Selective metric collection with Prometheus

On my way to getting to know Prometheus I’ve noticed that many exporters for it expose tons of data, most of which I’d probably never use in my monitoring and alerting setup. The node exporter, for instance, also includes data about the exporter itself (like how many go-routines are active etc.). I guess collecting all that data for no practical reason won’t be all that beneficial with regards to memory and disk usage so I started to look at option to get Prometheus to only collect what I actually want to work with later on.

Ignoring some weird proxying approaches there are probably only the following two ways to get what I want:

  1. Prevent the exporters to even expose the data or
  2. prevent Prometheus from storing them

The first approach completely depends on how configurable the exporters I use are. For many of my tools I’ll probably use the official Go client library, which also exposes generally useful metrics from the underlying run-time like the number of go-routines or GC behaviour. These are enabled by default but I found a slightly hacky way to get rid of them:

// ...
func main() {
	http.Handle("/metrics", prometheus.UninstrumentedHandler())
	http.ListenAndServe("", nil)

func init() {
    // Register your own collector

With other collectors (like if they are already used by an app like the node exporter or are simply not as configurable as the Go one) I will have to go with option 2. For this, Prometheus offers the metric_relabel_configs setting for each scraper, where the labels for the metrics that should be collected can be manipulated:

  scrape_interval: 5s
  - job_name: "myservice"
      - targets:
          - "localhost:9999"
      - source_labels: [__name__]
        regex: go_(.*)
        action: drop
      - source_labels: [__name__]
        regex: go_(.*)
        action: drop

This tells Prometheus to drop the metrics with names starting with go_ or http_. The label __name__ is an internal label of every metric that represents the name of that metric.

I hope that with a combination of these two approaches I will be able to keep the number of unused data within Prometheus to a minimum 😊

comments powered by Disqus