The Snowflake Cost Optimization Playbook

Eight levers, ranked by impact-to-effort.

Lever 1 — Auto-suspend by workload class

Default 600s auto-suspend is wasteful for ad-hoc analytics warehouses. Drop to 60s for interactive, keep 600s+ for batch ETL with frequent steps. Typical saving: 8–12%.

Lever 2 — Right-size, then right-cluster

Most warehouses are sized for peak. Run a week of EXPLAIN at the next size down on representative queries. If P99 latency stays inside SLA, downsize. Then enable multi-cluster with conservative scale-out.

Lever 3 — Result cache hygiene

Materialize analyst-favorite queries as views over result-cached subqueries. Free wins on dashboards.

Lever 4 — Clone instead of copy

Zero-copy cloning is genuinely free until divergence. Stop copying tables for dev environments.

Lever 5 — Storage tiering

Move history to permanent storage with reduced time-travel. The 90-day default is overkill for most fact tables.

Lever 6 — Query tagging + showback

Tag every query with team/use-case. Build a showback dashboard. Behavior changes within a quarter when teams see their own bills.

Lever 7 — Search optimization, surgically

It's expensive. Apply only to point-lookup tables that drive interactive queries.

Lever 8 — Materialized views, sparingly

They auto-maintain (good) but charge for it (bad). Use for query patterns that beat the maintenance cost — usually high read-to-write ratios on dimensional data.

A note on Cortex and AI workloads

If your roadmap includes Snowflake Cortex — Cortex Search, Cortex Analyst, or model serving — the same cost discipline applies. Cortex compute is metered separately and can grow quickly once a copilot starts seeing real traffic. Tag every Cortex call, watch the warehouse cost on the embedding refresh job, and add Cortex usage to the same showback dashboard. The Cortex deep-dive walks through how we set those guardrails up alongside the build.

Full PDF includes worked examples and a sizing worksheet.