Send tracing spans to Prometheus

api/agent/agent.go

@@ -188,6 +188,8 @@ func (a *agent) Submit(callI Call) error {
 	ctx := call.req.Context()
 
 	span, ctx := opentracing.StartSpanFromContext(ctx, "agent_submit")
+	span.SetBaggageItem("fn_appname", callI.Model().AppName)
+	span.SetBaggageItem("fn_path", callI.Model().Path)
 	defer span.Finish()
 
 	// start the timer STAT! TODO add some wiggle room

api/server/fntracer.go

@@ -0,0 +1,130 @@
+package server
+
+import (
+	"github.com/opentracing/opentracing-go"
+	"github.com/opentracing/opentracing-go/log"
+	"strings"
+)
+
+// FnTracer is a custom Tracer which wraps another another tracer
+// its main purpose is to wrap the underlying Span in a FnSpan,
+// which adds some extra behaviour required for sending tracing spans to prometheus
+type FnTracer struct {
+	wrappedTracer opentracing.Tracer
+}
+
+// NewFnTracer returns a new FnTracer which wraps the specified Tracer
+func NewFnTracer(tracerToWrap opentracing.Tracer) opentracing.Tracer {
+	newTracer := &FnTracer{}
+	newTracer.wrappedTracer = tracerToWrap
+	return newTracer
+}
+
+// FnTracer implements opentracing.Tracer
+func (thisFnTracer FnTracer) StartSpan(operationName string, opts ...opentracing.StartSpanOption) opentracing.Span {
+	return NewFnSpan(thisFnTracer.wrappedTracer.StartSpan(operationName, opts...))
+}
+
+// FnTracer implements opentracing.Tracer
+func (thisFnTracer FnTracer) Inject(sm opentracing.SpanContext, format interface{}, carrier interface{}) error {
+	return thisFnTracer.wrappedTracer.Inject(sm, format, carrier)
+
+}
+
+// FnTracer implements opentracing.Tracer
+func (thisFnTracer FnTracer) Extract(format interface{}, carrier interface{}) (opentracing.SpanContext, error) {
+	return thisFnTracer.wrappedTracer.Extract(format, carrier)
+}
+
+// FnSpan is a custom Span that wraps another span
+// which adds some extra behaviour required for sending tracing spans to prometheus
+type FnSpan struct {
+	wrappedSpan opentracing.Span
+}
+
+// NewFnSpan returns a new FnSpan which wraps the specified Span
+func NewFnSpan(spanToWrap opentracing.Span) opentracing.Span {
+	newSpan := &FnSpan{}
+	newSpan.wrappedSpan = spanToWrap
+	return newSpan
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) Finish() {
+	thisFnSpan.copyBaggageItemsToTags()
+	thisFnSpan.wrappedSpan.Finish()
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) FinishWithOptions(opts opentracing.FinishOptions) {
+	thisFnSpan.copyBaggageItemsToTags()
+	thisFnSpan.wrappedSpan.FinishWithOptions(opts)
+}
+
+func (thisFnSpan FnSpan) copyBaggageItemsToTags() {
+	// copy baggage items (which are inherited from the parent) with keys starting with "fn" to tags
+	// the PrometheusCollector will send these to Prometheus
+	// need to do this because the collector can't access baggage items, but it can access tags
+	// whereas here we can access the parent's baggage items, but not its tags
+	thisFnSpan.Context().ForeachBaggageItem(func(k, v string) bool {
+		if strings.HasPrefix(k, "fn") {
+			thisFnSpan.SetTag(k, v)
+		}
+		return true
+	})
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) Context() opentracing.SpanContext {
+	return thisFnSpan.wrappedSpan.Context()
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) SetOperationName(operationName string) opentracing.Span {
+	return thisFnSpan.wrappedSpan.SetOperationName(operationName)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) SetTag(key string, value interface{}) opentracing.Span {
+	return thisFnSpan.wrappedSpan.SetTag(key, value)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) LogFields(fields ...log.Field) {
+	thisFnSpan.wrappedSpan.LogFields(fields...)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) LogKV(alternatingKeyValues ...interface{}) {
+	thisFnSpan.wrappedSpan.LogKV(alternatingKeyValues...)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) SetBaggageItem(restrictedKey, value string) opentracing.Span {
+	return thisFnSpan.wrappedSpan.SetBaggageItem(restrictedKey, value)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) BaggageItem(restrictedKey string) string {
+	return thisFnSpan.wrappedSpan.BaggageItem(restrictedKey)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) Tracer() opentracing.Tracer {
+	return thisFnSpan.wrappedSpan.Tracer()
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) LogEvent(event string) {
+	thisFnSpan.wrappedSpan.LogEvent(event)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) LogEventWithPayload(event string, payload interface{}) {
+	thisFnSpan.wrappedSpan.LogEventWithPayload(event, payload)
+}
+
+// FnSpan implements opentracing.Span
+func (thisFnSpan FnSpan) Log(data opentracing.LogData) {
+	thisFnSpan.wrappedSpan.Log(data)
+}

api/server/prom_zip_collector.go

@@ -0,0 +1,80 @@
+package server
+
+import (
+	"github.com/fnproject/fn/api/agent"
+	"github.com/openzipkin/zipkin-go-opentracing"
+	"github.com/openzipkin/zipkin-go-opentracing/thrift/gen-go/zipkincore"
+	"github.com/prometheus/client_golang/prometheus"
+	"strings"
+)
+
+// Each span name is published as a separate Histogram metric
+// Using metric names of the form fn_span_<span-name>_duration_seconds
+var histogramVecMap = make(map[string]*prometheus.HistogramVec)
+
+// Return the HistogramVec corresponding to the specified spanName.
+// If a HistogramVec does not already exist for specified spanName then one is created and configured with the specified labels
+// otherwise the labels parameter is ignored.
+func getHistogramVecForSpanName(spanName string, labels []string) *prometheus.HistogramVec {
+	thisHistogramVec, found := histogramVecMap[spanName]
+	if !found {
+		thisHistogramVec = prometheus.NewHistogramVec(
+			prometheus.HistogramOpts{
+				Name: "fn_span_" + spanName + "_duration_seconds",
+				Help: "Span " + spanName + " duration, by span name",
+			},
+			labels,
+		)
+		histogramVecMap[spanName] = thisHistogramVec
+		prometheus.MustRegister(thisHistogramVec)
+	}
+	return thisHistogramVec
+}
+
+// PrometheusCollector is a custom Collector
+// which sends ZipKin traces to Prometheus
+type PrometheusCollector struct {
+	a agent.Agent
+}
+
+// NewPrometheusCollector returns a new PrometheusCollector
+func NewPrometheusCollector(agent agent.Agent) (zipkintracer.Collector, error) {
+	pc := &PrometheusCollector{}
+	pc.a = agent
+	return pc, nil
+}
+
+// PrometheusCollector implements Collector.
+func (pc PrometheusCollector) Collect(span *zipkincore.Span) error {
+
+	// extract any label values from the span
+	labelKeys, labelValueMap := getLabels(span)
+
+	getHistogramVecForSpanName(span.GetName(), labelKeys).With(labelValueMap).Observe(float64(span.GetDuration()) / 1000000)
+
+	return nil
+}
+
+// extract from the specified span the key/value pairs that we want to add as labels to the Prometheus metric for this span
+// returns an array of keys, and a map of key-value pairs
+func getLabels(span *zipkincore.Span) ([]string, map[string]string) {
+
+	var keys []string
+	labelMap := make(map[string]string)
+
+	// extract any tags whose key starts with "fn" from the span
+	binaryAnnotations := span.GetBinaryAnnotations()
+	for _, thisBinaryAnnotation := range binaryAnnotations {
+		key := thisBinaryAnnotation.GetKey()
+		if thisBinaryAnnotation.GetAnnotationType() == zipkincore.AnnotationType_STRING && strings.HasPrefix(key, "fn") {
+			keys = append(keys, key)
+			value := string(thisBinaryAnnotation.GetValue()[:])
+			labelMap[key] = value
+		}
+	}
+
+	return keys, labelMap
+}
+
+// PrometheusCollector implements Collector.
+func (PrometheusCollector) Close() error { return nil }

api/server/server.go

@@ -86,7 +86,7 @@ func New(ctx context.Context, ds models.Datastore, mq models.MessageQueue, logDB
 	}
 
 	setMachineId()
-	setTracer()
+	s.setTracer()
 	s.Router.Use(loggerWrap, traceWrap, panicWrap)
 	s.bindHandlers(ctx)
 
@@ -117,7 +117,7 @@ func traceWrap(c *gin.Context) {
 	c.Next()
 }
 
-func setTracer() {
+func (s *Server) setTracer() {
 	var (
 		debugMode          = false
 		serviceName        = "fn-server"
@@ -126,17 +126,29 @@ func setTracer() {
 		// ex: "http://zipkin:9411/api/v1/spans"
 	)
 
-	if zipkinHTTPEndpoint == "" {
-		return
+	var collector zipkintracer.Collector
+
+	// custom Zipkin collector to send tracing spans to Prometheus
+	promCollector, promErr := NewPrometheusCollector(s.Agent)
+	if promErr != nil {
+		logrus.WithError(promErr).Fatalln("couldn't start Prometheus trace collector")
 	}
 
 	logger := zipkintracer.LoggerFunc(func(i ...interface{}) error { logrus.Error(i...); return nil })
 
-	collector, err := zipkintracer.NewHTTPCollector(zipkinHTTPEndpoint, zipkintracer.HTTPLogger(logger))
-	if err != nil {
-		logrus.WithError(err).Fatalln("couldn't start trace collector")
+	if zipkinHTTPEndpoint != "" {
+		// Custom PrometheusCollector and Zipkin HTTPCollector
+		httpCollector, zipErr := zipkintracer.NewHTTPCollector(zipkinHTTPEndpoint, zipkintracer.HTTPLogger(logger))
+		if zipErr != nil {
+			logrus.WithError(zipErr).Fatalln("couldn't start Zipkin trace collector")
 		}
-	tracer, err := zipkintracer.NewTracer(zipkintracer.NewRecorder(collector, debugMode, serviceHostPort, serviceName),
+		collector = zipkintracer.MultiCollector{httpCollector, promCollector}
+	} else {
+		// Custom PrometheusCollector only
+		collector = promCollector
+	}
+
+	ziptracer, err := zipkintracer.NewTracer(zipkintracer.NewRecorder(collector, debugMode, serviceHostPort, serviceName),
 		zipkintracer.ClientServerSameSpan(true),
 		zipkintracer.TraceID128Bit(true),
 	)
@@ -144,7 +156,10 @@ func setTracer() {
 		logrus.WithError(err).Fatalln("couldn't start tracer")
 	}
 	
-	opentracing.SetGlobalTracer(tracer)
+	// wrap the Zipkin tracer in a FnTracer which will also send spans to Prometheus
+	fntracer := NewFnTracer(ziptracer)
+
+	opentracing.SetGlobalTracer(fntracer)
 	logrus.WithFields(logrus.Fields{"url": zipkinHTTPEndpoint}).Info("started tracer")
 }
 

examples/grafana/README.md

@@ -90,11 +90,20 @@ Create a datasource to obtain metrics from Promethesus:
 * Click **Add** and then **Save and test**
 
 Import the example dashboard that displays metrics from the Fn server:
-* Click on the main menu at the top left and choose **Dashboards** and then **Home**
-* Click on **Home** at the top and then **Import dashboard**
+* Click on the main menu at the top left and choose **Dashboards** and then **Import**
 * In the dialog that opens, click **Upload .json file** and specify `fn_grafana_dashboard.json` in this example's directory.
 * Specify the Prometheus data source that you just created
 * Click **Import**
 
 You should then see the dashboard shown above. Now execute some functions and see the graphs update.
 
+## Tracing metrics
+
+Tracing spans from the Fn server are available as Prometheus metrics. Each span has a name that describes the operation being performed (for example `docker_wait_container`), and its duration in seconds. Each span name is represented by a separate histogram metric, which has a name of the form `fn_span_<span-name>_duration_seconds`. 
+
+If the span is associated with a specific function invocation, the corresponding metric is given the labels `fn_app` and `fn_path` which are set to the application name and function path respectively.
+
+A second example dashboard shows rate and duration data for a selection of tracing spans. This is `fn_grafana_dashboard2.json` in this example's directory. 
+
+<img src="../../docs/assets/GrafanaDashboard.png" width="800">
+