Instead of Prometheus local storage: VictoriaMetrics

Why VictoriaMetrics is better

• Prometheus local storage is not clustered or replicated and keeps 15 days by default. VictoriaMetrics is built to hold long retention and has a cluster version.
• VictoriaMetrics reports using up to 7x less RAM and up to 7x less storage than Prometheus, Thanos or Cortex. These are its own figures.
• VictoriaMetrics supports the Prometheus query API, so dashboards keep working, and its MetricsQL extends PromQL.
• The single-node version is one Apache-licensed binary, against the several components of Thanos or Cortex.

When to keep Prometheus local storage

Keep plain Prometheus local storage when a few weeks of history is enough, or on a single server where its retention settings cover your needs. Prometheus remains the standard scraper and alerting engine either way.

Switching

Add a remote_write block pointing at VictoriaMetrics to prometheus.yml, then add VictoriaMetrics as a Prometheus data source in Grafana. Historical data can be imported with vmctl. Later, vmagent can replace Prometheus for scraping, since it reads the same scrape config.

Watch out

MetricsQL is not identical to PromQL in every edge case, so check alert rules after switching. Some features, such as downsampling and multi-retention, are in the paid Enterprise edition. The project is run by one company, VictoriaMetrics Inc.

Also worth a look

• Grafana Mimir A horizontally scalable, multi-tenant Prometheus store on object storage; pick it for large multi-team setups already on Grafana.
• Thanos Adds long-term object storage and a global query view on top of existing Prometheus servers; pick it to keep Prometheus as is.