Capture and expose notification delivery errors (#31)

This PR makes it possible to store the last error for each receiver in case of notification delivery failure. These errors are exposed via the `/api/v2/receivers` endpoint.

Co-authored-by: gotjosh <josue.abreu@gmail.com>

Author: 2 people

.gitignore

@@ -21,3 +21,7 @@
 !/.promu.yml
 !/api/v2/openapi.yaml
 !.github/workflows/*.yml
+
+# Editor
+.vscode
+.DS_Store

api/api.go

@@ -29,6 +29,7 @@ import (
 	"github.com/prometheus/alertmanager/cluster"
 	"github.com/prometheus/alertmanager/config"
 	"github.com/prometheus/alertmanager/dispatch"
+	"github.com/prometheus/alertmanager/notify"
 	"github.com/prometheus/alertmanager/provider"
 	"github.com/prometheus/alertmanager/silence"
 	"github.com/prometheus/alertmanager/types"
@@ -186,8 +187,8 @@ func (api *API) Register(r *route.Router, routePrefix string) *http.ServeMux {
 
 // Update config and resolve timeout of each API. APIv2 also needs
 // setAlertStatus to be updated.
-func (api *API) Update(cfg *config.Config, setAlertStatus func(model.LabelSet)) {
-	api.v2.Update(cfg, setAlertStatus)
+func (api *API) Update(cfg *config.Config, receivers []*notify.Receiver, setAlertStatus func(model.LabelSet)) {
+	api.v2.Update(cfg, setAlertStatus, receivers)
 }
 
 func (api *API) limitHandler(h http.Handler) http.Handler {

api/v2/api.go

@@ -33,7 +33,6 @@ import (
 	"github.com/prometheus/common/version"
 	"github.com/rs/cors"
 
-	"github.com/prometheus/alertmanager/api/metrics"
 	open_api_models "github.com/prometheus/alertmanager/api/v2/models"
 	"github.com/prometheus/alertmanager/api/v2/restapi"
 	"github.com/prometheus/alertmanager/api/v2/restapi/operations"
@@ -42,10 +41,13 @@ import (
 	general_ops "github.com/prometheus/alertmanager/api/v2/restapi/operations/general"
 	receiver_ops "github.com/prometheus/alertmanager/api/v2/restapi/operations/receiver"
 	silence_ops "github.com/prometheus/alertmanager/api/v2/restapi/operations/silence"
+
+	"github.com/prometheus/alertmanager/api/metrics"
 	"github.com/prometheus/alertmanager/cluster"
 	"github.com/prometheus/alertmanager/config"
 	"github.com/prometheus/alertmanager/dispatch"
 	"github.com/prometheus/alertmanager/matchers/compat"
+	"github.com/prometheus/alertmanager/notify"
 	"github.com/prometheus/alertmanager/pkg/labels"
 	"github.com/prometheus/alertmanager/provider"
 	"github.com/prometheus/alertmanager/silence"
@@ -73,7 +75,8 @@ type API struct {
 	logger log.Logger
 	m      *metrics.Alerts
 
-	Handler http.Handler
+	Handler   http.Handler
+	receivers []*notify.Receiver
 }
 
 type (
@@ -155,13 +158,14 @@ func (api *API) requestLogger(req *http.Request) log.Logger {
 }
 
 // Update sets the API struct members that may change between reloads of alertmanager.
-func (api *API) Update(cfg *config.Config, setAlertStatus setAlertStatusFn) {
+func (api *API) Update(cfg *config.Config, setAlertStatus setAlertStatusFn, receivers []*notify.Receiver) {
 	api.mtx.Lock()
 	defer api.mtx.Unlock()
 
 	api.alertmanagerConfig = cfg
 	api.route = dispatch.NewRoute(cfg.Route, nil)
 	api.setAlertStatus = setAlertStatus
+	api.receivers = receivers
 }
 
 func (api *API) getStatusHandler(params general_ops.GetStatusParams) middleware.Responder {
@@ -222,11 +226,40 @@ func (api *API) getStatusHandler(params general_ops.GetStatusParams) middleware.
 
 func (api *API) getReceiversHandler(params receiver_ops.GetReceiversParams) middleware.Responder {
 	api.mtx.RLock()
-	defer api.mtx.RUnlock()
+	configReceivers := api.receivers
+	api.mtx.RUnlock()
+
+	receivers := make([]*open_api_models.Receiver, 0, len(configReceivers))
+	for _, r := range configReceivers {
+		integrations := make([]*open_api_models.Integration, 0, len(r.Integrations()))
+
+		for _, integration := range r.Integrations() {
+			notify, duration, err := integration.GetReport()
+			iname := integration.String()
+			sendResolved := integration.SendResolved()
+			integrations = append(integrations, &open_api_models.Integration{
+				Name:                      &iname,
+				SendResolved:              &sendResolved,
+				LastNotifyAttempt:         strfmt.DateTime(notify.UTC()),
+				LastNotifyAttemptDuration: duration.String(),
+				LastNotifyAttemptError: func() string {
+					if err != nil {
+						return err.Error()
+					}
+					return ""
+				}(),
+			})
+		}
+
+		rName := r.Name()
+		active := r.Active()
+		model := &open_api_models.Receiver{
+			Name:         &rName,
+			Active:       &active,
+			Integrations: integrations,
+		}
 
-	receivers := make([]*open_api_models.Receiver, 0, len(api.alertmanagerConfig.Receivers))
-	for i := range api.alertmanagerConfig.Receivers {
-		receivers = append(receivers, &open_api_models.Receiver{Name: &api.alertmanagerConfig.Receivers[i].Name})
+		receivers = append(receivers, model)
 	}
 
 	return receiver_ops.NewGetReceiversOK().WithPayload(receivers)

api/v2/api_test.go

@@ -33,6 +33,7 @@ import (
 	receiver_ops "github.com/prometheus/alertmanager/api/v2/restapi/operations/receiver"
 	silence_ops "github.com/prometheus/alertmanager/api/v2/restapi/operations/silence"
 	"github.com/prometheus/alertmanager/config"
+	"github.com/prometheus/alertmanager/notify"
 	"github.com/prometheus/alertmanager/pkg/labels"
 	"github.com/prometheus/alertmanager/silence"
 	"github.com/prometheus/alertmanager/silence/silencepb"
@@ -484,14 +485,18 @@ receivers:
 		uptime:             time.Now(),
 		logger:             log.NewNopLogger(),
 		alertmanagerConfig: cfg,
+		receivers: []*notify.Receiver{
+			notify.NewReceiver("team-X", true, nil),
+			notify.NewReceiver("team-Y", true, nil),
+		},
 	}
 
 	for _, tc := range []struct {
 		body         string
 		expectedCode int
 	}{
 		{
-			`[{"name":"team-X"},{"name":"team-Y"}]`,
+			`[{"active":true,"integrations":[],"name":"team-X"},{"active":true,"integrations":[],"name":"team-Y"}]`,
 			200,
 		},
 	} {

api/v2/models/integration.go

@@ -510,8 +510,36 @@ definitions:
     properties:
       name:
         type: string
+      active:
+        type: boolean
+      integrations:
+        type: array
+        items:
+          $ref: '#/definitions/integration'
+    required:
+      - name
+      - active
+      - integrations
+  integration:
+    type: object
+    properties:
+      name:
+        type: string
+      sendResolved:
+        type: boolean
+      lastNotifyAttempt:
+        description: A timestamp indicating the last attempt to deliver a notification regardless of the outcome.
+        type: string
+        format: date-time
+      lastNotifyAttemptDuration:
+        description: Duration of the last attempt to deliver a notification in humanized format (`1s` or `15ms`, etc).
+        type: string
+      lastNotifyAttemptError:
+        description: Error string for the last attempt to deliver a notification. Empty if the last attempt was successful.
+        type: string
     required:
       - name
+      - sendResolved
   labelSet:
     type: object
     additionalProperties: