feat: Add support for upstream PreconnectPolicy

api/v1alpha1/connection_types.go

@@ -70,6 +70,43 @@ type BackendConnection struct {
 	// +optional
 	// +notImplementedHide
 	SocketBufferLimit *resource.Quantity `json:"socketBufferLimit,omitempty"`
+
+	// Preconnect configures proactive upstream connections to reduce latency by establishing
+	// connections before they’re needed and avoiding connection establishment overhead.
+	//
+	// If unset, Envoy will fetch connections as needed to serve in-flight requests.
+	//
+	// +optional
+	Preconnect *PreconnectPolicy `json:"preconnect,omitempty"`
+}
+
+// Preconnect configures proactive upstream connections to avoid
+// connection establishment overhead and reduce latency.
+type PreconnectPolicy struct {
+	// PerUpstreamPercent configures how many additional connections to maintain per
+	// upstream endpoint, useful for high-QPS or latency sensitive services. Expressed as a
+	// percentage over the connections required by active streams
+	// (e.g. 0 = preconnect disabled 5 = 1.05x connections per-upstream, 100 = 2.00×).
+	//
+	// Maximum allowed is 200. When both PerUpstreamPercent and
+	// PredictivePercent are set, Envoy ensures both are satisfied (max of the two).
+	//
+	// +kubebuilder:validation:Maximum=200
+	// +optional
+	PerUpstreamPercent *uint32 `json:"perUpstreamPercent,omitempty"`
+
+	// PredictivePercent configures how many additional connections to maintain
+	// across the cluster by anticipating which upstream endpoint the load balancer
+	// will select next, useful for low-QPS services. Relies on deterministic
+	// loadbalancing and is only supported with Random or RoundRobin.
+	// Expressed as a percentage over the connections required by active streams
+	// (e.g. 0 = preconnect disabled, 5 = 1.05× connections across the cluster, 100 = 2.00×).
+	//
+	// When both PerUpstreamPercent and PredictivePercent are
+	// set Envoy ensures both are satisfied per host (max of the two).
+	//
+	// +optional
+	PredictivePercent *uint32 `json:"predictivePercent,omitempty"`
 }
 
 type ConnectionLimit struct {

api/v1alpha1/shared_types.go

@@ -585,6 +585,8 @@ type BackendCluster struct {
 
 // ClusterSettings provides the various knobs that can be set to control how traffic to a given
 // backend will be configured.
+//
+// +kubebuilder:validation:XValidation:rule="!((has(self.connection) && has(self.connection.preconnect) && has(self.connection.preconnect.predictivePercent)) && !(has(self.loadBalancer) && has(self.loadBalancer.type) && self.loadBalancer.type in ['Random', 'RoundRobin']))",message="predictivePercent in preconnect policy only works with RoundRobin or Random load balancers"
 type ClusterSettings struct {
 	// LoadBalancer policy to apply when routing traffic from the gateway to
 	// the backend endpoints. Defaults to `LeastRequest`.

charts/gateway-crds-helm/templates/generated/gateway.envoyproxy.io_backendtrafficpolicies.yaml

@@ -191,6 +191,39 @@ spec:
                       For example, 20Mi, 1Gi, 256Ki etc.
                       Note: that when the suffix is not provided, the value is interpreted as bytes.
                     x-kubernetes-int-or-string: true
+                  preconnect:
+                    description: |-
+                      Preconnect configures proactive upstream connections to reduce latency by establishing
+                      connections before they’re needed and avoiding connection establishment overhead.
+
+                      If unset, Envoy will fetch connections as needed to serve in-flight requests.
+                    properties:
+                      perUpstreamPercent:
+                        description: |-
+                          PerUpstreamPercent configures how many additional connections to maintain per
+                          upstream endpoint, useful for high-QPS or latency sensitive services. Expressed as a
+                          percentage over the connections required by active streams
+                          (e.g. 0 = preconnect disabled 5 = 1.05x connections per-upstream, 100 = 2.00×).
+
+                          Maximum allowed is 200. When both PerUpstreamPercent and
+                          PredictivePercent are set, Envoy ensures both are satisfied (max of the two).
+                        format: int32
+                        maximum: 200
+                        type: integer
+                      predictivePercent:
+                        description: |-
+                          PredictivePercent configures how many additional connections to maintain
+                          across the cluster by anticipating which upstream endpoint the load balancer
+                          will select next, useful for low-QPS services. Relies on deterministic
+                          loadbalancing and is only supported with Random or RoundRobin.
+                          Expressed as a percentage over the connections required by active streams
+                          (e.g. 0 = preconnect disabled, 5 = 1.05× connections across the cluster, 100 = 2.00×).
+
+                          When both PerUpstreamPercent and PredictivePercent are
+                          set Envoy ensures both are satisfied per host (max of the two).
+                        format: int32
+                        type: integer
+                    type: object
                   socketBufferLimit:
                     allOf:
                     - pattern: ^(\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))))?$
@@ -2265,6 +2298,12 @@ spec:
                 : true'
             - message: either compression or compressor can be set, not both
               rule: '!has(self.compression) || !has(self.compressor)'
+            - message: predictivePercent in preconnect policy only works with RoundRobin
+                or Random load balancers
+              rule: '!((has(self.connection) && has(self.connection.preconnect) &&
+                has(self.connection.preconnect.predictivePercent)) && !(has(self.loadBalancer)
+                && has(self.loadBalancer.type) && self.loadBalancer.type in [''Random'',
+                ''RoundRobin'']))'
           status:
             description: status defines the current status of BackendTrafficPolicy.
             properties:

charts/gateway-crds-helm/templates/generated/gateway.envoyproxy.io_envoyextensionpolicies.yaml

@@ -302,6 +302,39 @@ spec:
                                 For example, 20Mi, 1Gi, 256Ki etc.
                                 Note: that when the suffix is not provided, the value is interpreted as bytes.
                               x-kubernetes-int-or-string: true
+                            preconnect:
+                              description: |-
+                                Preconnect configures proactive upstream connections to reduce latency by establishing
+                                connections before they’re needed and avoiding connection establishment overhead.
+
+                                If unset, Envoy will fetch connections as needed to serve in-flight requests.
+                              properties:
+                                perUpstreamPercent:
+                                  description: |-
+                                    PerUpstreamPercent configures how many additional connections to maintain per
+                                    upstream endpoint, useful for high-QPS or latency sensitive services. Expressed as a
+                                    percentage over the connections required by active streams
+                                    (e.g. 0 = preconnect disabled 5 = 1.05x connections per-upstream, 100 = 2.00×).
+
+                                    Maximum allowed is 200. When both PerUpstreamPercent and
+                                    PredictivePercent are set, Envoy ensures both are satisfied (max of the two).
+                                  format: int32
+                                  maximum: 200
+                                  type: integer
+                                predictivePercent:
+                                  description: |-
+                                    PredictivePercent configures how many additional connections to maintain
+                                    across the cluster by anticipating which upstream endpoint the load balancer
+                                    will select next, useful for low-QPS services. Relies on deterministic
+                                    loadbalancing and is only supported with Random or RoundRobin.
+                                    Expressed as a percentage over the connections required by active streams
+                                    (e.g. 0 = preconnect disabled, 5 = 1.05× connections across the cluster, 100 = 2.00×).
+
+                                    When both PerUpstreamPercent and PredictivePercent are
+                                    set Envoy ensures both are satisfied per host (max of the two).
+                                  format: int32
+                                  type: integer
+                              type: object
                             socketBufferLimit:
                               allOf:
                               - pattern: ^(\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))(([KMGTPE]i)|[numkMGTPE]|([eE](\+|-)?(([0-9]+(\.[0-9]*)?)|(\.[0-9]+))))?$
@@ -1059,6 +1092,13 @@ spec:
                               type: object
                           type: object
                       type: object
+                      x-kubernetes-validations:
+                      - message: predictivePercent in preconnect policy only works
+                          with RoundRobin or Random load balancers
+                        rule: '!((has(self.connection) && has(self.connection.preconnect)
+                          && has(self.connection.preconnect.predictivePercent)) &&
+                          !(has(self.loadBalancer) && has(self.loadBalancer.type)
+                          && self.loadBalancer.type in [''Random'', ''RoundRobin'']))'
                     failOpen:
                       default: false
                       description: |-