fix(flowcontrol): apply effective defaults for DefaultRequestTTL and per-band MaxRequests

apix/config/v1alpha1/endpointpickerconfig_types.go

@@ -351,23 +351,25 @@ type FlowControlConfig struct {
 	// levels. If this limit is exceeded, new requests will be rejected even if their specific
 	// priority band has capacity.
 	// Accepts standard Kubernetes resource quantities (e.g., "1Gi", "500M").
-	// If omitted, no global byte limit is enforced.
+	// If omitted or "0", no global byte limit is enforced.
 	MaxBytes *resource.Quantity `json:"maxBytes,omitempty"`
 
 	// +optional
 	// MaxRequests defines the global maximum number of concurrent requests across all priority
 	// levels. If this limit is exceeded, new requests will be rejected even if their specific
 	// priority band has capacity.
 	// Accepts standard Kubernetes resource quantities (e.g., "100", "1k").
-	// If omitted, no global request limit is enforced.
+	// If omitted or "0", no global request limit is enforced.
 	MaxRequests *resource.Quantity `json:"maxRequests,omitempty"`
 
 	// +optional
-	// DefaultRequestTTL serves as a fallback timeout for requests that do not specify their own
-	// deadline.
-	// It ensures that requests do not hang indefinitely in the queue.
-	// If 0 or omitted, it defaults to the client context deadline, meaning requests may wait
-	// indefinitely unless cancelled by the client.
+	// DefaultRequestTTL bounds how long a request may wait in the queue before it is evicted.
+	// If omitted, it defaults to 60s. This is a queue-wait budget: a request that cannot dispatch
+	// within it is shed with a retryable backpressure error rather than served late, and it is the
+	// only bound on queue wait when neither the client nor the gateway enforces a request deadline.
+	// Where such deadlines exist and fire sooner, they evict the request first (client disconnect).
+	// An explicit "0s" disables the TTL: requests then wait until client disconnect or controller
+	// shutdown.
 	DefaultRequestTTL *metav1.Duration `json:"defaultRequestTTL,omitempty"`
 
 	// +optional
@@ -382,7 +384,10 @@ type FlowControlConfig struct {
 	// +optional
 	// DefaultNegativePriorityBand allows you to define a separate template for priority levels
 	// strictly below zero. This enables designating negative-priority traffic as sheddable by
-	// setting lower capacity limits (e.g., maxBytes: "0" to drop immediately).
+	// setting lower capacity limits (e.g., a small maxRequests, so that under saturation the band
+	// fills quickly and subsequent requests are rejected immediately rather than queued).
+	// Note that a value of "0" is treated as unset and receives the system default, not zero
+	// capacity.
 	// If not specified, negative priorities fall back to DefaultPriorityBand.
 	DefaultNegativePriorityBand *PriorityBandConfig `json:"defaultNegativePriorityBand,omitempty"`
 
@@ -457,13 +462,15 @@ type PriorityBandConfig struct {
 	// +optional
 	// MaxBytes is the maximum number of bytes allowed for this priority band.
 	// Accepts standard Kubernetes resource quantities (e.g., "1Gi", "500M").
-	// If omitted, the system default is used.
+	// If omitted or "0", the system default (1G) is used. Per-band limits are always bounded; to
+	// effectively remove the bound, set an explicit large value.
 	MaxBytes *resource.Quantity `json:"maxBytes,omitempty"`
 
 	// +optional
 	// MaxRequests is the maximum number of concurrent requests allowed for this priority band.
 	// Accepts standard Kubernetes resource quantities (e.g., "100", "1k").
-	// If omitted, no request limit is enforced.
+	// If omitted or "0", the system default (5000) is used. Per-band limits are always bounded; to
+	// effectively remove the bound, set an explicit large value.
 	MaxRequests *resource.Quantity `json:"maxRequests,omitempty"`
 
 	// +optional

pkg/epp/config/loader/flowcontrol_test.go

@@ -154,10 +154,13 @@ func TestBuildRegistryConfig(t *testing.T) {
 			apiConfig: nil,
 			assertion: func(t *testing.T, cfg *registry.Config) {
 				assert.Equal(t, uint64(0), cfg.MaxBytes, "Default global limit should be 0 (unlimited)")
+				assert.Equal(t, uint64(0), cfg.MaxRequests, "Default global request limit should be 0 (unlimited)")
 				require.NotNil(t, cfg.DefaultPriorityBand,
 					"Default priority band template should be initialized automatically")
 				assert.Equal(t, uint64(1_000_000_000) /* registry default: 1 GB */, cfg.DefaultPriorityBand.MaxBytes,
 					"Default template should use system default capacity")
+				assert.Equal(t, uint64(5000) /* registry default */, cfg.DefaultPriorityBand.MaxRequests,
+					"Default template should use the system default request-count capacity")
 			},
 		},
 
@@ -207,6 +210,49 @@ func TestBuildRegistryConfig(t *testing.T) {
 					"Explicit 0 in DefaultPriorityBand template should be treated as 'Use Default'")
 			},
 		},
+		{
+			name: "ShouldApplyDefault_WhenBandMaxRequestsIsNilOrZero",
+			apiConfig: &configapi.FlowControlConfig{
+				PriorityBands: []configapi.PriorityBandConfig{
+					{
+						Priority: 1,
+						// MaxRequests omitted
+					},
+					{
+						Priority:    2,
+						MaxRequests: ptr.To(resource.MustParse("0")), // Explicitly zero
+					},
+					{
+						Priority:    3,
+						MaxRequests: ptr.To(resource.MustParse("250")),
+					},
+				},
+			},
+			assertion: func(t *testing.T, cfg *registry.Config) {
+				require.Contains(t, cfg.PriorityBands, 1)
+				assert.Equal(t, uint64(5000) /* registry default */, cfg.PriorityBands[1].MaxRequests,
+					"Omitted MaxRequests (nil) should result in the system default (5000)")
+				require.Contains(t, cfg.PriorityBands, 2)
+				assert.Equal(t, uint64(5000) /* registry default */, cfg.PriorityBands[2].MaxRequests,
+					"Explicit MaxRequests (0) should be treated as 'Use Default' (5000)")
+				require.Contains(t, cfg.PriorityBands, 3)
+				assert.Equal(t, uint64(250), cfg.PriorityBands[3].MaxRequests,
+					"Explicit MaxRequests should be preserved")
+			},
+		},
+		{
+			name: "ShouldApplyDefault_WhenNegativeBandTemplateMaxRequestsIsZero",
+			apiConfig: &configapi.FlowControlConfig{
+				DefaultNegativePriorityBand: &configapi.PriorityBandConfig{
+					MaxRequests: ptr.To(resource.MustParse("0")), // Explicitly zero
+				},
+			},
+			assertion: func(t *testing.T, cfg *registry.Config) {
+				require.NotNil(t, cfg.DefaultNegativePriorityBand)
+				assert.Equal(t, uint64(5000) /* registry default */, cfg.DefaultNegativePriorityBand.MaxRequests,
+					"Explicit 0 in the negative band template should be treated as 'Use Default', not zero capacity")
+			},
+		},
 
 		// --- Validation Errors ---
 		{
@@ -299,37 +345,8 @@ func TestBuildRegistryConfig(t *testing.T) {
 		},
 
 		// --- MaxRequests: Defaulting Logic ---
-		{
-			name: "ShouldDefaultToZero_WhenMaxRequestsIsNil",
-			apiConfig: &configapi.FlowControlConfig{
-				PriorityBands: []configapi.PriorityBandConfig{
-					{
-						Priority: 1,
-					},
-				},
-			},
-			assertion: func(t *testing.T, cfg *registry.Config) {
-				require.Contains(t, cfg.PriorityBands, 1)
-				assert.Equal(t, uint64(0), cfg.PriorityBands[1].MaxRequests,
-					"Omitted MaxRequests should default to 0 (no request limit)")
-			},
-		},
-		{
-			name: "ShouldDefaultToZero_WhenMaxRequestsIsZero",
-			apiConfig: &configapi.FlowControlConfig{
-				PriorityBands: []configapi.PriorityBandConfig{
-					{
-						Priority:    1,
-						MaxRequests: ptr.To(resource.MustParse("0")),
-					},
-				},
-			},
-			assertion: func(t *testing.T, cfg *registry.Config) {
-				require.Contains(t, cfg.PriorityBands, 1)
-				assert.Equal(t, uint64(0), cfg.PriorityBands[1].MaxRequests,
-					"Explicit MaxRequests=0 should remain 0 (no request limit)")
-			},
-		},
+		// Nil and explicit-zero band MaxRequests defaulting is covered by
+		// ShouldApplyDefault_WhenBandMaxRequestsIsNilOrZero above.
 
 		// --- MaxRequests: Validation Errors ---
 		{

pkg/epp/flowcontrol/controller/config.go

@@ -24,6 +24,13 @@ import (
 )
 
 const (
+	// defaultRequestTTL is the default Time-To-Live applied to queued requests when the configuration
+	// does not specify one. It is a queue-wait budget: a request still undispatched after this long is
+	// shed with a retryable backpressure signal instead of being served with severely degraded
+	// time-to-first-token. It is also the only bound on queue wait when neither the client nor the
+	// gateway enforces a request deadline (the well-lit guides configure no gateway request timeout);
+	// where such deadlines exist and fire sooner, context cancellation evicts the request first.
+	defaultRequestTTL = 60 * time.Second
 	// defaultExpiryCleanupInterval is the default frequency for scanning for expired items.
 	defaultExpiryCleanupInterval = 1 * time.Second
 	// defaultEnqueueChannelBufferSize is the default size of a worker's incoming request buffer.
@@ -32,9 +39,12 @@ const (
 
 // Config holds the configuration for the `FlowController`.
 type Config struct {
-	// DefaultRequestTTL is the default Time-To-Live applied to requests that do not
-	// specify their own TTL hint.
-	// Optional: If zero, no TTL is applied by default and we rely solely on request context cancellation.
+	// DefaultRequestTTL is the default Time-To-Live applied to requests that do not specify their own
+	// TTL hint. Because the admission adapter does not currently plumb a per-request hint, this value
+	// governs every request entering flow control.
+	// Optional: Defaults to `defaultRequestTTL` (60s). An explicit zero disables the TTL entirely, in
+	// which case queued requests are bounded only by request context cancellation (client disconnect
+	// or gateway timeout).
 	DefaultRequestTTL time.Duration
 
 	// ExpiryCleanupInterval is the interval at which each processor scans its queues for expired items.
@@ -76,6 +86,7 @@ func NewConfigFromAPI(apiConfig *configapi.FlowControlConfig) (*Config, error) {
 // NewConfig creates a new Config with the given options, applying defaults and validation.
 func NewConfig(opts ...ConfigOption) (*Config, error) {
 	c := &Config{
+		DefaultRequestTTL:        defaultRequestTTL,
 		ExpiryCleanupInterval:    defaultExpiryCleanupInterval,
 		EnqueueChannelBufferSize: defaultEnqueueChannelBufferSize,
 	}

pkg/epp/flowcontrol/controller/config_test.go

@@ -40,6 +40,18 @@ func TestNewConfig(t *testing.T) {
 			name:      "Defaults_ShouldBeApplied_WhenNoOptionsProvided",
 			opts:      nil,
 			expectErr: false,
+			expectedCfg: Config{
+				DefaultRequestTTL:        defaultRequestTTL,
+				ExpiryCleanupInterval:    defaultExpiryCleanupInterval,
+				EnqueueChannelBufferSize: defaultEnqueueChannelBufferSize,
+			},
+		},
+		{
+			name: "ZeroDefaultRequestTTL_ShouldDisableTTL",
+			opts: []ConfigOption{
+				WithDefaultRequestTTL(0),
+			},
+			expectErr: false,
 			expectedCfg: Config{
 				DefaultRequestTTL:        0,
 				ExpiryCleanupInterval:    defaultExpiryCleanupInterval,
@@ -136,7 +148,7 @@ func TestNewConfigFromAPI(t *testing.T) {
 					"ExpiryCleanupInterval should be defaulted")
 				assert.Equal(t, defaultEnqueueChannelBufferSize, cfg.EnqueueChannelBufferSize,
 					"EnqueueChannelBufferSize should be defaulted")
-				assert.Equal(t, time.Duration(0), cfg.DefaultRequestTTL, "DefaultRequestTTL should default to 0 (disabled)")
+				assert.Equal(t, defaultRequestTTL, cfg.DefaultRequestTTL, "DefaultRequestTTL should be defaulted when unset")
 			},
 		},
 		{

pkg/epp/flowcontrol/registry/config.go

@@ -39,9 +39,14 @@ const (
 )
 
 const (
-	// defaultPriorityBandMaxBytes is the default global capacity for a priority band if not explicitly configured.
-	// It is set to 1 GB.
+	// defaultPriorityBandMaxBytes is the default byte-size capacity for a priority band if not explicitly
+	// configured. It is set to 1 GB.
 	defaultPriorityBandMaxBytes uint64 = 1_000_000_000
+	// defaultPriorityBandMaxRequests is the default request-count capacity for a priority band if not
+	// explicitly configured. Together with the default request TTL (60s), it only binds when a band's
+	// arrival rate under a dispatch halt exceeds maxRequests/TTL (~83 req/s); below that, TTL expiry
+	// bounds occupancy first.
+	defaultPriorityBandMaxRequests uint64 = 5000
 	// defaultFlowGCTimeout is the default duration of inactivity after which an idle flow is garbage collected.
 	// This also serves as the interval for the periodic garbage collection scan.
 	defaultFlowGCTimeout time.Duration = 5 * time.Minute
@@ -85,14 +90,15 @@ type Config struct {
 	DefaultPriorityBand *PriorityBandConfig
 
 	// DefaultNegativePriorityBand is an optional template for dynamically provisioning priority bands when a request
-	// arrives with a priority level strictly below zero. This allows setting lower capacity limits (or zero) for
-	// negative-priority traffic to designate it as sheddable.
+	// arrives with a priority level strictly below zero. This allows setting lower capacity limits for
+	// negative-priority traffic to designate it as sheddable (a value of 0 is treated as unset and receives the
+	// default).
 	// If nil, negative priorities fall back to DefaultPriorityBand.
 	DefaultNegativePriorityBand *PriorityBandConfig
 
 	// FlowGCTimeout defines the interval at which the registry scans for and garbage collects idle flows.
 	// A flow is collected if it has been observed to be Idle for at least one full scan interval.
-	// Optional: Defaults to `defaultFlowGCTimeout` (1 hour).
+	// Optional: Defaults to `defaultFlowGCTimeout` (5 minutes).
 	FlowGCTimeout time.Duration
 
 	// PriorityBandGCTimeout defines the duration of inactivity after which a dynamically provisioned priority band
@@ -136,12 +142,15 @@ type PriorityBandConfig struct {
 	FairnessPolicy flowcontrol.FairnessPolicy
 
 	// MaxBytes defines the maximum total byte size for this priority band.
-	// Optional: Defaults to defaultPriorityBandMaxBytes (1 GB).
+	// Optional: Defaults to defaultPriorityBandMaxBytes (1 GB). A value of 0 is treated as unset and receives the
+	// default; per-band limits are always bounded, unlike the optional global limits. To effectively remove the
+	// bound, set an explicit large value.
 	MaxBytes uint64
 
 	// MaxRequests defines the maximum total request count for this priority band.
-	// A value of 0 signifies no request-count limit is enforced.
-	// Optional: Defaults to defaultPriorityBandMaxRequests (5000).
+	// Optional: Defaults to defaultPriorityBandMaxRequests (5000). A value of 0 is treated as unset and receives
+	// the default; per-band limits are always bounded, unlike the optional global limits. To effectively remove
+	// the bound, set an explicit large value.
 	MaxRequests uint64
 }
 
@@ -386,6 +395,9 @@ func (p *PriorityBandConfig) applyDefaults(defaults PriorityBandPolicyDefaults)
 	if p.MaxBytes == 0 {
 		p.MaxBytes = defaultPriorityBandMaxBytes
 	}
+	if p.MaxRequests == 0 {
+		p.MaxRequests = defaultPriorityBandMaxRequests
+	}
 	if p.FairnessPolicy == nil {
 		if defaults.FairnessPolicy == nil {
 			return fmt.Errorf("no default fairness policy provided and none set on band %d", p.Priority)

pkg/epp/flowcontrol/registry/config_test.go

@@ -88,6 +88,7 @@ func TestNewConfig(t *testing.T) {
 				require.NotNil(t, band.FairnessPolicy)
 				assert.Equal(t, DefaultFairnessPolicyRef, band.FairnessPolicy.TypedName().Name)
 				assert.Equal(t, defaultPriorityBandMaxBytes, band.MaxBytes)
+				assert.Equal(t, defaultPriorityBandMaxRequests, band.MaxRequests)
 			},
 		},
 		{
@@ -129,9 +130,17 @@ func TestNewConfig(t *testing.T) {
 			assertion: func(t *testing.T, cfg *Config) {
 				assert.Len(t, cfg.PriorityBands, 1, "PriorityBands should contain only the always-injected priority 0")
 				assert.Contains(t, cfg.PriorityBands, 0, "PriorityBands should contain priority 0")
+				assert.Equal(t, defaultPriorityBandMaxBytes, cfg.PriorityBands[0].MaxBytes,
+					"Auto-provisioned priority 0 band should receive the byte-size default")
+				assert.Equal(t, defaultPriorityBandMaxRequests, cfg.PriorityBands[0].MaxRequests,
+					"Auto-provisioned priority 0 band should receive the request-count default")
 				require.NotNil(t, cfg.DefaultPriorityBand, "DefaultPriorityBand template must be initialized")
 				assert.NotNil(t, cfg.DefaultPriorityBand.FairnessPolicy)
 				assert.Equal(t, DefaultFairnessPolicyRef, cfg.DefaultPriorityBand.FairnessPolicy.TypedName().Name)
+				assert.Equal(t, defaultPriorityBandMaxBytes, cfg.DefaultPriorityBand.MaxBytes,
+					"Dynamic provisioning template should receive the byte-size default")
+				assert.Equal(t, defaultPriorityBandMaxRequests, cfg.DefaultPriorityBand.MaxRequests,
+					"Dynamic provisioning template should receive the request-count default")
 			},
 		},
 		{
@@ -372,15 +381,16 @@ func TestNewConfig_DefaultNegativePriorityBand(t *testing.T) {
 			"Defaults should be applied to DefaultNegativePriorityBand")
 	})
 
-	t.Run("ShouldAllowZeroMaxBytes_ForSheddableTraffic", func(t *testing.T) {
+	t.Run("ShouldApplyCapacityDefaults_ToEmptyNegativeBandTemplate", func(t *testing.T) {
 		t.Parallel()
 		cfg, err := NewConfig(defaults,
 			WithDefaultNegativePriorityBand(&PriorityBandConfig{}),
 		)
 		require.NoError(t, err)
 		require.NotNil(t, cfg.DefaultNegativePriorityBand)
-		// MaxBytes=0 gets defaulted to 1GB via applyDefaults
+		// Zero capacity values are treated as unset: bands are always bounded, never zero-capacity.
 		assert.Equal(t, defaultPriorityBandMaxBytes, cfg.DefaultNegativePriorityBand.MaxBytes)
+		assert.Equal(t, defaultPriorityBandMaxRequests, cfg.DefaultNegativePriorityBand.MaxRequests)
 	})
 
 	t.Run("ShouldCloneNegativeBandTemplate", func(t *testing.T) {

pkg/epp/requestcontrol/admission.go

@@ -194,7 +194,13 @@ func (r *flowControlRequest) ID() string {
 	}
 	return r.inferenceRequest.RequestID
 }
-func (r *flowControlRequest) InitialEffectiveTTL() time.Duration { return 0 } // Use controller default.
+
+// InitialEffectiveTTL returns 0 to defer to the controller-level default TTL, which is therefore the only TTL
+// source for every request today.
+// TODO(https://github.com/llm-d/llm-d-router/issues/1090): plumb more specific TTL scopes and resolve
+// most-specific-wins: per-request (clamped), then per-band (effectively priority band config, eventually
+// codifiable in the InferenceObjective CRD), then the controller default.
+func (r *flowControlRequest) InitialEffectiveTTL() time.Duration { return 0 }
 func (r *flowControlRequest) ByteSize() uint64                   { return r.requestByteSize }
 
 func (r *flowControlRequest) InferenceRequest() *scheduling.InferenceRequest {