Prometheus remote write enables the capability to send metrics to a remote Prometheus storage server. This is primarily intended for long term storage.

• By default, Prometheus remote write sends all metrics stored in local Prometheus to a remote server.
• There’s a high likely chance you don’t need to send every metric that you have to a remote server only to increase the storage of that server unnecessarily.
• Another scenario maybe that you’re using remote write to send the metrics to a SaaS vendor like Grafana cloud / Datadog. Keep in mind that they charge for the amount of metrics you send and hence if you don’t cut down on what you send, expect a hefty bill on your way.

In this article, let’s see how to send only metrics from a specific scrape job and also only set of particular metrics that matches a name prefix.

Prerequisite:

• Docker
• Docker-compose

Sending metrics only from a specific scrape job :

• Let’s image a scenario where you’re scraping multiple services, and you want to send all the metrics from one specific job, for the sake of simplicity, let’s say you’re scraping 3 services metrics: Cadvisor, node_exporter & Prometheus itself.
• If you only want to send Cadvisor metrics via remote write and drop all others, then you can specify this via write_relabel_configs config block.
• Below is my docker-compose stack which is running 3 services, namely Prometheus, Cadvisor & node_exporter.

version: '3.8'
services:
  prometheus:
    image: prom/prometheus
    container_name: prometheus
    ports:
    - 9090:9090
    command:
    - --config.file=/etc/prometheus/prometheus.yml
    volumes:
    - ./prometheus.yaml:/etc/prometheus/prometheus.yml:ro
    depends_on:
    - cadvisor
    - node_exporter
  cadvisor:
    image: gcr.io/cadvisor/cadvisor
    container_name: cadvisor
    ports:
    - 8080:8080
    volumes:
    - /:/rootfs:ro
    - /var/run:/var/run:rw
    - /sys:/sys:ro
  node_exporter:
    image: quay.io/prometheus/node-exporter:latest
    container_name: node_exporter
    command:
      - '--path.rootfs=/host'
    pid: host
    volumes:
      - '/:/host:ro'
    ports:
      - 9100:9100

• Now let’s create a prometheus.yaml config to scrape the above targets and remote write to Grafana Cloud.

# prometheus global config
global:
  scrape_interval:  1s
  external_labels:
    environment: demo

scrape_configs:
  - job_name: prometheus
    static_configs:
      - targets: ['prometheus:9090']

  - job_name: cadvisor
    static_configs:
      - targets: ['cadvisor:8080']

  - job_name: node_exporter
    static_configs:
      - targets: ['node_exporter:9100']

remote_write:
- url: https://your-server-endpoint.grafana.net/api/prom/push
  basic_auth:
    username: your-username
    password: your-grafana-api-key
  write_relabel_configs:
  - source_labels: [job]
    regex: 'cadvisor'
    action: keep

• I’ve Chosen Grafana cloud because it's free and very easy to setup for a remote Prometheus server. This demo works for any remote write enabled Prometheus server.
• The magic is happening in write_relabel_configs block.
  • source_labels : The source labels select values from existing labels. In our case, all metrics scraped via the job Cadvisor will have the label Cadvisor.
  • regex : this is required is you want to match the metrics / label values returned via source_labels. In our case, we don't need any regex pattern. Just cadvisor will do as we need to match exactly that.
  • action : Action to perform based on regex matching, keep or drop. If you mention drop, all Cadvisor metrics will be dropped and all other job metrics will be sent.
• If you mention the action as keep, only Cadvsior metrics will be sent and all other metrics will be dropped.
• You can just change the name to any other job name which you want to send the metrics from.
• Now, you can start the stack by running : docker-compose up -d

Verification

• To verify how many metrics we have scraped from Cadvisor job along with others, query this metric from your local Prometheus : scrape_samples_scraped

• As we can see, there are 2269 time series from Cadvisor Job.
• We can now query the same metric scrape_samples_scraped from Remote Prometheus, which should result in the same number and should not list any other job in result:

Sending metrics which starts with a prefix :

• For example, If you only want to send metrics starting with node_memory i.e. all memory related metrics of a node only, , then you can use the config similar to below in your prometheus.yaml :

  write_relabel_configs:
    - source_labels: [__name__]
      regex: 'node_memory(.*)'
      action: keep

• The above config will only send all metrics which starts with node_memory prefix and will drop all other metrics.

---

Tip :

• If you want to check what are the labels available in your Prometheus, there’s an API endpoint exists for that : /api/v1/labels
• So you can make a curl request and get that list like this :

curl -s 'localhost:9090/api/v1/labels' | tr ',' '\n'

#result
"address"
"bios_date"
"job"
...

• Once you know what label you want to query, you can get the values for that specific label, for example for job label, we have 3 values available :

curl -s 'localhost:9090/api/v1/label/job/values’

#result
"data":["cadvisor","node_exporter","prometheus"]

---

Reference:

• https://tanmay-bhat.github.io/posts/how-to-prevent-metrics-explosion-in-prometheus/
• https://grafana.com/docs/grafana-cloud/metrics-control-usage/control-prometheus-metrics-usage/usage-reduction/