Merge pull request #953 from ravikumarrvit/master

k8s-ci-robot · web-flow · commit 4f99aa34e868 · 2026-03-20T00:44:34.000+05:30
feat: Add configurable pgRejectThreshold to Coscheduling plugin
diff --git a/apis/config/scheme/scheme_test.go b/apis/config/scheme/scheme_test.go
@@ -97,6 +97,7 @@ profiles:
 							Name: coscheduling.Name,
 							Args: &config.CoschedulingArgs{
 								PermitWaitingTimeSeconds: 60,
+								PodGroupRejectPercentage: 10,
 							},
 						},
 						{
@@ -380,6 +381,7 @@ profiles:
       kind: CoschedulingArgs
       permitWaitingTimeSeconds: 10
       podGroupBackoffSeconds: 0
+      podGroupRejectPercentage: 0
     name: Coscheduling
   - args:
       apiVersion: kubescheduler.config.k8s.io/v1
diff --git a/apis/config/types.go b/apis/config/types.go
@@ -32,6 +32,13 @@ type CoschedulingArgs struct {
 	PermitWaitingTimeSeconds int64
 	// PodGroupBackoffSeconds is the backoff time in seconds before a pod group can be scheduled again.
 	PodGroupBackoffSeconds int64
+	// PodGroupRejectPercentage is the percentage (0-100) of unassigned pods relative to
+	// minMember below which PostFilter will not reject the PodGroup.
+	// When the percentage of unassigned pods is at or below this value, PostFilter
+	// optimistically allows remaining pods to retry scheduling instead of rejecting the group.
+	// Default: 10 (10%). Set to 0 to always reject on any failure.
+	// Set to 100 to never reject (disable PostFilter group rejection).
+	PodGroupRejectPercentage int32
 }
 
 // ModeType is a "string" type.
diff --git a/apis/config/v1/defaults.go b/apis/config/v1/defaults.go
@@ -29,6 +29,7 @@ import (
 var (
 	defaultPermitWaitingTimeSeconds int64 = 60
 	defaultPodGroupBackoffSeconds   int64 = 0
+	defaultPodGroupRejectPercentage int32 = 10
 
 	defaultNodeResourcesAllocatableMode = Least
 
@@ -112,6 +113,9 @@ func SetDefaults_CoschedulingArgs(obj *CoschedulingArgs) {
 	if obj.PodGroupBackoffSeconds == nil {
 		obj.PodGroupBackoffSeconds = &defaultPodGroupBackoffSeconds
 	}
+	if obj.PodGroupRejectPercentage == nil {
+		obj.PodGroupRejectPercentage = &defaultPodGroupRejectPercentage
+	}
 }
 
 // SetDefaults_NodeResourcesAllocatableArgs sets the defaults parameters for NodeResourceAllocatable.
diff --git a/apis/config/v1/defaults_test.go b/apis/config/v1/defaults_test.go
@@ -41,17 +41,20 @@ func TestSchedulingDefaults(t *testing.T) {
 			expect: &CoschedulingArgs{
 				PermitWaitingTimeSeconds: pointer.Int64Ptr(60),
 				PodGroupBackoffSeconds:   pointer.Int64Ptr(0),
+				PodGroupRejectPercentage: pointer.Int32Ptr(10),
 			},
 		},
 		{
 			name: "set non default CoschedulingArgs",
 			config: &CoschedulingArgs{
 				PermitWaitingTimeSeconds: pointer.Int64Ptr(60),
 				PodGroupBackoffSeconds:   pointer.Int64Ptr(20),
+				PodGroupRejectPercentage: pointer.Int32Ptr(50),
 			},
 			expect: &CoschedulingArgs{
 				PermitWaitingTimeSeconds: pointer.Int64Ptr(60),
 				PodGroupBackoffSeconds:   pointer.Int64Ptr(20),
+				PodGroupRejectPercentage: pointer.Int32Ptr(50),
 			},
 		},
 		{
diff --git a/apis/config/v1/types.go b/apis/config/v1/types.go
@@ -32,6 +32,11 @@ type CoschedulingArgs struct {
 	PermitWaitingTimeSeconds *int64 `json:"permitWaitingTimeSeconds,omitempty"`
 	// PodGroupBackoffSeconds is the backoff time in seconds before a pod group can be scheduled again.
 	PodGroupBackoffSeconds *int64 `json:"podGroupBackoffSeconds,omitempty"`
+	// PodGroupRejectPercentage is the percentage (0-100) of unassigned pods relative to
+	// minMember below which PostFilter will not reject the PodGroup.
+	// Default: 10 (10%). Set to 0 to always reject on any failure.
+	// Set to 100 to never reject (disable PostFilter group rejection).
+	PodGroupRejectPercentage *int32 `json:"podGroupRejectPercentage,omitempty"`
 }
 
 // ModeType is a type "string".
diff --git a/apis/config/v1/zz_generated.conversion.go b/apis/config/v1/zz_generated.conversion.go
diff --git a/apis/config/v1/zz_generated.deepcopy.go b/apis/config/v1/zz_generated.deepcopy.go
diff --git a/apis/config/validation/validation_pluginargs.go b/apis/config/validation/validation_pluginargs.go
@@ -104,6 +104,10 @@ func ValidateCoschedulingArgs(args *config.CoschedulingArgs, _ *field.Path) erro
 		allErrs = append(allErrs, field.Invalid(field.NewPath("podGroupBackoffSeconds"),
 			args.PodGroupBackoffSeconds, "must be greater than 0"))
 	}
+	if args.PodGroupRejectPercentage < 0 || args.PodGroupRejectPercentage > 100 {
+		allErrs = append(allErrs, field.Invalid(field.NewPath("podGroupRejectPercentage"),
+			args.PodGroupRejectPercentage, "must be between 0 and 100"))
+	}
 	if len(allErrs) == 0 {
 		return nil
 	}
diff --git a/apis/config/validation/validation_plugingargs_test.go b/apis/config/validation/validation_plugingargs_test.go
@@ -83,6 +83,25 @@ func TestValidateCoschedulingArgs(t *testing.T) {
 			args: &config.CoschedulingArgs{
 				PermitWaitingTimeSeconds: 30,
 				PodGroupBackoffSeconds:   60,
+				PodGroupRejectPercentage: 10,
+			},
+			expectedErr: nil,
+		},
+		{
+			description: "valid PodGroupRejectPercentage at boundaries",
+			args: &config.CoschedulingArgs{
+				PermitWaitingTimeSeconds: 30,
+				PodGroupBackoffSeconds:   0,
+				PodGroupRejectPercentage: 0,
+			},
+			expectedErr: nil,
+		},
+		{
+			description: "valid PodGroupRejectPercentage at upper boundary",
+			args: &config.CoschedulingArgs{
+				PermitWaitingTimeSeconds: 30,
+				PodGroupBackoffSeconds:   0,
+				PodGroupRejectPercentage: 100,
 			},
 			expectedErr: nil,
 		},
@@ -91,6 +110,7 @@ func TestValidateCoschedulingArgs(t *testing.T) {
 			args: &config.CoschedulingArgs{
 				PermitWaitingTimeSeconds: -10,
 				PodGroupBackoffSeconds:   60,
+				PodGroupRejectPercentage: 10,
 			},
 			expectedErr: fmt.Errorf("permitWaitingTimeSeconds: Invalid value: %v: must be greater than 0", -10),
 		},
@@ -99,14 +119,34 @@ func TestValidateCoschedulingArgs(t *testing.T) {
 			args: &config.CoschedulingArgs{
 				PermitWaitingTimeSeconds: 30,
 				PodGroupBackoffSeconds:   -20,
+				PodGroupRejectPercentage: 10,
 			},
 			expectedErr: fmt.Errorf("podGroupBackoffSeconds: Invalid value: %v: must be greater than 0", -20),
 		},
+		{
+			description: "invalid PodGroupRejectPercentage (negative value)",
+			args: &config.CoschedulingArgs{
+				PermitWaitingTimeSeconds: 30,
+				PodGroupBackoffSeconds:   0,
+				PodGroupRejectPercentage: -1,
+			},
+			expectedErr: fmt.Errorf("podGroupRejectPercentage: Invalid value: %v: must be between 0 and 100", -1),
+		},
+		{
+			description: "invalid PodGroupRejectPercentage (greater than 100)",
+			args: &config.CoschedulingArgs{
+				PermitWaitingTimeSeconds: 30,
+				PodGroupBackoffSeconds:   0,
+				PodGroupRejectPercentage: 150,
+			},
+			expectedErr: fmt.Errorf("podGroupRejectPercentage: Invalid value: %v: must be between 0 and 100", 150),
+		},
 		{
 			description: "both PermitWaitingTimeSeconds and PodGroupBackoffSeconds are negative",
 			args: &config.CoschedulingArgs{
 				PermitWaitingTimeSeconds: -30,
 				PodGroupBackoffSeconds:   -20,
+				PodGroupRejectPercentage: 10,
 			},
 			expectedErr: fmt.Errorf(
 				"[permitWaitingTimeSeconds: Invalid value: %v: must be greater than 0, podGroupBackoffSeconds: Invalid value: %v: must be greater than 0]",
diff --git a/kep/42-podgroup-coscheduling/README.md b/kep/42-podgroup-coscheduling/README.md
@@ -15,6 +15,7 @@
     - [QueueSort](#queuesort)
     - [PreFilter](#prefilter)
     - [PostFilter](#postfilter)
+    - [Backoff](#backoff)
     - [Permit](#permit)
   - [Known Limitations](#known-limitations)
 <!-- /toc -->
@@ -158,7 +159,26 @@ For any pod that gets rejected, their pod group would be added to a backoff list
 
 #### PostFilter
 
-If the gap to reach the quorum of a PodGroup is greater than 10%, we reject the whole PodGroup. Note that this plugin should be configured as the last one among PostFilter plugins.
+PostFilter handles scheduling failures for pods that belong to a PodGroup. When a pod fails Filter, PostFilter evaluates whether the PodGroup should be rejected based on how far it is from meeting its quorum:
+
+1. If the number of assigned pods already meets `minMember`, no action is taken.
+2. If the fraction of unassigned pods is at or below `podGroupRejectPercentage` (default: 10%), PostFilter returns `Unschedulable` without rejecting the group — the remaining pods get another scheduling attempt.
+3. If the fraction of unassigned pods exceeds the threshold, PostFilter rejects all waiting pods in the group and optionally triggers backoff (see below).
+
+The `podGroupRejectPercentage` parameter (default: `10`) is configurable in the scheduler's `CoschedulingArgs`. Set it to `0` to always reject on any failure, or `100` to never reject.
+
+Note that this plugin should be configured as the last one among PostFilter plugins.
+
+#### Backoff
+
+When `podGroupBackoffSeconds` is set to a positive value in `CoschedulingArgs`, PostFilter places a failed PodGroup into a time-based backoff cache after rejection. During the backoff window, PreFilter immediately rejects all pods from the PodGroup with `UnschedulableAndUnresolvable`, preventing wasteful scheduling cycles.
+
+Backoff is triggered only when all of the following conditions are met:
+- `podGroupBackoffSeconds > 0`
+- The fraction of unassigned pods exceeds `podGroupRejectPercentage`
+- The total number of pods with the PodGroup label is at least `minMember`
+
+The backoff state is stored in a TTL-based in-memory cache that auto-evicts entries after the configured duration.
 
 #### Permit
 
diff --git a/pkg/coscheduling/coscheduling.go b/pkg/coscheduling/coscheduling.go
@@ -40,11 +40,12 @@ import (
 
 // Coscheduling is a plugin that schedules pods in a group.
 type Coscheduling struct {
-	logger           klog.Logger
-	frameworkHandler framework.Handle
-	pgMgr            core.Manager
-	scheduleTimeout  *time.Duration
-	pgBackoff        *time.Duration
+	logger            klog.Logger
+	frameworkHandler  framework.Handle
+	pgMgr             core.Manager
+	scheduleTimeout   *time.Duration
+	pgBackoff         *time.Duration
+	pgRejectThreshold float64
 }
 
 var _ framework.QueueSortPlugin = &Coscheduling{}
@@ -94,10 +95,11 @@ func New(ctx context.Context, obj runtime.Object, handle framework.Handle) (fram
 		handle.SharedInformerFactory().Core().V1().Pods(),
 	)
 	plugin := &Coscheduling{
-		logger:           lh,
-		frameworkHandler: handle,
-		pgMgr:            pgMgr,
-		scheduleTimeout:  &scheduleTimeDuration,
+		logger:            lh,
+		frameworkHandler:  handle,
+		pgMgr:             pgMgr,
+		scheduleTimeout:   &scheduleTimeDuration,
+		pgRejectThreshold: float64(args.PodGroupRejectPercentage) / 100.0,
 	}
 	if args.PodGroupBackoffSeconds < 0 {
 		err := fmt.Errorf("parse arguments failed")
@@ -176,10 +178,10 @@ func (cs *Coscheduling) PostFilter(ctx context.Context, state fwk.CycleState, po
 		return &framework.PostFilterResult{}, fwk.NewStatus(fwk.Unschedulable)
 	}
 
-	// If the gap is less than/equal 10%, we may want to try subsequent Pods
+	// If the gap is less than/equal the reject threshold, we may want to try subsequent Pods
 	// to see they can satisfy the PodGroup
-	notAssignedPercentage := float32(int(pg.Spec.MinMember)-assigned) / float32(pg.Spec.MinMember)
-	if notAssignedPercentage <= 0.1 {
+	notAssignedPercentage := float64(int(pg.Spec.MinMember)-assigned) / float64(pg.Spec.MinMember)
+	if notAssignedPercentage <= cs.pgRejectThreshold {
 		lh.V(4).Info("A small gap of pods to reach the quorum", "podGroup", klog.KObj(pg), "percentage", notAssignedPercentage)
 		return &framework.PostFilterResult{}, fwk.NewStatus(fwk.Unschedulable)
 	}
diff --git a/pkg/coscheduling/coscheduling_test.go b/pkg/coscheduling/coscheduling_test.go
diff --git a/site/content/en/docs/plugins/coscheduling.md b/site/content/en/docs/plugins/coscheduling.md

Original file line number	Diff line number	Diff line change
`@@ -29,6 +29,7 @@ import (`
`29`	`29`	`var (`
`30`	`30`	`defaultPermitWaitingTimeSeconds int64 = 60`
`31`	`31`	`defaultPodGroupBackoffSeconds int64 = 0`
	`32`	`+ defaultPodGroupRejectPercentage int32 = 10`
`32`	`33`
`33`	`34`	`defaultNodeResourcesAllocatableMode = Least`
`34`	`35`
`@@ -112,6 +113,9 @@ func SetDefaults_CoschedulingArgs(obj *CoschedulingArgs) {`
`112`	`113`	`if obj.PodGroupBackoffSeconds == nil {`
`113`	`114`	`obj.PodGroupBackoffSeconds = &defaultPodGroupBackoffSeconds`
`114`	`115`	`}`
	`116`	`+ if obj.PodGroupRejectPercentage == nil {`
	`117`	`+ obj.PodGroupRejectPercentage = &defaultPodGroupRejectPercentage`
	`118`	`+ }`
`115`	`119`	`}`
`116`	`120`
`117`	`121`	`// SetDefaults_NodeResourcesAllocatableArgs sets the defaults parameters for NodeResourceAllocatable.`
Original file line number	Diff line number	Diff line change
`@@ -104,6 +104,10 @@ func ValidateCoschedulingArgs(args config.CoschedulingArgs, _ field.Path) erro`
`104`	`104`	`allErrs = append(allErrs, field.Invalid(field.NewPath("podGroupBackoffSeconds"),`
`105`	`105`	`args.PodGroupBackoffSeconds, "must be greater than 0"))`
`106`	`106`	`}`
	`107`	`+ if args.PodGroupRejectPercentage < 0 \|\| args.PodGroupRejectPercentage > 100 {`
	`108`	`+ allErrs = append(allErrs, field.Invalid(field.NewPath("podGroupRejectPercentage"),`
	`109`	`+ args.PodGroupRejectPercentage, "must be between 0 and 100"))`
	`110`	`+ }`
`107`	`111`	`if len(allErrs) == 0 {`
`108`	`112`	`return nil`
`109`	`113`	`}`