A modern data stack can burn six figures a month if nobody is watching. Cost is a design constraint, not an afterthought.
Partition and prune
Z-ordering, liquid clustering, partition pruning — the difference between scanning 12 TB and scanning 12 GB.
Right-size compute
Autoscaling with aggressive scale-down. Snowflake auto-suspend at 60s. Databricks job clusters over interactive whenever possible.
Showback, then chargeback
You can't optimize what you can't attribute. Tag every job, every query, every cluster with team, product, and pipeline.