Data Platform Cost

Cost driver

• Compute — query time × cost/sec (Snowflake credits).
• Storage — TB stored × $/TB-month.
• Egress — cross-region transfer.
• Vendor seat — Looker license per user.
• Pipeline run — Fivetran rows synced.

Optimization

• Query: avoid SELECT *, partition, cluster, materialized view.
• Storage: lifecycle policy — move cold to cheap tier.
• Compute: right-size warehouse, auto-suspend.
• Caching: result cache for repeat query.
• Pipeline: incremental dbt models, batch larger.

Cost monitoring

• Tag query by team/project.
• Dashboard cost by team monthly.
• Alert on anomaly (10× spike).
• Quota per team — kick when exceeded.

FinOps practice

• Showback (visibility) → Chargeback (bill them).
• Right-size review quarterly.
• Reserved capacity for predictable workload.
• Spot/preemptible for batch.

VN concern

• BigQuery on-demand cheap entry, expensive scale → switch to slots.
• Snowflake credits hide usage — must monitor.
• USD billing — depreciation VND impact.

Anti-pattern

• "Cost is engineering problem only" — business consumer needs to know cost of their query.
• Ignore cost until shock — invest visibility upfront.

Saving examples

• Cluster by date → 70% query cost cut for time-series.
• Materialized view for 100x repeat query → 95% cut.
• Compress before load → storage 30-50% saved.