feat(application): add application metrics for failed

cmd/tke-application-controller/controller-manager.go

@@ -20,10 +20,13 @@ package main
 
 import (
 	"math/rand"
+	"net/http"
 	"os"
 	"runtime"
 	"time"
 
+	"github.com/prometheus/client_golang/prometheus/promhttp"
+	"k8s.io/klog"
 	"tkestack.io/tke/cmd/tke-application-controller/app"
 )
 
@@ -33,5 +36,13 @@ func main() {
 		runtime.GOMAXPROCS(runtime.NumCPU())
 	}
 
+	mux := http.NewServeMux()
+	mux.Handle("/metrics", promhttp.Handler())
+	go func() {
+		if err := http.ListenAndServe(":9464", mux); err != nil {
+			klog.Fatal(err)
+		}
+	}()
+
 	app.NewApp("tke-application-controller").Run()
 }

pkg/application/controller/app/action/install.go

@@ -33,6 +33,7 @@ import (
 	"tkestack.io/tke/pkg/application/util"
 	chartpath "tkestack.io/tke/pkg/application/util/chartpath/v1"
 	"tkestack.io/tke/pkg/util/log"
+	"tkestack.io/tke/pkg/util/metrics"
 )
 
 // Install installs a chart archive
@@ -57,6 +58,7 @@ func Install(ctx context.Context,
 		if updateStatusFunc != nil {
 			if app.Status.Phase == applicationv1.AppPhaseInstallFailed {
 				log.Error(fmt.Sprintf("install app failed, helm pull err: %s", err.Error()))
+				metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 				// delayed retry, queue.AddRateLimited does not meet the demand
 				return app, nil
 			}
@@ -65,6 +67,7 @@ func Install(ctx context.Context,
 			newStatus.Reason = err.Error()
 			newStatus.LastTransitionTime = metav1.Now()
 			_, updateStatusErr := updateStatusFunc(ctx, app, &app.Status, newStatus)
+			metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 			if updateStatusErr != nil {
 				return nil, updateStatusErr
 			}
@@ -101,13 +104,15 @@ func Install(ctx context.Context,
 		if err != nil {
 			if app.Status.Phase == applicationv1.AppPhaseInstallFailed {
 				log.Error(fmt.Sprintf("install app failed, helm install err: %s", err.Error()))
+				metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 				// delayed retry, queue.AddRateLimited does not meet the demand
 				return app, nil
 			}
 			newStatus.Phase = applicationv1.AppPhaseInstallFailed
 			newStatus.Message = "install app failed"
 			newStatus.Reason = err.Error()
 			newStatus.LastTransitionTime = metav1.Now()
+			metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		} else {
 			newStatus.Phase = applicationv1.AppPhaseSucceeded
 			newStatus.Message = ""

pkg/application/controller/app/action/rollback.go

@@ -29,6 +29,7 @@ import (
 	helmaction "tkestack.io/tke/pkg/application/helm/action"
 	applicationprovider "tkestack.io/tke/pkg/application/provider/application"
 	"tkestack.io/tke/pkg/application/util"
+	"tkestack.io/tke/pkg/util/metrics"
 )
 
 // Rollback roll back to the previous release
@@ -62,6 +63,7 @@ func Rollback(ctx context.Context,
 			newStatus.Message = "rollback app failed"
 			newStatus.Reason = err.Error()
 			newStatus.LastTransitionTime = metav1.Now()
+			metrics.GaugeApplicationRollbackFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		} else {
 			newStatus.Phase = applicationv1.AppPhaseRolledBack
 			newStatus.Message = ""

pkg/application/controller/app/action/upgrade.go

@@ -33,6 +33,7 @@ import (
 	"tkestack.io/tke/pkg/application/util"
 	chartpath "tkestack.io/tke/pkg/application/util/chartpath/v1"
 	"tkestack.io/tke/pkg/util/log"
+	"tkestack.io/tke/pkg/util/metrics"
 )
 
 // Upgrade upgrade a helm release
@@ -58,6 +59,7 @@ func Upgrade(ctx context.Context,
 		if updateStatusFunc != nil {
 			if app.Status.Phase == applicationv1.AppPhaseUpgradFailed {
 				log.Error(fmt.Sprintf("upgrade app failed, helm pull err: %s", err.Error()))
+				metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 				// delayed retry, queue.AddRateLimited does not meet the demand
 				return app, nil
 			}
@@ -66,6 +68,7 @@ func Upgrade(ctx context.Context,
 			newStatus.Reason = err.Error()
 			newStatus.LastTransitionTime = metav1.Now()
 			updateStatusFunc(ctx, app, &app.Status, newStatus)
+			metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		}
 		return nil, err
 	}
@@ -102,13 +105,15 @@ func Upgrade(ctx context.Context,
 		if err != nil {
 			if app.Status.Phase == applicationv1.AppPhaseUpgradFailed {
 				log.Error(fmt.Sprintf("upgrade app failed, helm upgrade err: %s", err.Error()))
+				metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 				// delayed retry, queue.AddRateLimited does not meet the demand
 				return app, nil
 			}
 			newStatus.Phase = applicationv1.AppPhaseUpgradFailed
 			newStatus.Message = "upgrade app failed"
 			newStatus.Reason = err.Error()
 			newStatus.LastTransitionTime = metav1.Now()
+			metrics.GaugeApplicationUpgradeFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		} else {
 			newStatus.Phase = applicationv1.AppPhaseSucceeded
 			newStatus.Message = ""

pkg/application/controller/app/app_controller.go

@@ -373,26 +373,30 @@ func (c *Controller) syncAppFromRelease(ctx context.Context, cachedApp *cachedAp
 	if err != nil {
 		if app.Status.Phase == applicationv1.AppPhaseSyncFailed {
 			log.Error(fmt.Sprintf("sync app failed, helm list failed, err: %s", err.Error()))
+			metrics.GaugeApplicationSyncFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 			// delayed retry, queue.AddRateLimited does not meet the demand
 			return app, nil
 		}
 		newStatus.Phase = applicationv1.AppPhaseSyncFailed
 		newStatus.Message = "sync app failed"
 		newStatus.Reason = err.Error()
 		newStatus.LastTransitionTime = metav1.Now()
+		metrics.GaugeApplicationSyncFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		return c.updateStatus(ctx, app, &app.Status, newStatus)
 	}
 	rel, found := helmutil.Filter(rels, app.Spec.TargetNamespace, app.Spec.Name)
 	if !found {
 		if app.Status.Phase == applicationv1.AppPhaseSyncFailed {
 			log.Error(fmt.Sprintf("sync app failed, release not found: %s/%s", app.Spec.TargetNamespace, app.Spec.Name))
+			metrics.GaugeApplicationSyncFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 			// delayed retry, queue.AddRateLimited does not meet the demand
 			return app, nil
 		}
 		newStatus.Phase = applicationv1.AppPhaseSyncFailed
 		newStatus.Message = "sync app failed"
 		newStatus.Reason = fmt.Sprintf("release not found: %s/%s", app.Spec.TargetNamespace, app.Spec.Name)
 		newStatus.LastTransitionTime = metav1.Now()
+		metrics.GaugeApplicationSyncFailed.WithLabelValues(app.Spec.TargetCluster, app.Name).Set(1)
 		return c.updateStatus(ctx, app, &app.Status, newStatus)
 	}
 	app.Spec.Chart.ChartVersion = rel.Chart.Metadata.Version

pkg/util/metrics/util.go

@@ -79,3 +79,27 @@ func RegisterMetricAndTrackRateLimiterUsage(ownerName string, rateLimiter flowco
 	// }, updatePeriod, rateLimiterMetrics[ownerName].stopCh)
 	return nil
 }
+
+var (
+	GaugeApplicationInstallFailed = prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: "applicationInstallFailed",
+		Help: "application install failed count",
+	}, []string{"cluster", "application"})
+	GaugeApplicationUpgradeFailed = prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: "applicationUpgradeFailed",
+		Help: "application upgrade failed count",
+	}, []string{"cluster", "application"})
+	GaugeApplicationRollbackFailed = prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: "applicationRollbackFailed",
+		Help: "application rollback failed count",
+	}, []string{"cluster", "application"})
+	GaugeApplicationSyncFailed = prometheus.NewGaugeVec(prometheus.GaugeOpts{
+		Name: "applicationSyncFailed",
+		Help: "application sync failed count",
+	}, []string{"cluster", "application"})
+)
+
+func init() {
+	// Register the summary and the histogram with Prometheus's default registry.
+	prometheus.MustRegister(GaugeApplicationInstallFailed, GaugeApplicationUpgradeFailed, GaugeApplicationRollbackFailed, GaugeApplicationSyncFailed)
+}