[release-1.13] fix(scheduler): latch PodGroup Completed phase to prevent quota re-reservation after Pods deleted

pkg/scheduler/framework/session.go

@@ -376,6 +376,28 @@ func closeSession(ssn *Session) {
 }
 
 func getPodGroupPhase(jobInfo *api.JobInfo, unschedulable bool) scheduling.PodGroupPhase {
+	// Latch the terminal phase: once a PodGroup reached Completed, keep it
+	// Completed while every remaining member task is itself in a terminal
+	// state. Pod deletion (TTL/GC/workload controller) is not atomic and
+	// usually spans multiple sessions; without this latch a partially or
+	// fully deleted group is recomputed below minMember, downgraded back to
+	// Pending/Inqueue, re-admitted by the enqueue action, and the
+	// capacity/proportion plugins reserve its MinResources queue quota again
+	// even though it can never run again. A remaining non-completed task
+	// (e.g. a fresh Pending pod) still falls through to the normal recompute.
+	if jobInfo.PodGroup.Status.Phase == scheduling.PodGroupCompleted {
+		allCompleted := true
+		for _, task := range jobInfo.Tasks {
+			if !api.CompletedStatus(task.Status) {
+				allCompleted = false
+				break
+			}
+		}
+		if allCompleted {
+			return scheduling.PodGroupCompleted
+		}
+	}
+
 	// If running tasks && unschedulable, unknown phase
 	if len(jobInfo.TaskStatusIndex[api.Running]) != 0 && unschedulable {
 		return scheduling.PodGroupUnknown

pkg/scheduler/framework/session_test.go

@@ -91,6 +91,38 @@ func TestGetPodGroupPhase(t *testing.T) {
 				newTask("task-2", api.Succeeded, "node-2")),
 			expected: scheduling.PodGroupCompleted,
 		},
+		{
+			name:     "completed podgroup with no tasks left keeps Completed",
+			job:      newJob(2, scheduling.PodGroupCompleted),
+			expected: scheduling.PodGroupCompleted,
+		},
+		{
+			name: "partially deleted completed podgroup keeps Completed",
+			job: newJob(2, scheduling.PodGroupCompleted,
+				newTask("task-1", api.Succeeded, "node-1")),
+			expected: scheduling.PodGroupCompleted,
+		},
+		{
+			name: "completed podgroup with failed remaining task keeps Completed",
+			job: newJob(2, scheduling.PodGroupCompleted,
+				newTask("task-1", api.Succeeded, "node-1"),
+				newTask("task-2", api.Failed, "node-2")),
+			expected: scheduling.PodGroupCompleted,
+		},
+		{
+			name: "completed podgroup with remaining running task recomputes",
+			job: newJob(2, scheduling.PodGroupCompleted,
+				newTask("task-1", api.Succeeded, "node-1"),
+				newTask("task-2", api.Running, "node-2")),
+			expected: scheduling.PodGroupRunning,
+		},
+		{
+			name: "completed podgroup with new pending task recomputes",
+			job: newJob(2, scheduling.PodGroupCompleted,
+				newTask("task-1", api.Succeeded, "node-1"),
+				newTask("task-2", api.Pending, "")),
+			expected: scheduling.PodGroupPending,
+		},
 		{
 			name: "scheduled releasing tasks below minMember fall to Pending",
 			job: newJob(2, scheduling.PodGroupRunning,