test(binder): cover extended-resource claim rollback in UnAllocate

Add unit coverage for the DRA extended-resource rollback path: after a
successful bind creates the synthetic annotated ResourceClaim, a later
bind failure triggers UnAllocate, which must delete the claim and clear
pod.Status.ExtendedResourceClaimStatus. Includes idempotency cases for
repeated rollback and rollback with nothing to clean up.

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

Author: gshaibi

pkg/binder/plugins/k8s-plugins/dynamicresources/dra_test.go

@@ -0,0 +1,124 @@
+// Copyright 2025 NVIDIA CORPORATION
+// SPDX-License-Identifier: Apache-2.0
+
+package dynamicresources
+
+import (
+	"context"
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	corev1 "k8s.io/api/core/v1"
+	resourceapi "k8s.io/api/resource/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/apimachinery/pkg/types"
+	"k8s.io/client-go/kubernetes/fake"
+	k8stesting "k8s.io/client-go/testing"
+
+	"github.com/kai-scheduler/KAI-scheduler/pkg/apis/scheduling/v1alpha2"
+)
+
+func TestDynamicResources(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Binder DynamicResources Suite")
+}
+
+var _ = Describe("dynamicResourcesPlugin extended-resource rollback", func() {
+	const (
+		namespace = "test-ns"
+		podName   = "test-pod"
+		podUID    = "pod-uid-123"
+	)
+
+	var (
+		ctx     context.Context
+		client  *fake.Clientset
+		plugin  *dynamicResourcesPlugin
+		pod     *corev1.Pod
+		request *v1alpha2.BindRequest
+	)
+
+	listClaims := func() []resourceapi.ResourceClaim {
+		list, err := client.ResourceV1().ResourceClaims(namespace).List(ctx, metav1.ListOptions{})
+		Expect(err).NotTo(HaveOccurred())
+		return list.Items
+	}
+
+	getPod := func() *corev1.Pod {
+		p, err := client.CoreV1().Pods(namespace).Get(ctx, podName, metav1.GetOptions{})
+		Expect(err).NotTo(HaveOccurred())
+		return p
+	}
+
+	BeforeEach(func() {
+		ctx = context.Background()
+		pod = &corev1.Pod{
+			ObjectMeta: metav1.ObjectMeta{
+				Name:      podName,
+				Namespace: namespace,
+				UID:       types.UID(podUID),
+			},
+		}
+		client = fake.NewSimpleClientset(pod)
+		// The fake tracker does not populate Name from GenerateName; do it here so the
+		// claim created during Bind is retrievable by the rollback path.
+		client.PrependReactor("create", "resourceclaims", func(action k8stesting.Action) (bool, runtime.Object, error) {
+			claim := action.(k8stesting.CreateAction).GetObject().(*resourceapi.ResourceClaim)
+			if claim.Name == "" && claim.GenerateName != "" {
+				claim.Name = claim.GenerateName + "generated"
+			}
+			return false, claim, nil
+		})
+
+		plugin = &dynamicResourcesPlugin{client: client, bindTimeout: 5}
+		request = &v1alpha2.BindRequest{
+			Spec: v1alpha2.BindRequestSpec{
+				ExtendedResourceClaimAllocation: &v1alpha2.ExtendedResourceClaimAllocation{
+					Allocation:     &resourceapi.AllocationResult{},
+					DeviceRequests: []resourceapi.DeviceRequest{{Name: "req0"}},
+					ContainerMappings: []corev1.ContainerExtendedResourceRequest{
+						{ContainerName: "main", ResourceName: "nvidia.com/gpu", RequestName: "req0"},
+					},
+				},
+			},
+		}
+	})
+
+	It("rolls back the created claim and clears pod status after a later bind failure", func() {
+		Expect(plugin.Bind(ctx, pod, request, nil)).To(Succeed())
+
+		claims := listClaims()
+		Expect(claims).To(HaveLen(1))
+		Expect(claims[0].Annotations).To(HaveKeyWithValue(resourceapi.ExtendedResourceClaimAnnotation, "true"))
+		Expect(getPod().Status.ExtendedResourceClaimStatus).NotTo(BeNil())
+
+		// A later plugin's Bind fails: the framework rolls back by calling UnAllocate.
+		plugin.UnAllocate(ctx, pod, "node", nil)
+
+		Expect(listClaims()).To(BeEmpty())
+		Expect(getPod().Status.ExtendedResourceClaimStatus).To(BeNil())
+	})
+
+	It("is idempotent when rollback runs more than once", func() {
+		Expect(plugin.Bind(ctx, pod, request, nil)).To(Succeed())
+		Expect(listClaims()).To(HaveLen(1))
+
+		plugin.UnAllocate(ctx, pod, "node", nil)
+		Expect(listClaims()).To(BeEmpty())
+		Expect(getPod().Status.ExtendedResourceClaimStatus).To(BeNil())
+
+		// Second rollback must be a no-op: the claim is already gone and status already nil.
+		Expect(func() { plugin.UnAllocate(ctx, pod, "node", nil) }).NotTo(Panic())
+		Expect(listClaims()).To(BeEmpty())
+		Expect(getPod().Status.ExtendedResourceClaimStatus).To(BeNil())
+	})
+
+	It("does nothing when there is no extended-resource claim to roll back", func() {
+		plugin.UnAllocate(ctx, pod, "node", nil)
+		Expect(listClaims()).To(BeEmpty())
+		Expect(getPod().Status.ExtendedResourceClaimStatus).To(BeNil())
+	})
+})