Cloud bills explode from idle clusters, full-table scans, hot storage of cold data, and chatty APIs. Here are five high-value strategies interviewers like:
1. Right-size and shut down idle compute
Stop/terminate EMR/Databricks clusters when jobs end. Prefer ephemeral jobs (Glue, job clusters) over always-on interactive clusters for production ETL. Use autoscaling.
2. Lifecycle object storage
Move old S3/GCS/ADLS data to IA / Nearline / Cool / Archive classes. Delete temp scratch prefixes. Partition so you can expire by date.
3. Reduce query scan bytes
Partition and cluster warehouse tables. Select only needed columns (columnar formats help). Avoid SELECT * on huge fact tables. Cache or materialize heavy marts.
4. Pick serverless vs provisioned by utilization
Spiky workloads → serverless / on-demand. Steady high utilization → reserved / provisioned capacity (savings plans, slot commitments, reserved instances).
5. Observe and allocate
Turn on cost tags / labels per team/env. Budgets + alerts. Review top jobs weekly (Spark stage metrics, BigQuery bytes billed, warehouse credit burn).
Before: always-on cluster + Standard storage forever + SELECT * After: job clusters + lifecycle rules + partitioned queries + tags/alerts
Interview tip: List five concrete levers (idle compute, storage tiering, scan reduction, pricing model fit, cost observability). One example each beats a vague "be careful."