Teams that run Kubecost Enterprise with a unified multi-cluster view may choose to run Kubecost in “agent-mode” on non-Master clusters. This means that the Kubecost deployment on these clusters can be operated in a mode where they are only collecting metrics and sending to centralized durable storage.
It’s worth noting that Kubecost and it’s accompanying Prometheus always collect a reduced set of metrics that allow for lower resource consumption than a standard Prometheus deployment. The following configuration options further reduce resource consumption when not using the Kubecost frontend:
Disable product caching with the following parameters:
kubecostModel.warmCache=false kubecostModel.warmSavingsCache=false
kubecostModel.etl=false
Note: disabling these has UI performance implications
Disable grafana by setting global.grafana.enabled=false && global.grafana.proxy=false
When using durable storage…
Set prometheus.server.retention to < 7d, can go down to under 1d if not viewing UI on this cluster
Potentially reduce prometheus.server.persistentVolume.size depending on scale, or outright disable persistent storage
Disable thanos.query and thanos.store components (more info)
Reduce Prometheus query concurrency with prometheus.server.extraArgs.query.max-concurrency=1 if using Kubecost bundled Prometheus or directly in our product if using an external Prometheus/Thanos/Cortex (values flag)
For reference, here’s a sample list of helm set command line arguments when running Kubecost in agent mode:
--set kubecostModel.warmCache=false
--set kubecostModel.warmSavingsCache=false
--set kubecostModel.etl=false
--set prometheus.server.retention=”2d”
--set prometheus.alertmanager.enabled=false
There are potentially other configuration options based on the specifics of different deployment requirements. Reach out to [email protected] if you have any questions!
Edit this doc on Github