Fix a cluster-agent crash at startup (#28282)

DataDog · Aug 8, 2024 · d3901a2 · d3901a2
1 parent 1a835ad
commit d3901a2
Show file tree

Hide file tree

Showing 3 changed files with 18 additions and 19 deletions.
diff --git a/cmd/cluster-agent/subcommands/start/command.go b/cmd/cluster-agent/subcommands/start/command.go
@@ -324,7 +324,7 @@ func start(log log.Component,
 		DatadogClient:          dc,
 	}
 
-	if aggErr := controllers.StartControllers(ctx); aggErr != nil {
+	if aggErr := controllers.StartControllers(&ctx); aggErr != nil {
 		for _, err := range aggErr.Errors() {
 			pkglog.Warnf("Error while starting controller: %v", err)
 		}

diff --git a/flakes.yaml b/flakes.yaml
@@ -12,10 +12,6 @@ test/new-e2e/tests/containers:
   - TestEKSSuite/TestCPU/metric___container.cpu.usage{^kube_deployment:stress-ng$,^kube_namespace:workload-cpustress$}
   - TestKindSuite/TestCPU/metric___container.cpu.usage{^kube_deployment:stress-ng$,^kube_namespace:workload-cpustress$}
   - TestECSSuite
-  - TestEKSSuite/Test00UpAndRunning/agent_pods_are_ready_and_not_restarting
-  - TestEKSSuite/TestZZUpAndRunning/agent_pods_are_ready_and_not_restarting
-  - TestKindSuite/Test00UpAndRunning/agent_pods_are_ready_and_not_restarting
-  - TestKindSuite/TestZZUpAndRunning/agent_pods_are_ready_and_not_restarting
 
 test/new-e2e/tests/installer:
   - TestPackages/upgrade_scenario_ubuntu_22_04_x86_64/TestUpgradeSuccessful
diff --git a/pkg/util/kubernetes/apiserver/controllers/controllers.go b/pkg/util/kubernetes/apiserver/controllers/controllers.go
@@ -33,7 +33,7 @@ const autoscalerNowHandleMsgEvent = "Autoscaler is now handled by the Cluster-Ag
 
 var errIsEmpty = errors.New("entity is empty") //nolint:revive
 
-type startFunc func(ControllerContext, chan error)
+type startFunc func(*ControllerContext, chan error)
 
 type controllerFuncs struct {
 	enabled func() bool
@@ -64,6 +64,7 @@ var controllerCatalog = map[controllerName]controllerFuncs{
 // ControllerContext holds all the attributes needed by the controllers
 type ControllerContext struct {
 	informers              map[apiserver.InformerName]cache.SharedInformer
+	informersMutex         sync.Mutex
 	InformerFactory        informers.SharedInformerFactory
 	DynamicClient          dynamic.Interface
 	DynamicInformerFactory dynamicinformer.DynamicSharedInformerFactory
@@ -77,7 +78,7 @@ type ControllerContext struct {
 
 // StartControllers runs the enabled Kubernetes controllers for the Datadog Cluster Agent. This is
 // only called once, when we have confirmed we could correctly connect to the API server.
-func StartControllers(ctx ControllerContext) k8serrors.Aggregate {
+func StartControllers(ctx *ControllerContext) k8serrors.Aggregate {
 	ctx.informers = make(map[apiserver.InformerName]cache.SharedInformer)
 
 	var wg sync.WaitGroup
@@ -126,9 +127,7 @@ func StartControllers(ctx ControllerContext) k8serrors.Aggregate {
 
 // startMetadataController starts the informers needed for metadata collection.
 // The synchronization of the informers is handled by the controller.
-//
-//nolint:revive // TODO(CAPP) Fix revive linter
-func startMetadataController(ctx ControllerContext, c chan error) {
+func startMetadataController(ctx *ControllerContext, _ chan error) {
 	metaController := newMetadataController(
 		ctx.InformerFactory.Core().V1().Endpoints(),
 		ctx.WorkloadMeta,
@@ -138,7 +137,7 @@ func startMetadataController(ctx ControllerContext, c chan error) {
 
 // startAutoscalersController starts the informers needed for autoscaling.
 // The synchronization of the informers is handled by the controller.
-func startAutoscalersController(ctx ControllerContext, c chan error) {
+func startAutoscalersController(ctx *ControllerContext, c chan error) {
 	var err error
 	if ctx.DatadogClient == nil {
 		c <- fmt.Errorf("datadog client is nil")
@@ -166,15 +165,19 @@ func startAutoscalersController(ctx ControllerContext, c chan error) {
 }
 
 // registerServicesInformer registers the services informer.
-//
-//nolint:revive // TODO(CAPP) Fix revive linter
-func registerServicesInformer(ctx ControllerContext, c chan error) {
-	ctx.informers[servicesInformer] = ctx.InformerFactory.Core().V1().Services().Informer()
+func registerServicesInformer(ctx *ControllerContext, _ chan error) {
+	informer := ctx.InformerFactory.Core().V1().Services().Informer()
+
+	ctx.informersMutex.Lock()
+	ctx.informers[servicesInformer] = informer
+	ctx.informersMutex.Unlock()
 }
 
 // registerEndpointsInformer registers the endpoints informer.
-//
-//nolint:revive // TODO(CAPP) Fix revive linter
-func registerEndpointsInformer(ctx ControllerContext, c chan error) {
-	ctx.informers[endpointsInformer] = ctx.InformerFactory.Core().V1().Endpoints().Informer()
+func registerEndpointsInformer(ctx *ControllerContext, _ chan error) {
+	informer := ctx.InformerFactory.Core().V1().Endpoints().Informer()
+
+	ctx.informersMutex.Lock()
+	ctx.informers[endpointsInformer] = informer
+	ctx.informersMutex.Unlock()
 }