				"autoscaling.knative.dev/min-scale": minScale,
 			},
 		},
 		Spec:   knative.RevisionSpec{},
@@ -197,7 +207,7 @@ func TestGetPodGroupMetadata_MinScale(t *testing.T) {
 	assert.Equal(t, "pg-revision-2", metadata.Name)
 	assert.Equal(t, constants.InferencePriorityClass, metadata.PriorityClassName)
 	assert.Equal(t, "test_queue", metadata.Queue)
-	assert.Equal(t, "3", strconv.Itoa(int(metadata.MinAvailable)))
+	assert.Equal(t, expectedMinAvailable, metadata.MinAvailable)
 }
 
 func TestGetPodGroupMetadataBackwardsCompatibility(t *testing.T) {

pkg/scheduler/api/podgroup_info/job_info.go

@@ -204,9 +204,9 @@ func (pgi *PodGroupInfo) setSubGroups(podGroup *enginev2alpha2.PodGroup) {
 	}
 	defaultSubGroup, found := pgi.SubGroups[DefaultSubGroup]
 	if !found {
-		pgi.SubGroups[DefaultSubGroup] = NewSubGroupInfo(DefaultSubGroup, max(podGroup.Spec.MinMember, 1))
+		pgi.SubGroups[DefaultSubGroup] = NewSubGroupInfo(DefaultSubGroup, max(podGroup.Spec.MinMember, 0))
 	} else {
-		defaultSubGroup.SetMinAvailable(max(podGroup.Spec.MinMember, 1))
+		defaultSubGroup.SetMinAvailable(max(podGroup.Spec.MinMember, 0))
 	}
 }
 

pkg/scheduler/api/podgroup_info/job_info_test.go

@@ -115,6 +115,36 @@ func TestAddTaskInfo(t *testing.T) {
 	}
 }
 
+func TestSetPodGroupMinMember(t *testing.T) {
+	tests := []struct {
+		name             string
+		minMember        int32
+		expectedMinAvail int32
+	}{
+		{"zero minMember is honored", 0, 0},
+		{"positive minMember is honored", 3, 3},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			info := NewPodGroupInfo("group-1")
+			info.SetPodGroup(&v2alpha2.PodGroup{
+				ObjectMeta: metav1.ObjectMeta{
+					Name:      "group-1",
+					Namespace: "ns-1",
+				},
+				Spec: v2alpha2.PodGroupSpec{
+					Queue:     "queue-1",
+					MinMember: test.minMember,
+				},
+			})
+
+			if got := info.SubGroups[DefaultSubGroup].GetMinAvailable(); got != test.expectedMinAvail {
+				t.Errorf("expected minAvailable %d, got %d", test.expectedMinAvail, got)
+			}
+		})
+	}
+}
+
 func TestDeleteTaskInfo(t *testing.T) {
 	// case1
 	case01_uid := common_info.PodGroupID("owner1")

test/e2e/suites/integrations/third_party/knative/knative_test.go

@@ -118,6 +118,32 @@ var _ = Describe("Knative integration", Ordered, func() {
 				"Expected minmember of podgroup to be 1")
 		})
 
+		It("knative service with scale-to-zero - minscale 0", func(ctx context.Context) {
+			namespace := queue.GetConnectedNamespaceToQueue(testCtx.Queues[0])
+			serviceName := "knative-service-" + utils.GenerateRandomK8sName(10)
+
+			knativeService := getKnativeServiceObject(serviceName, namespace, testCtx.Queues[0].Name)
+
+			knativeService.Spec.ConfigurationSpec.Template.Annotations[minScaleAnnotation] = "0"
+
+			Expect(testCtx.ControllerClient.Create(ctx, knativeService)).To(Succeed())
+			defer func() {
+				Expect(testCtx.ControllerClient.Delete(ctx, knativeService)).To(Succeed())
+			}()
+
+			// Knative still creates the initial pod on deployment even with min-scale 0
+			pods := getServicePods(ctx, testCtx, serviceName, namespace, 1)
+
+			wait.ForPodsScheduled(ctx, testCtx.ControllerClient, namespace, pods)
+
+			var podGroups v2alpha2.PodGroupList
+			Expect(testCtx.ControllerClient.List(ctx, &podGroups, client.InNamespace(namespace))).To(Succeed())
+			Expect(len(podGroups.Items)).To(Equal(1),
+				"Expected one podgroup for the revision")
+			Expect(podGroups.Items[0].Spec.MinMember).To(Equal(int32(0)),
+				"Expected minmember of podgroup to be 0")
+		})
+
 		Context("Non-Gang scheduling", func() {
 			BeforeAll(func(ctx context.Context) {
 				setKnativeGangAndWait(ctx, testCtx, pointer.Bool(false))