Consider combining the xds-translator and the xds-server runners into 1 runner

[arkodg]

Description:

Describe the issue.

• The xds-translator runners subscribes to Xds IR and translates it to Xds https://github.com/envoyproxy/gateway/blob/main/internal/xds/translator/runner/runner.go
• The xds-server runner subscribes to Xds and services Xds requests from Envoy clients https://github.com/envoyproxy/gateway/blob/main/internal/xds/server/runner/runner.go

these two runners can be combined into 1 to reduce the extra layer of Xds memory saved in watchable

gateway/internal/message/types.go

Line 126 in 847d554

// Xds message

[optional Relevant Links:]

Any extra documentation required to understand the issue.

[arkodg]

wdyt @envoyproxy/gateway-maintainers ?

[zhaohuabing]

@arkodg the intention is to save the memory for the xds as we don't have to publish and subscript to the generated xds? Do you have an estimate of how much memory we can save by merging these two runners? I suggest finding the memory bottle first before making an optimizing decision.

[arkodg]

yes, the goal is to eliminate the Xds watchable layer. the optimization will be visible in the benchmarking report, the effort here is minimal

[qicz]

IMO needs to split the resources too. some resources for subscription like gateway API resources and EG APIs, but others are just kept in memory. like follows:

for subscription:

type WatchableResources struct {
	// This field is only used for marshalling/unmarshalling purposes and is not used by
	// the translator

	// EnvoyProxyForGatewayClass holds EnvoyProxy attached to GatewayClass
	EnvoyProxyForGatewayClass *egv1a1.EnvoyProxy `json:"envoyProxyForGatewayClass,omitempty" yaml:"envoyProxyForGatewayClass,omitempty"`
	// EnvoyProxiesForGateways holds EnvoyProxiesForGateways attached to Gateways
	EnvoyProxiesForGateways []*egv1a1.EnvoyProxy `json:"envoyProxiesForGateways,omitempty" yaml:"envoyProxiesForGateways,omitempty"`

	GatewayClass            *gwapiv1.GatewayClass          `json:"gatewayClass,omitempty" yaml:"gatewayClass,omitempty"`
	Gateways                []*gwapiv1.Gateway             `json:"gateways,omitempty" yaml:"gateways,omitempty"`
	HTTPRoutes              []*gwapiv1.HTTPRoute           `json:"httpRoutes,omitempty" yaml:"httpRoutes,omitempty"`
	GRPCRoutes              []*gwapiv1.GRPCRoute           `json:"grpcRoutes,omitempty" yaml:"grpcRoutes,omitempty"`
	TLSRoutes               []*gwapiv1a2.TLSRoute          `json:"tlsRoutes,omitempty" yaml:"tlsRoutes,omitempty"`
	TCPRoutes               []*gwapiv1a2.TCPRoute          `json:"tcpRoutes,omitempty" yaml:"tcpRoutes,omitempty"`
	UDPRoutes               []*gwapiv1a2.UDPRoute          `json:"udpRoutes,omitempty" yaml:"udpRoutes,omitempty"`
	ReferenceGrants         []*gwapiv1b1.ReferenceGrant    `json:"referenceGrants,omitempty" yaml:"referenceGrants,omitempty"`
	
	ExtensionRefFilters     []unstructured.Unstructured    `json:"extensionRefFilters,omitempty" yaml:"extensionRefFilters,omitempty"`
	EnvoyPatchPolicies      []*egv1a1.EnvoyPatchPolicy     `json:"envoyPatchPolicies,omitempty" yaml:"envoyPatchPolicies,omitempty"`
	ClientTrafficPolicies   []*egv1a1.ClientTrafficPolicy  `json:"clientTrafficPolicies,omitempty" yaml:"clientTrafficPolicies,omitempty"`
	BackendTrafficPolicies  []*egv1a1.BackendTrafficPolicy `json:"backendTrafficPolicies,omitempty" yaml:"backendTrafficPolicies,omitempty"`
	SecurityPolicies        []*egv1a1.SecurityPolicy       `json:"securityPolicies,omitempty" yaml:"securityPolicies,omitempty"`
	BackendTLSPolicies      []*gwapiv1a3.BackendTLSPolicy  `json:"backendTLSPolicies,omitempty" yaml:"backendTLSPolicies,omitempty"`
	EnvoyExtensionPolicies  []*egv1a1.EnvoyExtensionPolicy `json:"envoyExtensionPolicies,omitempty" yaml:"envoyExtensionPolicies,omitempty"`
	ExtensionServerPolicies []unstructured.Unstructured    `json:"extensionServerPolicies,omitempty" yaml:"extensionServerPolicies,omitempty"`
	Backends                []*egv1a1.Backend              `json:"backends,omitempty" yaml:"backends,omitempty"`
}

just in memory:(the Secret, EndpointSlice that too large bad for watable compare)

type InMemoryResources struct {
	
	Namespaces              []*corev1.Namespace            `json:"namespaces,omitempty" yaml:"namespaces,omitempty"`
	Services                []*corev1.Service              `json:"services,omitempty" yaml:"services,omitempty"`
	ServiceImports          []*mcsapiv1a1.ServiceImport    `json:"serviceImports,omitempty" yaml:"serviceImports,omitempty"`
	EndpointSlices          []*discoveryv1.EndpointSlice   `json:"endpointSlices,omitempty" yaml:"endpointSlices,omitempty"`
	Secrets                 []*corev1.Secret               `json:"secrets,omitempty" yaml:"secrets,omitempty"`
	ConfigMaps              []*corev1.ConfigMap            `json:"configMaps,omitempty" yaml:"configMaps,omitempty"`
}

cc @arkodg

[shawnh2]

Like @qicz 's idea.

One possible way to reduce memory is by triming unused field for these resources / CRD. And this achievable via the Transform func defined in client-go,

https://github.com/kubernetes/client-go/blob/v0.31.1/informers/factory.go#L104

for controller-runtime is in

https://github.com/kubernetes-sigs/controller-runtime/blob/main/pkg/cache/cache.go#L207

[arkodg]

@shawnh2 @qicz can we use a separate GH issue to discuss this additional design that could increase memory savings ?

[shawnh2]

+1 on this!

From heap flame graph in #3698, combining the xds-translator and the xds-server runners into 1 runner will prevent lots of deepcopy for message.Xds, which will result in saving ~30% memory.