Adding custom charts to components

StackState Kubernetes Troubleshooting

Overview

StackState provides already many metric charts by default on most types of components that represent Kubernetes resources. Extra metric charts can be added to any set of components whenever needed. When adding metrics to components there are 2 options:

The metrics are already collected by StackState but aren't visualized on a component by default
The metrics aren't yet collected by StackState at all and therefore aren't available yet

For option 1 the steps below will instruct you how to create a metric binding which will configure StackState to add a specific metric to a specific set of components.

In case of option 2, first make sure that the metrics are available in StackState, by sending them to StackState using the Prometheus remote write protocol. Only after that continue by adding charts for the metrics to the components.

Steps

Steps to create a metric binding:

As an example the steps will add a metric binding for the Replica counts of Kubernetes deployments. This is just an example, this metric binding already exists in StackState by default.

Write the outline of the metric binding

Create a new YAML file called metric-bindings.yaml and add this YAML template to it to create your own metric binding. Open it in your favorite code editor to change it throughout this guide. At the end the StackState CLI will be used to create and update the metric bindings in StackState.

nodes:
- _type: MetricBinding
  chartType: line
  enabled: true
  tags: {}
  unit: 
  name: 
  description: 
  priority: 
  identifier: urn:custom:metric-binding:...
  queries:
    - expression:
      alias:
  scope:
  layout:
    weight:
    layouts:
      - pespective: Metrics
        tab: 
          name: 
          weight:
        section: 
          name:
          weight:

The fields in this template are:

_type: StackState needs to know this is a metric binding so, value always needs to be MetricBinding
chartType: StackState will support different chart types (line, bar, etc.), currently only line is supported
enabled: Set to false to keep the metric binding but not show it to users
tags: Will be used to organize metrics in the user interface, can be left empty using {}
unit: The unit of the values in the time series returned by the query or queries, used to render the Y-axis of the chart. See the supported units reference for all units.
name: The name for the metric binding
description: Optional description, displayed on-hover of the name
priority: [Deprecated] One of HIGH, MEDIUM, or LOW. Main sort order for metrics on a component (in the order they're mentioned here), secondary sort order is the name.
identifier: A URN (universal resource identifier), used as the unique identifier of the metric binding. It must start with urn:custom:metric-binding:, the remainder is free-format as long as it's unique amongst all metric bindings.
queries: A list of queries to show in the chart for the metric binding (see also the following sections)
scope: The topology scope of the metric binding, a topology query that selects the components on which this metric binding will be shown.
layout: How to groups charts on different perspective views, e.g. on Metrics perspective

Fill in all the parts already known first (with the deployment replica counts as the example)

_type: MetricBinding
chartType: line
enabled: true
tags: {}
unit: short
name: Replica counts
priority: MEDIUM
identifier: urn:custom:metric-binding:my-deployment-replica-counts
queries:
  - expression:
    alias:
scope:

The queries and scope section will be filled in the next steps. Note that the unit used is short, which will simply render a numeric value. In case you're not sure yet about the unit of the metric, you can leave it open and decide on the correct unit when writing the PromQL query.

Write the topology query

Use the Explore view of the Topology perspective, http://your-stackstate-instance/#/views/explore, and select the components that need to show the new metric. Both the basic and advanced views can be used to make the selection. The most common fields to select topology on for metric bindings are type for the component type and label for selecting all the labels. For example for the deployments:

type = "deployment" and label = "stackpack:kubernetes"

The type filter selects all deployments, while the label filter selects only components created by the Kubernetes stackpack (label name is stackpack and label value is kubernetes). The latter can also be omitted to get the same result.

Switch to the advanced mode to copy the resulting topology query and put it in the scope field of the metric binding.

Metric bindings only support the query filters, query functions like withNeighborsOf aren't supported and can't be used.

Write the PromQL query

Go to the metric explorer of your StackState instance, http://your-stackstate-instance/#/metrics, and use it to query for the metric of interest. The explorer has auto-completion for metrics, labels, label values but also PromQL functions, and operators to help you out. Start with a short time range of, for example, an hour to get the best results.

For the total number of replicas use the kubernetes_state_deployment_replicas metric. To make the charts shown for this metric representative for the time series data extend the query to do an aggregation using the ${__interval} parameter:

max_over_time(kubernetes_state_deployment_replicas[${__interval}])

In this specific case use max_over_time to make sure the chart always shows the highest number of replicas at any given time. For longer time ranges this means that a short dip in replicas won't be shown, to emphasize the lowest number of replicas use min_over_time instead.

Copy the query into the expression property of the first entry in the queries field of the metric binding. Use Total replicas as an alias. This is the name that will show in the chart legend.

In StackState the size of the metric chart automatically determines the granularity of the metric shown in the chart. PromQL queries can adjusted to make optimal use of this behavior to get a representative chart for the metric. Writing PromQL for charts explains this in detail.

Bind the correct time series to each component

The metric binding with all fields filled in:

_type: MetricBinding
chartType: line
enabled: true
tags: {}
unit: short
name: Replica counts
priority: MEDIUM
identifier: urn:custom:metric-binding:my-deployment-replica-counts
queries:
  - expression: max_over_time(kubernetes_state_deployment_replicas[${__interval}])
    alias: Total replicas
scope: type = "deployment" and label = "stackpack:kubernetes"

Creating it in StackState and viewing the "Replica count" chart on a deployment component gives an unexpected result. The chart shows the replica counts for all deployments. Logically one would expect only 1 time series: the replica count for this specific deployment.

To fix this make the PromQL query specific for a component using information from the component. Filter on enough metric labels to select only the specific time series for the component. This is the "binding" of the correct time series to the component. For anyone experienced in making Grafana dashboards this is similar to a dashboard with parameters that are used in queries on the dashboard. Let's change the query in the metric binding to this:

max_over_time(kubernetes_state_deployment_replicas{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}", deployment="${name}"}[${__interval}])

The PromQL query now filters on 3 labels, cluster_name, namespace and deployment. Instead of specifying an actual value for these labels a variable reference to fields of the component is used. In this case the labels cluster-name and namespace are used, referenced using ${tags.cluster-name} and ${tags.namespace}. Further the component name is referenced with ${name}.

Supported variable references are:

Any component label, using ${tags.<label-name>}
The component name, using ${name}

The cluster name, namespace and a combination of the component type and name are ususally enough for selecting the metrics for a specific component from Kubernetes. These labels, or similar labels, are usually available on most metrics and components.

Create or update the metric binding in StackState

Use the StackState CLI to create the metric binding in StackState. Make sure the metric-bindings.yaml is saved and looks like this:

nodes:
- _type: MetricBinding
  chartType: line
  enabled: true
  tags: {}
  unit: short
  name: Replica counts
  priority: MEDIUM
  identifier: urn:custom:metric-binding:my-deployment-replica-counts
  queries:
    - expression: max_over_time(kubernetes_state_deployment_replicas{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}", deployment="${name}"}[${__interval}])
      alias: Total replicas
  scope: type = "deployment" and label = "stackpack:kubernetes"

Use the StackState CLI to create the metric binding:

sts settings apply -f metric-bindings.yaml

Verify the results in StackState by opening the metrics perspective for a deployment. If you're not happy with the result simply change the metric binding in the YAML file and run the command again to update it. The list of nodes supports adding many metric bindings. Simply add another metric binding entry to the YAML array using the same steps as before.

The identifier is used as the unique key of a metric binding. Changing the identifier will create a new metric binding instead of updating the existing one.

The sts settings command has more options, for example it can list all metric bindings:

sts settings list --type MetricBinding

Finally to delete a metric binding use

sts settings delete --ids <id>

The <id> in this command isn't the identifier but the number in the Id column of the sts settings list output.

The recommended way of working is to store metric bindings (and any other custom resources created in StackState) as YAML files in a Git repository. From there changes can be manually applied or it can be fully automated by using the StackState CLI in a CI/CD system like GitHub actions or GitLab pipelines.

Other options

More than 1 time series in a chart

There is only 1 unit for a metric binding (it gets plotted on the y-axis of the chart). As a result you should only combine queries that produce time series with the same unit in 1 metric binding. Sometimes it might be possible to convert the unit. For example, CPU usage might be reported in milli-cores or cores, milli-cores can be converted to cores by multiplying by 1000 like this (<original-query>) * 1000.

There are 2 ways to get more than 1 time series in a single metric binding and therefore in a single chart:

Write a PromQL query that returns multiple time series for a single component
Add more PromQL queries to the metric binding

For the first option an example is given in the next section. The second option can be useful for comparing related metrics. Some typical use-cases:

Comparing total replicas vs desired and available
Resource usage: limits, requests and usage in a single chart

To add more queries to a metric binding simply repeat steps 3. and 4. and add the query as an extra entry in the list of queries. For the deployment replica counts there are several related metrics that can be included in the same chart:

nodes:
- _type: MetricBinding
  chartType: line
  enabled: true
  tags: {}
  unit: short
  name: Replica counts
  priority: MEDIUM
  identifier: urn:custom:metric-binding:my-deployment-replica-counts
  queries:
    - expression: max_over_time(kubernetes_state_deployment_replicas{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}", deployment="${name}"}[${__interval}])
      alias: Total replicas
    - expression: max_over_time(kubernetes_state_deployment_replicas_available{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}",  deployment="${name}"}[${__interval}])
      alias: Available - ${cluster_name} - ${namespace} - ${deployment}
    - expression: max_over_time(kubernetes_state_deployment_replicas_unavailable{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}",  deployment="${name}"}[${__interval}])
      alias: Unavailable - ${cluster_name} - ${namespace} - ${deployment}
    - expression: min_over_time(kubernetes_state_deployment_replicas_desired{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}",  deployment="${name}"}[${__interval}])
      alias: Desired - ${cluster_name} - ${namespace} - ${deployment}
  scope: type = "deployment" and label = "stackpack:kubernetes"

Using metric labels in aliases

When a single query returns multiple time series per component, this will show as multiple lines in the chart. But in the legend they will all use the same alias. To be able to see the difference between the different time series the alias can include references to the metric labels using the ${label} syntax. For example here is a metric binding for the "Container restarts" metric on a pod, note that a pod can have multiple containers:

type: MetricBinding
chartType: line
enabled: true
id: -1
identifier: urn:custom:metric-binding:my-pod-restart-count
name: Container restarts
priority: MEDIUM
queries:
- alias: Restarts - ${container}
  expression: max by (cluster_name, namespace, pod_name, container) (kubernetes_state_container_restarts{cluster_name="${tags.cluster-name}", namespace="${tags.namespace}", pod_name="${name}"})
scope: (label = "stackpack:kubernetes" and type = "pod")
unit: short

Note that the alias references the container label of the metric. Make sure the label is present on the query result, when the label is missing the ${container} will be rendered as literal text to help troubleshooting.

Layouts

Each component can be associated with various technologies or protocols such as k8s, networking, runtime environments (e.g., JVM), protocols (HTTP, AMQP), etc. Consequently, a multitude of different metrics can be displayed for each component. For easier readability, StackState can organize these charts into tabs and sections. To display a chart (MetricBinding) within a specific tab or section, you need to configure the layout property. Any MetricsBinding without a specified layout will be displayed in a tab and section named Other.

Here is an example configuration:

layout:
  weight: 100
  layouts:
    - perspective: Metrics
      tab:
        name: AMQP
        weight: 150
      section:
        name: Performance
        weight: 300

Fields:

layout.weight - This represents the weight of the chart within sections. The charts are sorted in ascending order by weight, followed by alphabetical order.
layout.layouts - (array) This allows each chart to be added to multiple views (perspectives).
layout.layouts.perspective - This is the type of perspective to display the chart. Currently, only the Metrics perspective is supported.
- Metrics - display on Metrics perspective
layout.layouts.tab - This indicates the tab name and its weight. Tabs are sorted ascending by weight (the lowest one from all MetricBinding), followed by alphabetical order.
layout.layouts.section - This is the section name and its weight. Sections are sorted alphabetically.

PreviousCustom charts NextWriting PromQL queries for representative charts

Last updated 1 month ago