xdsclient: update watcher API as per gRFC A88

internal/xds/bootstrap/bootstrap.go

@@ -206,7 +206,7 @@ func (sc *ServerConfig) ServerFeatures() []string {
 //
 // This feature controls the behavior of the xDS client when the server deletes
 // a previously sent Listener or Cluster resource. If set, the xDS client will
-// not invoke the watchers' OnResourceDoesNotExist() method when a resource is
+// not invoke the watchers' ResourceError() method when a resource is
 // deleted, nor will it remove the existing resource value from its cache.
 func (sc *ServerConfig) ServerFeaturesIgnoreResourceDeletion() bool {
 	for _, sf := range sc.serverFeatures {

xds/csds/csds_e2e_test.go

@@ -71,49 +71,49 @@ func Test(t *testing.T) {
 
 type nopListenerWatcher struct{}
 
-func (nopListenerWatcher) OnUpdate(_ *xdsresource.ListenerResourceData, onDone xdsresource.OnDoneFunc) {
+func (nopListenerWatcher) ResourceChanged(_ *xdsresource.ListenerResourceData, onDone func()) {
 	onDone()
 }
-func (nopListenerWatcher) OnError(_ error, onDone xdsresource.OnDoneFunc) {
+func (nopListenerWatcher) ResourceError(_ error, onDone func()) {
 	onDone()
 }
-func (nopListenerWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (nopListenerWatcher) AmbientError(_ error, onDone func()) {
 	onDone()
 }
 
 type nopRouteConfigWatcher struct{}
 
-func (nopRouteConfigWatcher) OnUpdate(_ *xdsresource.RouteConfigResourceData, onDone xdsresource.OnDoneFunc) {
+func (nopRouteConfigWatcher) ResourceChanged(_ *xdsresource.RouteConfigResourceData, onDone func()) {
 	onDone()
 }
-func (nopRouteConfigWatcher) OnError(_ error, onDone xdsresource.OnDoneFunc) {
+func (nopRouteConfigWatcher) ResourceError(_ error, onDone func()) {
 	onDone()
 }
-func (nopRouteConfigWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (nopRouteConfigWatcher) AmbientError(_ error, onDone func()) {
 	onDone()
 }
 
 type nopClusterWatcher struct{}
 
-func (nopClusterWatcher) OnUpdate(_ *xdsresource.ClusterResourceData, onDone xdsresource.OnDoneFunc) {
+func (nopClusterWatcher) ResourceChanged(_ *xdsresource.ClusterResourceData, onDone func()) {
 	onDone()
 }
-func (nopClusterWatcher) OnError(_ error, onDone xdsresource.OnDoneFunc) {
+func (nopClusterWatcher) ResourceError(_ error, onDone func()) {
 	onDone()
 }
-func (nopClusterWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (nopClusterWatcher) AmbientError(_ error, onDone func()) {
 	onDone()
 }
 
 type nopEndpointsWatcher struct{}
 
-func (nopEndpointsWatcher) OnUpdate(_ *xdsresource.EndpointsResourceData, onDone xdsresource.OnDoneFunc) {
+func (nopEndpointsWatcher) ResourceChanged(_ *xdsresource.EndpointsResourceData, onDone func()) {
 	onDone()
 }
-func (nopEndpointsWatcher) OnError(_ error, onDone xdsresource.OnDoneFunc) {
+func (nopEndpointsWatcher) ResourceError(_ error, onDone func()) {
 	onDone()
 }
-func (nopEndpointsWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (nopEndpointsWatcher) AmbientError(_ error, onDone func()) {
 	onDone()
 }
 
@@ -127,31 +127,31 @@ func (nopEndpointsWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc)
 // for ADS stream level flow control), and was causing CSDS to not receive any
 // updates from the xDS client.
 type blockingListenerWatcher struct {
-	testCtxDone <-chan struct{}             // Closed when the test is done.
-	onDoneCh    chan xdsresource.OnDoneFunc // Channel to write the onDone callback to.
+	testCtxDone <-chan struct{} // Closed when the test is done.
+	onDoneCh    chan func()     // Channel to write the onDone callback to.
 }
 
 func newBlockingListenerWatcher(testCtxDone <-chan struct{}) *blockingListenerWatcher {
 	return &blockingListenerWatcher{
 		testCtxDone: testCtxDone,
-		onDoneCh:    make(chan xdsresource.OnDoneFunc, 1),
+		onDoneCh:    make(chan func(), 1),
 	}
 }
 
-func (w *blockingListenerWatcher) OnUpdate(_ *xdsresource.ListenerResourceData, onDone xdsresource.OnDoneFunc) {
+func (w *blockingListenerWatcher) ResourceChanged(_ *xdsresource.ListenerResourceData, onDone func()) {
 	writeOnDone(w.testCtxDone, w.onDoneCh, onDone)
 }
-func (w *blockingListenerWatcher) OnError(_ error, onDone xdsresource.OnDoneFunc) {
+func (w *blockingListenerWatcher) ResourceError(_ error, onDone func()) {
 	writeOnDone(w.testCtxDone, w.onDoneCh, onDone)
 }
-func (w *blockingListenerWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (w *blockingListenerWatcher) AmbientError(_ error, onDone func()) {
 	writeOnDone(w.testCtxDone, w.onDoneCh, onDone)
 }
 
 // writeOnDone attempts to write the onDone callback on the onDone channel. It
 // returns when it can successfully write to the channel or when the test is
 // done, which is signalled by testCtxDone being closed.
-func writeOnDone(testCtxDone <-chan struct{}, onDoneCh chan xdsresource.OnDoneFunc, onDone xdsresource.OnDoneFunc) {
+func writeOnDone(testCtxDone <-chan struct{}, onDoneCh chan func(), onDone func()) {
 	select {
 	case <-testCtxDone:
 	case onDoneCh <- onDone:

xds/internal/balancer/cdsbalancer/cdsbalancer.go

@@ -542,38 +542,27 @@ func (b *cdsBalancer) onClusterUpdate(name string, update xdsresource.ClusterUpd
 	}
 }
 
-// Handles an error Cluster update from the xDS client. Propagates the error
-// down to the child policy if one exists, or puts the channel in
-// TRANSIENT_FAILURE.
+// Handles an ambient error Cluster update from the xDS client to not stop
+// using the previously seen resource.
 //
 // Only executed in the context of a serializer callback.
-func (b *cdsBalancer) onClusterError(name string, err error) {
-	b.logger.Warningf("Cluster resource %q received error update: %v", name, err)
+func (b *cdsBalancer) onClusterAmbientError(name string, err error) {
+	b.logger.Warningf("Cluster resource %q received ambient error update: %v", name, err)
 
-	if b.childLB != nil {
-		if xdsresource.ErrType(err) != xdsresource.ErrorTypeConnection {
-			// Connection errors will be sent to the child balancers directly.
-			// There's no need to forward them.
-			b.childLB.ResolverError(err)
-		}
-	} else {
-		// If child balancer was never created, fail the RPCs with
-		// errors.
-		b.ccw.UpdateState(balancer.State{
-			ConnectivityState: connectivity.TransientFailure,
-			Picker:            base.NewErrPicker(fmt.Errorf("%q: %v", name, err)),
-		})
+	if xdsresource.ErrType(err) != xdsresource.ErrorTypeConnection && b.childLB != nil {
+		// Connection errors will be sent to the child balancers directly.
+		// There's no need to forward them.
+		b.childLB.ResolverError(err)
 	}
 }
 
-// Handles a resource-not-found error from the xDS client. Propagates the error
-// down to the child policy if one exists, or puts the channel in
-// TRANSIENT_FAILURE.
+// Handles an error Cluster update from the xDS client to stop using the
+// previously seen resource. Propagates the error down to the child policy
+// if one exists, and puts the channel in TRANSIENT_FAILURE.
 //
 // Only executed in the context of a serializer callback.
-func (b *cdsBalancer) onClusterResourceNotFound(name string) {
-	b.logger.Warningf("CDS watch for resource %q reported resource-does-not-exist error", name)
-	err := b.annotateErrorWithNodeID(xdsresource.NewErrorf(xdsresource.ErrorTypeResourceNotFound, "cluster %q not found", name))
+func (b *cdsBalancer) onClusterResourceError(name string, err error) {
+	b.logger.Warningf("CDS watch for resource %q reported resource error", name)
 	b.closeChildPolicyAndReportTF(err)
 }
 
@@ -677,6 +666,14 @@ func (b *cdsBalancer) generateDMsForCluster(name string, depth int, dms []cluste
 	return append(dms, dm), true, nil
 }
 
+func (b *cdsBalancer) onClusterError(name string, err error) {
+	if b.childLB != nil {
+		b.onClusterAmbientError(name, err)
+	} else {
+		b.onClusterResourceError(name, err)
+	}
+}
+
 // ccWrapper wraps the balancer.ClientConn passed to the CDS balancer at
 // creation and intercepts the NewSubConn() and UpdateAddresses() call from the
 // child policy to add security configuration required by xDS credentials.

xds/internal/balancer/cdsbalancer/cdsbalancer_test.go

@@ -1003,7 +1003,7 @@ func (s) TestClusterUpdate_ResourceNotFound(t *testing.T) {
 
 	// Ensure RPC fails with Unavailable status code and the error message is
 	// meaningful and contains the xDS node ID.
-	wantErr := fmt.Sprintf("cluster %q not found", clusterName)
+	wantErr := fmt.Sprintf("resource %q of type %q has been removed", clusterName, "ClusterResource")
 	_, err := client.EmptyCall(ctx, &testpb.Empty{})
 	if err := verifyRPCError(err, codes.Unavailable, wantErr, nodeID); err != nil {
 		t.Fatal(err)

xds/internal/balancer/cdsbalancer/cluster_watcher.go

@@ -32,19 +32,19 @@ type clusterWatcher struct {
 	parent *cdsBalancer
 }
 
-func (cw *clusterWatcher) OnUpdate(u *xdsresource.ClusterResourceData, onDone xdsresource.OnDoneFunc) {
+func (cw *clusterWatcher) ResourceChanged(u *xdsresource.ClusterResourceData, onDone func()) {
 	handleUpdate := func(context.Context) { cw.parent.onClusterUpdate(cw.name, u.Resource); onDone() }
 	cw.parent.serializer.ScheduleOr(handleUpdate, onDone)
 }
 
-func (cw *clusterWatcher) OnError(err error, onDone xdsresource.OnDoneFunc) {
-	handleError := func(context.Context) { cw.parent.onClusterError(cw.name, err); onDone() }
-	cw.parent.serializer.ScheduleOr(handleError, onDone)
+func (cw *clusterWatcher) ResourceError(err error, onDone func()) {
+	handleResourceError := func(context.Context) { cw.parent.onClusterResourceError(cw.name, err); onDone() }
+	cw.parent.serializer.ScheduleOr(handleResourceError, onDone)
 }
 
-func (cw *clusterWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
-	handleNotFound := func(context.Context) { cw.parent.onClusterResourceNotFound(cw.name); onDone() }
-	cw.parent.serializer.ScheduleOr(handleNotFound, onDone)
+func (cw *clusterWatcher) AmbientError(err error, onDone func()) {
+	handleError := func(context.Context) { cw.parent.onClusterAmbientError(cw.name, err); onDone() }
+	cw.parent.serializer.ScheduleOr(handleError, onDone)
 }
 
 // watcherState groups the state associated with a clusterWatcher.

xds/internal/balancer/clusterresolver/e2e_test/balancer_test.go

@@ -279,7 +279,7 @@ func (s) TestErrorFromParentLB_ResourceNotFound(t *testing.T) {
 	}
 
 	// Ensure that RPCs start to fail with expected error.
-	wantErr := fmt.Sprintf("cluster %q not found", clusterName)
+	wantErr := fmt.Sprintf("resource %q of type %q has been removed", clusterName, "ClusterResource")
 	for ; ctx.Err() == nil; <-time.After(defaultTestShortTimeout) {
 		sCtx, sCancel := context.WithTimeout(ctx, defaultTestShortTimeout)
 		defer sCancel()

xds/internal/balancer/clusterresolver/resource_resolver.go

@@ -38,7 +38,7 @@ type resourceUpdate struct {
 	priorities []priorityConfig
 	// To be invoked once the update is completely processed, or is dropped in
 	// favor of a newer update.
-	onDone xdsresource.OnDoneFunc
+	onDone func()
 }
 
 // topLevelResolver is used by concrete endpointsResolver implementations for
@@ -50,7 +50,7 @@ type topLevelResolver interface {
 	// endpointsResolver implementation. The onDone callback is to be invoked
 	// once the update is completely processed, or is dropped in favor of a
 	// newer update.
-	onUpdate(onDone xdsresource.OnDoneFunc)
+	onUpdate(onDone func())
 }
 
 // endpointsResolver wraps the functionality to resolve a given resource name to
@@ -282,7 +282,7 @@ func (rr *resourceResolver) stop(closing bool) {
 // clusterresolver LB policy.
 //
 // Caller must hold rr.mu.
-func (rr *resourceResolver) generateLocked(onDone xdsresource.OnDoneFunc) {
+func (rr *resourceResolver) generateLocked(onDone func()) {
 	var ret []priorityConfig
 	for _, rDM := range rr.children {
 		u, ok := rDM.r.lastUpdate()
@@ -312,7 +312,7 @@ func (rr *resourceResolver) generateLocked(onDone xdsresource.OnDoneFunc) {
 	rr.updateChannel <- &resourceUpdate{priorities: ret, onDone: onDone}
 }
 
-func (rr *resourceResolver) onUpdate(onDone xdsresource.OnDoneFunc) {
+func (rr *resourceResolver) onUpdate(onDone func()) {
 	handleUpdate := func(context.Context) {
 		rr.mu.Lock()
 		rr.generateLocked(onDone)

xds/internal/balancer/clusterresolver/resource_resolver_eds.go

@@ -75,8 +75,8 @@
 	return ret
 }
 
-// OnUpdate is invoked to report an update for the resource being watched.
-func (er *edsDiscoveryMechanism) OnUpdate(update *xdsresource.EndpointsResourceData, onDone xdsresource.OnDoneFunc) {
+// ResourceChanged is invoked to report an update for the resource being watched.
+func (er *edsDiscoveryMechanism) ResourceChanged(update *xdsresource.EndpointsResourceData, onDone func()) {
 	if er.stopped.HasFired() {
 		onDone()
 		return
@@ -89,54 +89,36 @@
 	er.topLevelResolver.onUpdate(onDone)
 }
 
-func (er *edsDiscoveryMechanism) OnError(err error, onDone xdsresource.OnDoneFunc) {
+func (er *edsDiscoveryMechanism) ResourceError(err error, onDone func()) {
 	if er.stopped.HasFired() {
 		onDone()
 		return
 	}
 
 	if er.logger.V(2) {
-		er.logger.Infof("EDS discovery mechanism for resource %q reported error: %v", er.nameToWatch, err)
+		er.logger.Infof("EDS discovery mechanism for resource %q reported resource error: %v", er.nameToWatch, err)
 	}
 
-	er.mu.Lock()
-	if er.update != nil {
-		// Continue using a previously received good configuration if one
-		// exists.
-		er.mu.Unlock()
-		onDone()
-		return
-	}
-
-	// Else report an empty update that would result in no priority child being
+	// Report an empty update that would result in no priority child being
 	// created for this discovery mechanism. This would result in the priority
 	// LB policy reporting TRANSIENT_FAILURE (as there would be no priorities or
 	// localities) if this was the only discovery mechanism, or would result in
 	// the priority LB policy using a lower priority discovery mechanism when
 	// that becomes available.
+	er.mu.Lock()
 	er.update = &xdsresource.EndpointsUpdate{}
 	er.mu.Unlock()
 
 	er.topLevelResolver.onUpdate(onDone)
 }
 
-func (er *edsDiscoveryMechanism) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
+func (er *edsDiscoveryMechanism) AmbientError(err error, onDone func()) {
 	if er.stopped.HasFired() {
 		onDone()
 		return
 	}
 
-	er.logger.Warningf("EDS discovery mechanism for resource %q reported resource-does-not-exist error", er.nameToWatch)
-
-	// Report an empty update that would result in no priority child being
-	// created for this discovery mechanism. This would result in the priority
-	// LB policy reporting TRANSIENT_FAILURE (as there would be no priorities or
-	// localities) if this was the only discovery mechanism, or would result in
-	// the priority LB policy using a lower priority discovery mechanism when
-	// that becomes available.
-	er.mu.Lock()
-	er.update = &xdsresource.EndpointsUpdate{}
-	er.mu.Unlock()
-
-	er.topLevelResolver.onUpdate(onDone)
+	if er.logger.V(2) {
+		er.logger.Infof("EDS discovery mechanism for resource %q reported ambient error: %v", er.nameToWatch, err)
+	}
 }

xds/internal/resolver/serviceconfig.go

@@ -141,8 +141,8 @@ type erroringConfigSelector struct {
 	err error
 }
 
-func newErroringConfigSelector(xdsNodeID string) *erroringConfigSelector {
-	return &erroringConfigSelector{err: annotateErrorWithNodeID(status.Errorf(codes.Unavailable, "no valid clusters"), xdsNodeID)}
+func newErroringConfigSelector(err error, xdsNodeID string) *erroringConfigSelector {
+	return &erroringConfigSelector{err: annotateErrorWithNodeID(status.Errorf(codes.Unavailable, err.Error()), xdsNodeID)}
 }
 
 func (cs *erroringConfigSelector) SelectConfig(iresolver.RPCInfo) (*iresolver.RPCConfig, error) {

xds/internal/resolver/watch_service.go

@@ -36,19 +36,19 @@ func newListenerWatcher(resourceName string, parent *xdsResolver) *listenerWatch
 	return lw
 }
 
-func (l *listenerWatcher) OnUpdate(update *xdsresource.ListenerResourceData, onDone xdsresource.OnDoneFunc) {
+func (l *listenerWatcher) ResourceChanged(update *xdsresource.ListenerResourceData, onDone func()) {
 	handleUpdate := func(context.Context) { l.parent.onListenerResourceUpdate(update.Resource); onDone() }
 	l.parent.serializer.ScheduleOr(handleUpdate, onDone)
 }
 
-func (l *listenerWatcher) OnError(err error, onDone xdsresource.OnDoneFunc) {
+func (l *listenerWatcher) ResourceError(err error, onDone func()) {
 	handleError := func(context.Context) { l.parent.onListenerResourceError(err); onDone() }
 	l.parent.serializer.ScheduleOr(handleError, onDone)
 }
 
-func (l *listenerWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
-	handleNotFound := func(context.Context) { l.parent.onListenerResourceNotFound(); onDone() }
-	l.parent.serializer.ScheduleOr(handleNotFound, onDone)
+func (l *listenerWatcher) AmbientError(err error, onDone func()) {
+	handleError := func(context.Context) { l.parent.onListenerResourceAmbientError(err); onDone() }
+	l.parent.serializer.ScheduleOr(handleError, onDone)
 }
 
 func (l *listenerWatcher) stop() {
@@ -68,22 +68,22 @@ func newRouteConfigWatcher(resourceName string, parent *xdsResolver) *routeConfi
 	return rw
 }
 
-func (r *routeConfigWatcher) OnUpdate(u *xdsresource.RouteConfigResourceData, onDone xdsresource.OnDoneFunc) {
+func (r *routeConfigWatcher) ResourceChanged(u *xdsresource.RouteConfigResourceData, onDone func()) {
 	handleUpdate := func(context.Context) {
 		r.parent.onRouteConfigResourceUpdate(r.resourceName, u.Resource)
 		onDone()
 	}
 	r.parent.serializer.ScheduleOr(handleUpdate, onDone)
 }
 
-func (r *routeConfigWatcher) OnError(err error, onDone xdsresource.OnDoneFunc) {
+func (r *routeConfigWatcher) ResourceError(err error, onDone func()) {
 	handleError := func(context.Context) { r.parent.onRouteConfigResourceError(r.resourceName, err); onDone() }
 	r.parent.serializer.ScheduleOr(handleError, onDone)
 }
 
-func (r *routeConfigWatcher) OnResourceDoesNotExist(onDone xdsresource.OnDoneFunc) {
-	handleNotFound := func(context.Context) { r.parent.onRouteConfigResourceNotFound(r.resourceName); onDone() }
-	r.parent.serializer.ScheduleOr(handleNotFound, onDone)
+func (r *routeConfigWatcher) AmbientError(err error, onDone func()) {
+	handleError := func(context.Context) { r.parent.onRouteConfigResourceAmbientError(r.resourceName, err); onDone() }
+	r.parent.serializer.ScheduleOr(handleError, onDone)
 }
 
 func (r *routeConfigWatcher) stop() {