Issue with current metrics #

We shouldn’t just say that the go-demo-5 application is slow. That would not provide enough information for us to quickly inspect the code in search of the exact cause of slowness. We should be able to do better and deduce which part of the application is misbehaving. Can we pinpoint a specific path that produces slow responses? Are all methods equally slow, or is the issue limited only to one? Do we know which function produces slowness? There are many similar questions we should be able to answer in situations like that, but we can’t with the current metrics. They are too generic, and they can usually only tell us that a specific Kubernetes resource is misbehaving. The metrics we’re collecting are too broad to answer application-specific questions.

The metrics we explored so far are a combination of exporters and instrumentations. Exporters are in charge of taking existing metrics and converting them into the Prometheus-friendly format. An example would be Node Exporter which is taking “standard” Linux metrics and converting them into Prometheus's time-series format. Another example is kube-state-metrics, which listens to the Kube API server and generates metrics with the state of the resources.

🔍 Instrumented metrics are baked into applications. They are an integral part of the code of our apps, and they are usually exposed through the /metrics endpoint.

The easiest way to add metrics to your applications is through one of the Prometheus client libraries. At the time of this writing, Go, Java and Scala, Python, and Ruby libraries are officially provided. On top of those, the community supports Bash, C++, Common Lisp, Elixir, Erlang, Haskell, Lua for Nginx, Lua for Tarantool, .NET / C#, Node.js, Perl, PHP, and Rust. Even if you code in a different language, you can easily provide Prometheus-friendly metrics by outputting results in a text-based exposition format.

The overhead of collecting metrics should be negligible and, since Prometheus pulls them periodically, outputting them should have a tiny footprint as well. Even if you choose not to use Prometheus, or to switch to something else, the format is becoming the standard, and your next metrics collector tool is likely to expect the same data.

All in all, there is no excuse not to bake metrics into your applications, and, as you’ll see soon, they provide invaluable information that we cannot obtain from outside.

Instrumented metrics in go-demo-5 #

Let’s take a look at an example of the instrumented metrics in go-demo-5.

🔍 The application is written in Go. Don’t worry if that’s not your language of choice. We’ll just take a quick look at a few examples as a way to understand the logic behind instrumentation, not the exact implementation.

The first interesting part is as follows.

var (
  histogram = prometheus.NewHistogramVec(prometheus.HistogramOpts{
    Subsystem: "http_server",
    Name:      "resp_time",
    Help:      "Request response time",
  }, []string{

Prometheus Histogram Vector #

We defined a variable that contains a Prometheus Histogram Vector with a few options. The Subsystem and the Name form the base metric http_server_resp_time. Since it is a histogram, the final metrics will be created by adding _bucket, _sum, and _count suffixes.

🔍 Please consult histogram documentation for more info about that Prometheus’ metric type.

The last part is a string array ([]string) that defines all the labels we’d like to add to the metric. In our case, those labels are service, code, method, and path. Labels can be anything we need, just as long as they provide sufficient information we might require when querying those metrics.

recordMetrics function #

The next point of interest is the recordMetrics function.

func recordMetrics(start time.Time, req *http.Request, code int) {
  duration := time.Since(start)
      "service": serviceName,
      "code":    fmt.Sprintf("%d", code),
      "method":  req.Method,
      "path":    req.URL.Path,

I created that as a helper function that can be called from different locations in the code. It accepts start time, the Request, and then return code as arguments. The function itself calculates duration by subtracting the current time with the start time. That duration is used in the Observe function and provides the value of the metric. There are also labels that will help us fine-tune our expressions later on.

Finally, we’ll take a look at one of the examples where the recordMetrics is invoked.

func HelloServer(w http.ResponseWriter, req *http.Request) {
  start := time.Now()
  defer func() { recordMetrics(start, req, http.StatusOK) }()

The HelloServer function is the one that returns the hello, world! response you already saw quite a few times. The details of that function are not important. In this context, the only part that matters is the line defer func() { recordMetrics(start, req, http.StatusOK) }(). In Go, defer allows us to execute something at the end of the function where it resides. In our case, that something is the invocation of the recordMetrics function that will record the duration of a request. In other words, before the execution leaves the HelloServer function, it’ll record the duration by invoking the recordMetrics function.

I won’t go further into the code that contains instrumentation since that would assume that you are interested in intricacies behind Go and I’m trying to keep this course language-agnostic. I’ll let you consult the documentation and examples from your favorite language. Instead, we’ll take a look at the go-demo-5 instrumented metrics in action.

kubectl -n metrics \
    run -it test \
    --image=appropriate/curl \
    --restart=Never \
    --rm \
    -- go-demo-5.go-demo-5:8080/metrics

We created a Pod based on the appropriate/curl image, and we sent a request through the Service using the address go-demo-5.go-demo-5:8080/metrics. The first go-demo-5 is the name of the Service, and the second is the Namespace where it resides. As a result, we got output with all the instrumented metrics available in that application. We won’t go through all of them, but only those created by the http_server_resp_time histogram.

The relevant parts of the output are as follows.

# HELP http_server_resp_time Request response time
# TYPE http_server_resp_time histogram
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.005"} 931
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.01"} 931
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.025"} 931
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.05"} 931
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.1"} 934
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.25"} 935
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="0.5"} 935
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="1"} 936
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="2.5"} 936
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="5"} 937
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="10"} 942
http_server_resp_time_bucket{code="200",method="GET",path="/demo/hello",service="go-demo",le="+Inf"} 942
http_server_resp_time_sum{code="200",method="GET",path="/demo/hello",service="go-demo"} 38.87928942600006
http_server_resp_time_count{code="200",method="GET",path="/demo/hello",service="go-demo"} 942

We can see that the Go library we used in the application code created quite a few metrics from the http_server_resp_time histogram. We got one for each of the twelve buckets (http_server_resp_time_bucket), one with the sum of the durations (http_server_resp_time_sum), and one with the count (http_server_resp_time_count). We would have much more if we made requests that would have different labels. For now, those fourteen metrics are all coming from requests that responded with the HTTP code 200, that used the GET method, that were sent to the /demo/hello path, and that is coming from the go-demo service (application). If we create requests with different methods (e.g., POST) or to different paths, the number of metrics would increase. Similarly, if we implement the same instrumented metric in other applications (but with different service labels), we’d have metrics with the same key (http_server_resp_time) that would provide insights into multiple apps. That raises the question of whether we should unify metric names across all the apps, or not.

Instrumented metrics of the same type #

I prefer having instrumented metrics of the same type with the same name across all the applications. For example, all those that collect response times can be called http_server_resp_time. That simplifies querying data in Prometheus. Instead of learning about instrumented metrics from each individual application, learning those from one provides knowledge about all. On the other hand, I am in favor of giving each team full control over their applications. That includes the decisions which metrics to implement, and how to call them.

Structure and responsibilities of teams #

All in all, it depends on the structure and responsibilities of the teams. If a team is entirely in charge of their applications and they debug problems specific to their apps, there is no inherent need for standardization of the names of instrumented metrics. On the other hand, if monitoring is centralized and the other teams might expect help from experts in that area, creating naming conventions is a must. Otherwise, we could easily end up with thousands of metrics with different names and types, even though most of them are providing the same information.

For the rest of this chapter, I will assume that we did agree to have http_server_resp_time histogram in all applications, where that’s applicable.

Prometheus discovery of Node Exporter and Kube State metrics #

Now, let’s see how we can tell Prometheus that it should pull the metrics from the go-demo-5 application. It would be even better if we could tell Prometheus to pull data from all the apps that have instrumented metrics. Actually, now when I think about it, we have not yet discussed how Prometheus finds Node Exporter and Kube State Metrics in the previous chapter. So, let’s go briefly through the discovery process.

A good starting point is the Prometheus target screen.

open "http://$PROM_ADDR/targets"

kubernetes-service-endpoints #

The most interesting group of targets is kubernetes-service-endpoints. If we take a closer look at the labels, we’ll see that each has kubernetes_name and that three of the targets have it set to go-demo-5. Prometheus somehow found that we have three replicas of the application and that metrics are available through the port 8080. If we look further, we’ll notice that prometheus-node-exporter is there as well, one for each node in the cluster. The same goes for prometheus-kube-state-metrics. There might be others in that group.

Get hands-on with 1000+ tech skills courses.