Most enterprise Databricks teams maintain 3 to 6 core compute policies in production, categorized by workload type: Interactive/Shared (multi-tenant all-purpose clusters for analysts), Job/ETL (cost-optimized automated job clusters enforcing Photon, spot instances, and worker limits), and ML/Analytics (GPU and single-node options). Key best practices include managing policies as code via Terraform or Databricks Asset Bundles (DABs), mandating tagging for cost attribution, enforcing aggressive auto-termination (20โ30 minutes) on interactive clusters, restricting high-cost instance types, and leveraging Serverless Compute wherever possible to minimize cluster management overhead.
