Implement KFServing v1beta1 controller (kubeflow#1042)

* Move existing controller to v1alpha2

* Implement v1beta1 controller

* Fix test crd path

* Add controller test for predictor

* Setup webhook for v1beta1 and v1alpha2

* Fix webhook setup

* Add api conversion

* Patch conversion webhook to crd

* Fix the configmap decoding

* Fix v1beta1 controller test

* Address comments

* Add transformer

* Transformer support

* Fix validate create for http test

* Reconcile kfserving ingress

* Add license

* Add rbac for ingress and svc

* Add rbac for ingress and svc

* Fix ingress host

* Switch to use virtual service

* Add conversion test

* Add crd conversion test

* Add explainer

* More conversion test

* Add custom predictor defaulter test

* Don't use nested inline

* Fix transformer and explainer

* Set ingress condition

* Add ingress asserts

* Fix explainer route

* Update ksvc traffic targets based on canary traffic split

* e2e test to work with v1beta1

* skip gpu test

* Fix canary traffic pointer type

* Download yq

* Add nil check for spec what miss component implementation

* Expose podspec on component

* Support http url in v1beta1

* fix typo

* Update traffic on routespec when canaryPercent is changed

* Add canary rollout test

* Add explainer test

* Address comments

* Use tritonserver

* Make container name and probe port field optional

* Add traffic percent in status

* Set xgboost nthread

* Fix updates in test

* Use default service name for backwards compatibility

* Call knative defaulter to avoid diffs during reconcilation

* Set SessionAffinity

* Check if configured as internal domain

.github/workflows/go.yml

@@ -29,5 +29,7 @@ jobs:
     - name: Test
       run: |
         export GOPATH=/home/runner/go
-        export PATH=$PATH:/usr/local/kubebuilder/bin
+        export PATH=$PATH:/usr/local/kubebuilder/bin:/home/runner/go/bin
+        wget -O $GOPATH/bin/yq https://github.com/mikefarah/yq/releases/download/3.3.2/yq_linux_amd64
+        chmod +x $GOPATH/bin/yq
         make test

Makefile

@@ -95,7 +95,7 @@ undeploy-dev:
 
 # Generate manifests e.g. CRD, RBAC etc.
 manifests: controller-gen
-	$(CONTROLLER_GEN) $(CRD_OPTIONS) paths=./pkg/apis/serving/v1alpha2/... output:crd:dir=config/crd
+	$(CONTROLLER_GEN) $(CRD_OPTIONS) paths=./pkg/apis/serving/... output:crd:dir=config/crd
 	$(CONTROLLER_GEN) rbac:roleName=kfserving-manager-role paths=./pkg/controller/... output:rbac:artifacts:config=config/rbac
 	$(CONTROLLER_GEN) object:headerFile="hack/boilerplate.go.txt" paths=./pkg/apis/serving/v1alpha2
 	$(CONTROLLER_GEN) object:headerFile="hack/boilerplate.go.txt" paths=./pkg/apis/serving/v1beta1
@@ -105,6 +105,17 @@ manifests: controller-gen
 	perl -pi -e 's/Any/string/g' config/crd/serving.kubeflow.org_inferenceservices.yaml
 	perl -pi -e 's/storedVersions: null/storedVersions: []/g' config/crd/serving.kubeflow.org_trainedmodels.yaml
 	perl -pi -e 's/conditions: null/conditions: []/g' config/crd/serving.kubeflow.org_trainedmodels.yaml
+	perl -pi -e 's/Any/string/g' config/crd/serving.kubeflow.org_trainedmodels.yaml
+	#TODO v1beta1 crd openAPIV3Schema is too big and kubectl client side apply takes long time to do diffs, need to use k8s 1.18's server side apply
+	#https://kubernetes.io/blog/2020/04/01/kubernetes-1.18-feature-server-side-apply-beta-2/#what-is-server-side-apply
+	#remove the required property on framework as name field needs to be optional
+	yq d -i config/crd/serving.kubeflow.org_inferenceservices.yaml 'spec.versions[1].schema.openAPIV3Schema.properties.spec.properties.*.properties.*.required'
+	#knative does not allow setting port on liveness or readiness probe
+	yq d -i config/crd/serving.kubeflow.org_inferenceservices.yaml 'spec.versions[1].schema.openAPIV3Schema.properties.spec.properties.*.properties.*.properties.readinessProbe.properties.httpGet.required'
+	yq d -i config/crd/serving.kubeflow.org_inferenceservices.yaml 'spec.versions[1].schema.openAPIV3Schema.properties.spec.properties.*.properties.*.properties.livenessProbe.properties.httpGet.required'
+	yq d -i config/crd/serving.kubeflow.org_inferenceservices.yaml 'spec.versions[1].schema.openAPIV3Schema.properties.spec.properties.*.properties.*.properties.readinessProbe.properties.tcpSocket.required'
+	yq d -i config/crd/serving.kubeflow.org_inferenceservices.yaml 'spec.versions[1].schema.openAPIV3Schema.properties.spec.properties.*.properties.*.properties.livenessProbe.properties.tcpSocket.required'
+
 
 # Run go fmt against code
 fmt:

cmd/manager/main.go

@@ -20,10 +20,9 @@ import (
 	"flag"
 	"github.com/kubeflow/kfserving/pkg/apis/serving/v1alpha2"
 	"github.com/kubeflow/kfserving/pkg/apis/serving/v1beta1"
-	v1alph2controller "github.com/kubeflow/kfserving/pkg/controller/inferenceservice"
+	v1beta1controller "github.com/kubeflow/kfserving/pkg/controller/v1beta1/inferenceservice"
 	trainedmodelcontroller "github.com/kubeflow/kfserving/pkg/controller/v1beta1/trainedmodel"
 	"github.com/kubeflow/kfserving/pkg/controller/v1beta1/trainedmodel/reconcilers/modelconfig"
-	"github.com/kubeflow/kfserving/pkg/webhook/admission/inferenceservice"
 	"github.com/kubeflow/kfserving/pkg/webhook/admission/pod"
 	"istio.io/client-go/pkg/apis/networking/v1alpha3"
 	v1 "k8s.io/api/core/v1"
@@ -96,23 +95,29 @@ func main() {
 		os.Exit(1)
 	}
 
+	log.Info("Setting up core scheme")
+	if err := v1.AddToScheme(mgr.GetScheme()); err != nil {
+		log.Error(err, "unable to add Core APIs to scheme")
+		os.Exit(1)
+	}
+
 	// Setup all Controllers
-	setupLog.Info("Setting up v1alpha2 controller")
+	setupLog.Info("Setting up v1beta1 controller")
 	eventBroadcaster := record.NewBroadcaster()
 	clientSet, err := kubernetes.NewForConfig(mgr.GetConfig())
 	if err != nil {
 		setupLog.Error(err, "unable to create clientSet")
 		os.Exit(1)
 	}
 	eventBroadcaster.StartRecordingToSink(&typedcorev1.EventSinkImpl{Interface: clientSet.CoreV1().Events("")})
-	if err = (&v1alph2controller.InferenceServiceReconciler{
+	if err = (&v1beta1controller.InferenceServiceReconciler{
 		Client: mgr.GetClient(),
-		Log:    ctrl.Log.WithName("v1alpha2Controllers").WithName("InferenceService"),
+		Log:    ctrl.Log.WithName("v1beta1Controllers").WithName("InferenceService"),
 		Scheme: mgr.GetScheme(),
 		Recorder: eventBroadcaster.NewRecorder(
-			mgr.GetScheme(), v1.EventSource{Component: "v1alpha2Controllers"}),
+			mgr.GetScheme(), v1.EventSource{Component: "v1beta1Controllers"}),
 	}).SetupWithManager(mgr); err != nil {
-		setupLog.Error(err, "unable to create controller", "v1alpha2Controller", "InferenceService")
+		setupLog.Error(err, "unable to create controller", "v1beta1Controller", "InferenceService")
 		os.Exit(1)
 	}
 
@@ -136,8 +141,19 @@ func main() {
 
 	log.Info("registering webhooks to the webhook server")
 	hookServer.Register("/mutate-pods", &webhook.Admission{Handler: &pod.Mutator{}})
-	hookServer.Register("/validate-inferenceservices", &webhook.Admission{Handler: &inferenceservice.Validator{}})
-	hookServer.Register("/mutate-inferenceservices", &webhook.Admission{Handler: &inferenceservice.Defaulter{}})
+
+	if err = ctrl.NewWebhookManagedBy(mgr).
+		For(&v1alpha2.InferenceService{}).
+		Complete(); err != nil {
+		setupLog.Error(err, "unable to create webhook", "webhook", "v1alpha2")
+		os.Exit(1)
+	}
+	if err = ctrl.NewWebhookManagedBy(mgr).
+		For(&v1beta1.InferenceService{}).
+		Complete(); err != nil {
+		setupLog.Error(err, "unable to create webhook", "webhook", "v1beta1")
+		os.Exit(1)
+	}
 
 	// Start the Cmd
 	log.Info("Starting the Cmd.")

config/configmap/inferenceservice.yaml

@@ -50,7 +50,7 @@ data:
         },
         "triton": {
             "image": "nvcr.io/nvidia/tritonserver",
-            "defaultImageVersion": "20.03-py3",
+            "defaultImageVersion": "20.08-py3",
             "supportedFrameworks": [
               "tensorrt",
               "tensorflow",