Cached stale Endpoints cannot be cleaned when Endpoints are filtered out by EndpointSlice conditions or TopologyAwareHints

[hongliangl]

Describe the bug
Enabled EndpointSlice feature gate in Antrea, failed on Kubernetes sig-network e2e test should create endpoints for unready pods

Root Cause
Assuming that there a Service whose PublishNotReadyAddresses is true, and there is a ready Endpoint of the Service.

• Make the Endpoint to be terminating and not serving
• Call

  antrea/pkg/agent/proxy/proxier.go

  Line 720 in 3b224c4

  func (p *proxier) OnEndpointSliceUpdate(oldEndpointSlice, newEndpointSlice *discovery.EndpointSlice) {

• Call

  antrea/pkg/agent/proxy/proxier.go

  Line 630 in 3b224c4

  func (p *proxier) syncProxyRules() {

  • In

    antrea/pkg/agent/proxy/proxier.go

    Line 418 in 3b224c4

    clusterEndpoints, localEndpoints, allReachableEndpoints, mcsLocalService := p.categorizeEndpoints(endpointsToInstall, svcInfo, svcPortName, serviceCIDRIPv4)

    , since the Endpoint is terminating and not serving, clusterEndpoints, localEndpoints and allReachableEndpoints are empty.
  • Corresponding OVS group will be updated in

    antrea/pkg/agent/proxy/proxier.go

    Line 503 in 3b224c4

    if err = p.ofClient.InstallServiceGroup(groupID, affinityTimeout != 0, mcsLocalService, allReachableEndpoints); err != nil {

    . The updated OVS group doesn't have any bucket since allReachableEndpoints is empty.
  • Note that, Endpoints for the Service in p.endpointsMap and p. endpointsInstalledMap are the same, as a result, when calling

    antrea/pkg/agent/proxy/proxier.go

    Line 246 in 3b224c4

    func (p *proxier) removeStaleEndpoints() {

    , stale Endpoint cannot be cleaned since the uninstalled Endpoints is still in p.endpointsMap.
  • When the Endpoint is ready, since there is a cache for the Endpoint, corresponding OVS group will not be updated, and the bucket for the Endpoint will not be synced to OVS.