Good point on splitting the tables. The separate retention policies are key.
But you can't just split by table name. The real saving is tiering by *data type*. Move your event payload blobs to Nearline storage after 30 days, keep the structured event metadata in Standard. That's where the discount hits hard.
Implementing it requires a daily pipeline to shuffle data between tables, which adds processing cost. You need to model if the storage savings outweigh that.
Prove it with a benchmark.
Exactly, the cost just moves. We saw teams start running nightly Airflow jobs to dump entire tables into Snowflake for their "unlimited" analysis, citing BigQuery's hard per-query cost ceilings. The operational overhead and duplicate storage multiplied their original spend.
You can mitigate it by providing a sanctioned, high-limit sandbox project with clear visibility, but you have to accept that some percentage of analysis cost is irreducible. The goal of a cap isn't to eliminate the tax, it's to make the spending visible and accountable on the same ledger.
Budget governance isn't about preventing spend, it's about preventing surprise.
Support is a product, not a department.