Promscale and Prometheus high availability

deprecation

Promscale has been discontinued. We strongly recommend that you do not use Promscale in a production environment. Learn more.

Promscale supports high availability alongside Prometheus high availability. Promscale is a stateless service, so it can use multiple replicas to run in a high availability (HA) mode. In this mode, a load balancer routes Promscale requests to any available replica.

To run Promscale in HA mode, you need multiple Prometheus instances running as a cluster. A cluster contains two or more identical Prometheus instances, each running on a different machine or container, with all instances scraping the same targets. This results in each Prometheus instance within the cluster having similar data. The data is not identical, because small differences can occur when data is scraped at slightly different times on each instance.

Promscale de-duplicates the data on each cluster after the data is ingested. One of the Prometheus instances is randomly elected leader, and all other instances only ingest data from that leader. This occurs until the leader stops sending data to Promscale for any reason. When the leader fails, Promscale selects a new Prometheus instance to act as leader instead, and the other instances ingest data from the new leader. Additionally, you can have multiple clusters sending data to the same set of Promscale instances, and in this case one leader is elected for each cluster.

To process data from Prometheus running in HA mode, Prometheus must be configured to communicate which cluster it belongs to. This is done using external labels. In particular, each Prometheus instance sends a cluster label to indicate which cluster it's in, and a __replica__ label that provides a unique identifier for the Prometheus instance within the cluster. Two Prometheus instances running as part of a HA cluster must send the same cluster label and different __replica__ labels.

In Kubernetes environments, it is useful to set the __replica__ label to match the pod name, and the cluster name as the Prometheus deployment or statefulset name. If you are using the Prometheus Operator, you can set it with these configuration parameters:

replicaExternalLabelName: "__replica__"
  prometheusExternalLabelName: "cluster"

In bare metal environments, and any environment where you are not using Kubernetes, you can configure external labels with the cluster and replica names in each Prometheus instance configuration, like this:

global:
  external_labels:
    __replica__: <REPLICA_NAME> (This should be unique name of the Prometheus instance)
    cluster: <CLUSTER_NAME> (This should be the name of the Prometheus deployment, which should be common across the Prometheus replica instances.)

When you have configured your Prometheus instances to send the correct labels, you can start Promscale with the --metrics.high-availability CLI flag, like this:

promscale --metrics.high-availability

Promscale automatically elects a single replica in each cluster as the current leader. Only data sent from that replica is ingested. If that leader replica stops sending data, then a new replica is elected as the leader.