10 trillion samples a day: Scaling beyond traditional monitoring infra at Databricks

Databricks built Pantheon, a custom Thanos-based monitoring platform handling 5 billion timeseries and 10 trillion daily samples.

• •Tiered storage keeps real-time data in-memory, 24-hour data on-disk, and older data on object storage for cost efficiency
• •Control plane manages three isolated StatefulSets with automatic rollout coordination, hashring routing, and self-healing
• •Metric aggregation removes expensive labels from serverless workloads while preserving fleet-wide observability
• •Pantheon reduced monitoring downtime by 5x and saved millions in annual cloud costs

This summary was automatically generated by AI based on the original article and may not be fully accurate.