prometheus-to-sd

❗ Note

For collecting Prometheus metrics on Google Kubernetes Engine (GKE) clusters, use Google Cloud Monitoring (fka Stackdriver) supported integration.

• Documentation
• Source Code

Google Cloud Monitoring supported integration is designed to meet the expectations of Prometheus users and to make it easy to run with Prometheus server. It is intended to monitor all your applications, Kubernetes and beyond.

The source code in this directory (prometheus-to-sd) is developed for Google Kubernetes Engine to collect metrics from system services in order to support Kubernetes users. We designed the tool to be lean when deployed as a sidecar in your pod. It's intended to support only the metrics the Kubernetes team at Google needs and is not meant for end-users.

⚠ Google internal use only

Overview

prometheus-to-sd is a simple component that can scrape metrics stored in prometheus text format from one or multiple components and push them to the Stackdriver. Main requirement: k8s cluster should run on GCE or GKE.

Container Image

Look at the following link: https://gcr.io/google-containers/prometheus-to-sd and pick the latest image.

Usage

For scraping metrics from the component it's name, host, port and metrics should passed through the flag source in the next format: component-name:http://host:port?whitelisted=a,b,c&whitelistedLabels=podIdLabel:podId1,podId2|containerLabelName:containerName1,containerName2. If whitelisted part is omitted, then all metrics that are scraped from the component will be pushed to the Stackdriver, unless flag auto-whitelist-metrics=true was passed. If whitelistedLabels part is omitted, then all metrics that are scraped from the component will be pushed to Stackdriver. Otherwise, only metrics with labels with all label values included in the whitelistedLabel values will be pushed. In the above example, only metrics with podIdLabel in [podId1, podId2] and containerLabelName in [containerName1, containerName2] will be pushed. Valid labels to filter against are containerLabelName, namespaceIdLabel, and podIdLabel.

Custom metrics

To be able to push custom metrics to the Stackdriver flag stackdriver-prefix=custom.googleapis.com has to be specified. In such case metric bar from the component foo is going to be pushed to the Stackdriver as custom.googleapis.com/foo/bar. Prefix can be also configured per component by metricsPrefix parameter in source flag, for example: source=foo:http://localhost:1000?metricsPrefix=custom.googleapis.com&whitelisted=bar

Metrics autodiscovery

If metric descriptors already exist on the Stackdriver (created manually or by different component) then autodiscovery feature could be used. In such case prometheus-to-sd will push metrics for which metric descriptors are available on the Stackdriver. To use this feature a flag auto-whitelist-metrics=true has to be passed.

Resource descriptor

Each pushed metric includes monitored resource descriptor. Fields, such as project_id, cluster_name, instance_id and zone are filled automatically by the prometheus-to-sd. Values of the namespace_id and pod_id can be passed to the component through the additional flags or omitted. container_name is always empty for now. Field zone is overridable via flag.

From 0.8.0, prometheus-to-sd can also take monitoredResourceLabels flag to explicitly specify all MonitoredResource labels as a k8s node. When this flag is used, prometheus-to-sd will still fetch labels from GCE metadata server, but all of them will be overwritten if explicitly specified. This flag needs to be used together with monitoredResourceTypePrefix all the time. If pod-id, namespace-id flags aren't specified, or Prometheus-to-sd isn't configured to extract them from metric labels, metrics will be written as node metrics. E.g, if monitored-resource-type-prefix = "k8s_", metrics will be written to "k8s_node" type, if namespace-id or pod-id isn't specified. If namespace-id and pod-id are specified, metrics will be written to "k8s_pod" type, with additional resource labels namespace_name and pod_name. If Prometheus-to-sd is configured to scrape container name from metrics, metrics will be written to "k8s_container" type, with extra resource label "container_name".

From 0.9.0, custom resource type and labels can be specified per source config. Within the source URL, two new flags are supported:

• customResourceType specifies the resource type override. When this flag is present, it will override other resource type settings (e.g. monitored-resource-type-prefix)
• customLabel[<label_name>] specifies the labels that are provided for given customResourceType. If customResourceType is not specified, this field is ignored. All the labels for custom resource type has to be explicitly listed in the source config. Providing empty value for the following labels will cause prometheus-to-sd to use a value from GCE config:
  • cluster_name
  • instance_id
  • location
  • node_name
  • project_id

Example config:

--source=my-container:http://localhost:123?customResourceType=k8s_node&customLabels[node_name]=my-node&customLabels[project_id]&customLabels[location]&customLabels[cluster_name]

This will scrape metrics from localhost:123 and send them against k8s_node resource type, providing my-node as the node name and defaulting remaining labels using information from GCE Metadata Server.

Scrape interval vs. export interval

There are two flags: scrape-interval and export-interval that allow specifying how often metrics are read from the sources and how often they are exported to Stackdriver, respectively. By default both are set to 1m. The scrapes can be more frequent than exports, however, to achieve grater precision for metrics being exported. For example, if metrics are exported once every minute and a container dies between scrapes, up to 1 minutes of metrics can be lost. Frequent scrapes mitigate that, at the cost of elevated resource usage.

Deployment

Example of pod used to monitor kube-state-metrics component, that is used to collect different metrics about the state of k8s cluster.

Container monitoring agents

Container monitoring agents, such as cAdvisor, collect metrics about containers other than itself. Monitoring agents use prometheus labels for the namespace, pod, and container to differentiate metrics from different containers. To use these prometheus labels as the monitored resource labels in stackdriver, include namespaceIdLabel, podIdLabel, and containerNameLabel in the source flag in this format: component-name:http://host:port?whitelisted=a,b,c&namespaceIdLabel=d&podIdLabel=e&containerNameLabel=f. Note that prometheus labels used for the monitored resource are not included as labels in stackdriver.

For example, if prom-to-sd scraped the prometheus metric: container_cpu_usage_seconds{d="my-namespace",e="abc123",f="my-app",g="production"} 1.02030405e+09 It would be displayed in stackdriver with the namespace, my-namespace, the pod id, abc123, and the container name, my-app. The only stackdriver label for this metric would be g="production".