Create HPA with Custom Metrics pulled through Instrumented metric

We confirmed that Prometheus' metrics, fetched by Prometheus Adapter and converted into Kubernetes’ custom metrics, can be used in HPA. So far, we used metrics pulled by Prometheus through exporters (nginx_ingress_controller_requests). Given that the adapter fetches metrics from Prometheus, it shouldn’t matter how they got there. Nevertheless, we’ll confirm that instrumented metrics can be used as well. That will give us an opportunity to cement what we learned so far and, at the same time, maybe learn a few new tricks.

Pull metrics through Instrumented metrics #

cat mon/prom-adapter-values-svc.yml

The output is yet another set of Prometheus Adapter Chart values.

  tag: v0.5.0
metricsRelistInterval: 90s
  url: http://prometheus-server.metrics.svc
  port: 80
  default: false
  - seriesQuery: 'http_server_resp_time_count{kubernetes_namespace!="",kubernetes_name!=""}'
        kubernetes_namespace: {resource: "namespace"}
        kubernetes_name: {resource: "service"}
      matches: "^(.*)server_resp_time_count"
      as: "${1}req_per_second_per_replica"
    metricsQuery: 'sum(rate(<<.Series>>{<<.LabelMatchers>>}[5m])) by (<<.GroupBy>>) / count(<<.Series>>{<<.LabelMatchers>>}) by (<<.GroupBy>>)'
  - seriesQuery: 'nginx_ingress_controller_requests'
        namespace: {resource: "namespace"}
        ingress: {resource: "ingress"}
      as: "http_req_per_second_per_replica"
    metricsQuery: 'sum(rate(<<.Series>>{<<.LabelMatchers>>}[5m])) by (<<.GroupBy>>) / sum(label_join(kube_deployment_status_replicas, "ingress", ",", "deployment")) by (<<.GroupBy>>)'

Combining both rules #

This time, we’re combining rules containing different metric series. The first rule is based on the http_server_resp_time_count instrumented metric that originates in go-demo-5. We used it in the Debugging Issues Discovered Through Metrics And Alerts chapter and there’s nothing truly extraordinary in its definition. It follows the same logic as the rules we used before. The second rule is a copy of one of the rules we used before.

What is interesting about those rules is that there are two completely different queries producing different results. However, the name is the same (http_req_per_second_per_replica) in both cases.

“Wait a minute”, you might say. The names are not the same. One is called ${1}req_per_second_per_replica while the other is http_req_per_second_per_replica. While that is true, the final name, excluding resource type, is indeed the same. I wanted to show you that you can use regular expressions to form a name. In the first rule, the name consists of matches and as entries. The (.*) part of the matches entry becomes the first variable (there can be others) which is later on used as part of the as value (${1}). Since the metric is http_server_resp_time_count, it will extract http_ from ^(.*)server_resp_time_count which, in the next line, is used instead of ${1}. The final result is http_req_per_second_per_replica, which is the same as the name of the second rule.

resources sections #

Now that we established that both rules will provide custom metrics with the same name, we might think that will result in a conflict. How will HPA know which metric to use, if both are called the same? Will the adapter have to discard one and keep the other? The answer lies in the resources sections.

A true identifier of a metric is a combination of its name and the resource it ties with. The first rule produces two custom metrics, one for Services and the other for Namespaces. The second also generates custom metrics for Namespaces, but for Ingresses as well. How many metrics is that in total? I’ll let you think about the answer before we check the result. To do that, we’ll have to upgrade the Chart for the new values to take effect.

helm upgrade prometheus-adapter \
    stable/prometheus-adapter \
    --version 1.4.0 \
    --namespace metrics \
    --values mon/prom-adapter-values-svc.yml

kubectl -n metrics \
    rollout status \
    deployment prometheus-adapter

We upgraded the Chart with the new values and waited until the Deployment rolls out.

Now we can go back to our pending question “how many custom metrics we’ve got?” Let’s see…

kubectl get --raw \
    "/apis/" \
    | jq "."

The output, limited to the relevant parts, is as follows.

      "name": "services/http_req_per_second_per_replica",
      "name": "namespaces/http_req_per_second_per_replica",
      "name": "ingresses.extensions/http_req_per_second_per_replica",

Now we have three custom metrics, not four. I already explained that the unique identifier is the name of the metric combined with the Kubernetes resource it’s tied to. All the metrics are called http_req_per_second_per_replica. But, since both rules override two resources, and namespace is set in both, one had to be discarded. We don’t know which one was removed and which stayed. Or maybe, they were merged. It does not matter since we shouldn’t override the same resource with the metrics with the same name. There was no practical reason for me to include namespace in adapter’s rule other than to show you that there can be multiple overrides and what happens when they are the same. Other than that silly reason, you can mentally ignore the namespaces/http_req_per_second_per_replica metric.

Two different custom metrics #

We used two different Prometheus expressions to create two different custom metrics, with the same name but related to other resources. One (based on nginx_ingress_controller_requests expression) comes from Ingress resources, while the other (based on http_server_resp_time_count) comes from Services. Even though the latter originates in go-demo-5 Pods, Prometheus discovered it through Services (as discussed in the previous chapter).

We can use the /apis/ endpoint not only to discover which custom metrics we have but also to inspect details, including values. For example, we can retrieve services/http_req_per_second_per_replica metric through the command that follows.

kubectl get --raw \
    "/apis/*/http_req_per_second_per_replica" \
    | jq .

The output is as follows.

  "kind": "MetricValueList",
  "apiVersion": "",
  "metadata": {
    "selfLink": "/apis/"
  "items": [
      "describedObject": {
        "kind": "Service",
        "namespace": "go-demo-5",
        "name": "go-demo-5",
        "apiVersion": "/v1"
      "metricName": "http_req_per_second_per_replica",
      "timestamp": "2018-10-27T23:49:58Z",
      "value": "1130m"

The describedObject section shows us the details of the items. Right now, we have only one Service with that metric. We can see that the Service resides in the go-demo-5 Namespace, that its name is go-demo-5, and that it is using v1 API version.

Further down, we can see the current value of the metric. In my case, it is 1130m, or slightly above one request per second. Since nobody is sending requests to the go-demo-5 Service, that value is as expected, considering that a health check is executed once a second.

Service-based metric used by updated HPA #

Next, we’ll explore the updated HPA definition that will use the Service-based metric.

cat mon/go-demo-5-hpa-svc.yml

The output is as follows.

apiVersion: autoscaling/v2beta1
kind: HorizontalPodAutoscaler
  name: go-demo-5
    apiVersion: apps/v1
    kind: Deployment
    name: go-demo-5
  minReplicas: 3
  maxReplicas: 10
  - type: Object
      metricName: http_req_per_second_per_replica
        kind: Service
        name: go-demo-5
      targetValue: 1500m

When compared with the previous definition, the only change is in the target and targetValue fields. Remember, the full identifier is a combination of the metricName and target. Therefore, this time we changed the kind to Service. We also had to change the targetValue since our application is receiving not only external requests through Ingress but also internal ones. They could be originating in other applications that might communicate with go-demo-5 or, as in our case, in Kubernetes’ health checks. Since their frequency is one second, we set the targetValue to 1500m, or 1.5 requests per second. That way, scaling will not be triggered if we do not send any requests to the application. Normally, you’d set a much bigger value. But, for now, we’re only trying to observe how it behaves before and after scaling.

Next, we’ll apply the changes to the HPA and describe it.

kubectl -n go-demo-5 \
    apply -f mon/go-demo-5-hpa-svc.yml

kubectl -n go-demo-5 \
    describe hpa go-demo-5

The output of the latter command, limited to the relevant parts, is as follows.

Metrics:                                                  ( current / target )
  "http_req_per_second_per_replica" on Service/go-demo-5: 1100m / 1500m
Deployment pods:                                           3 current / 3 desired
  Type    Reason             Age    From                       Message
  ----    ------             ----   ----                       -------
  Normal  SuccessfulRescale  12m    horizontal-pod-autoscaler  New size: 6; reason: Ingress metric http_req_per_second_per_replica above target
  Normal  SuccessfulRescale  9m20s  horizontal-pod-autoscaler  New size: 9; reason: Ingress metric http_req_per_second_per_replica above target
  Normal  SuccessfulRescale  4m20s  horizontal-pod-autoscaler  New size: 3; reason: All metrics below target

For now, there’s no reason for the HPA to scale up the Deployment. The current value is below the threshold. In my case, it’s 1100m.

Test autoscaling based on custom metrics #

Now we can test whether autoscaling based on custom metrics originating from instrumentation works as expected. Sending requests through Ingress might be slow, especially if our cluster runs in Cloud. The round-trip from our laptop all the way to the service might be too slow. So, we’ll send requests from inside the cluster, by spinning up a Pod and executing a request loop from inside it.

kubectl -n go-demo-5 \
    run -it test \
    --image=debian \
    --restart=Never \
    --rm \
    -- bash

Normally, I prefer alpine images since they are much small and efficient. However, for loops do not work from alpine, so we switched to debian instead. It doesn’t have curl though, so we’ll have to install it.

apt update

apt install -y curl

Now we can send requests that will generate enough traffic for HPA to trigger the scale-up process.

for i in {1..500}; do
    curl "http://go-demo-5:8080/demo/hello"


We sent five-hundred requests to /demo/hello endpoint, and we exited the container. Since we used the --rm argument when we created the Pod, it will be removed automatically from the system, so we do not need to execute any cleanup operation.

Let’s describe the HPA and see what happened.

kubectl -n go-demo-5 \
    describe hpa go-demo-5

The output, limited to the relevant parts, is as follows.

Reference:                                                Deployment/go-demo-5
Metrics:                                                  ( current / target )
  "http_req_per_second_per_replica" on Service/go-demo-5: 1794m / 1500m
Min replicas:                                             3
Max replicas:                                             10
Deployment pods:                                          3 current / 4 desired
... Message
... -------
... New size: 6; reason: Ingress metric http_req_per_second_per_replica above target
... New size: 9; reason: Ingress metric http_req_per_second_per_replica above target
... New size: 3; reason: All metrics below target
... New size: 4; reason: Service metric http_req_per_second_per_replica above target

HPA detected that the current value is above the target (in my case it’s 1794m) and changed the desired number of replicas from 3 to 4. We can observe that from the last event as well. If, in your case, the desired number of replicas is still 3, please wait for a few moments for the next iteration of HPA evaluations and repeat the describe command.

If we need an additional confirmation that scaling indeed worked as expected, we can retrieve the Pods in the go-demo-5 Namespace.

kubectl -n go-demo-5 get pods

The output is as follows.

go-demo-5-db-0 2/2   Running 0        33m
go-demo-5-db-1 2/2   Running 0        32m
go-demo-5-db-2 2/2   Running 0        32m
go-demo-5-...  1/1   Running 2        33m
go-demo-5-...  1/1   Running 0        53s
go-demo-5-...  1/1   Running 2        33m
go-demo-5-...  1/1   Running 2        33m

There’s probably no need to confirm that the HPA will soon scale down the go-demo-5 Deployment after we stopped sending requests.

Get hands-on with 1200+ tech skills courses.