Expose metrics about checks for prometheus

elb-instance-status.yml

@@ -1,16 +1,21 @@
 ---
 
-- name: Ensure there are at least 30% free inodes on /
+root_free_inodes:
+  name: Ensure there are at least 30% free inodes on /
   command: test $(df -i | grep "/$" | xargs | cut -d ' ' -f 5 | sed "s/%//") -lt 70
  
-- name: Ensure there are at least 30% free inodes on /var/lib/docker
+docker_free_inodes:
+  name: Ensure there are at least 30% free inodes on /var/lib/docker
   command: test $(df -i | grep "/var/lib/docker$" | xargs | cut -d ' ' -f 5 | sed "s/%//") -lt 70
  
-- name: Ensure there is at least 30% free disk space on /var/lib/docker
+docker_free_diskspace:
+  name: Ensure there is at least 30% free disk space on /var/lib/docker
   command: test $(df | grep "/var/lib/docker$" | xargs | cut -d ' ' -f 5 | sed "s/%//") -lt 70
  
-- name: Ensure volume on /var/lib/docker is mounted
+docker_mounted:
+  name: Ensure volume on /var/lib/docker is mounted
   command: mount | grep -q /var/lib/docker
  
-- name: Ensure docker can start a small container
+docker_start_container:
+  name: Ensure docker can start a small container
   command: docker run --rm alpine /bin/sh -c "echo testing123" | grep -q testing123

main.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/Luzifer/rconfig"
 	"github.com/gorilla/mux"
+	"github.com/prometheus/client_golang/prometheus"
 	"github.com/robfig/cron"
 )
 
@@ -30,7 +31,7 @@ var (
 
 	version = "dev"
 
-	checks               = []checkCommand{}
+	checks               = map[string]checkCommand{}
 	checkResults         = map[string]*checkResult{}
 	checkResultsLock     sync.RWMutex
 	lastResultRegistered time.Time
@@ -78,17 +79,19 @@ func main() {
 
 	r := mux.NewRouter()
 	r.HandleFunc("/status", handleELBHealthCheck)
+	r.Handle("/metrics", prometheus.Handler())
 	http.ListenAndServe(cfg.Listen, r)
 }
 
 func spawnChecks() {
-	for i := range checks {
+	for id := range checks {
-		go executeAndRegisterCheck(i)
+		go executeAndRegisterCheck(id)
 	}
 }
 
-func executeAndRegisterCheck(checkIndex int) {
+func executeAndRegisterCheck(checkID string) {
-	check := checks[checkIndex]
+	check := checks[checkID]
+	start := time.Now()
 
 	cmd := exec.Command("/bin/bash", "-c", check.Command)
 	err := cmd.Run()
@@ -97,21 +100,28 @@ func executeAndRegisterCheck(checkIndex int) {
 
 	checkResultsLock.Lock()
 
-	if _, ok := checkResults[check.Name]; !ok {
+	if _, ok := checkResults[checkID]; !ok {
-		checkResults[check.Name] = &checkResult{
+		checkResults[checkID] = &checkResult{
 			Check: check,
 		}
 	}
 
-	if success == checkResults[check.Name].IsSuccess {
+	if success == checkResults[checkID].IsSuccess {
-		checkResults[check.Name].Streak++
+		checkResults[checkID].Streak++
 	} else {
-		checkResults[check.Name].IsSuccess = success
+		checkResults[checkID].IsSuccess = success
-		checkResults[check.Name].Streak = 1
+		checkResults[checkID].Streak = 1
 	}
 
 	lastResultRegistered = time.Now()
 
+	if success {
+		checkPassing.WithLabelValues(checkID).Set(1)
+	} else {
+		checkPassing.WithLabelValues(checkID).Set(0)
+	}
+	checkExecutionTime.WithLabelValues(checkID).Observe(float64(time.Since(start).Nanoseconds()) / float64(time.Microsecond))
+
 	checkResultsLock.Unlock()
 }
 
@@ -121,7 +131,7 @@ func handleELBHealthCheck(res http.ResponseWriter, r *http.Request) {
 	buf := bytes.NewBuffer([]byte{})
 
 	checkResultsLock.RLock()
-	for cn, cr := range checkResults {
+	for _, cr := range checkResults {
 		state := ""
 		switch {
 		case cr.IsSuccess:
@@ -134,15 +144,17 @@ func handleELBHealthCheck(res http.ResponseWriter, r *http.Request) {
 			state = "CRIT"
 			healthy = false
 		}
-		fmt.Fprintf(buf, "[%s] %s\n", state, cn)
+		fmt.Fprintf(buf, "[%s] %s\n", state, cr.Check.Name)
 	}
 	checkResultsLock.RUnlock()
 
 	res.Header().Set("X-Collection-Parsed-In", strconv.FormatInt(time.Since(start).Nanoseconds()/int64(time.Microsecond), 10)+"ms")
 	res.Header().Set("X-Last-Result-Registered-At", lastResultRegistered.Format(time.RFC1123))
 	if healthy {
+		currentStatusCode.Set(http.StatusOK)
 		res.WriteHeader(http.StatusOK)
 	} else {
+		currentStatusCode.Set(http.StatusInternalServerError)
 		res.WriteHeader(http.StatusInternalServerError)
 	}
 

metrics.go

@@ -0,0 +1,50 @@
+package main
+
+import (
+	"log"
+	"os"
+
+	"github.com/prometheus/client_golang/prometheus"
+)
+
+var (
+	checkPassing       *prometheus.GaugeVec
+	checkExecutionTime *prometheus.SummaryVec
+	currentStatusCode  prometheus.Gauge
+
+	dynamicLabels = []string{"check_id"}
+)
+
+func init() {
+	hostname, err := os.Hostname()
+	if err != nil {
+		log.Fatalf("Unable to determine own hostname: %s", err)
+	}
+
+	co := prometheus.GaugeOpts{
+		Subsystem:   "elb_instance_status",
+		ConstLabels: prometheus.Labels{"hostname": hostname},
+	}
+
+	co.Name = "check_passing"
+	co.Help = "Bit showing whether the check PASSed (=1) or FAILed (=0), WARNs are also reported as FAILs"
+
+	cp := prometheus.NewGaugeVec(co, dynamicLabels)
+
+	co.Name = "status_code"
+	co.Help = "Contains the current HTTP status code the ELB is seeing"
+
+	csc := prometheus.NewGauge(co)
+
+	cet := prometheus.NewSummaryVec(prometheus.SummaryOpts{
+		Namespace:   co.Namespace,
+		Subsystem:   co.Subsystem,
+		ConstLabels: co.ConstLabels,
+		Name:        "check_execution_time",
+		Help:        "Timespan in µs the execution of the check took",
+	}, dynamicLabels)
+
+	checkPassing = prometheus.MustRegisterOrGet(cp).(*prometheus.GaugeVec)
+	currentStatusCode = prometheus.MustRegisterOrGet(csc).(prometheus.Gauge)
+	checkExecutionTime = prometheus.MustRegisterOrGet(cet).(*prometheus.SummaryVec)
+}