Snowflake FinOps: The Compute Credit Trap and How to Stop It
Set auto‑suspend to 30–60 seconds for all warehouses, right‑size based on QUERY_HISTORY, cap multi‑cluster to 3–5, and enable query attribution to track per‑warehouse credits.
Update Snowflake warehouse configurations: set auto‑suspend to 30s, adjust sizes per query history, set max clusters to 3, and enable query attribution.
Summary
A large Snowflake warehouse running 24/7 can cost $11,520 per month on AWS Standard Edition, and a typical data team with four warehouses can spend $46,080 per month before tuning. The same workload, when auto‑suspend is enabled, warehouses are right‑sized, multi‑cluster scaling capped, and query‑level attribution in place, can run for $18,000 to $25,000 per month, a 60% savings. The article explains four levers: auto‑suspend, warehouse sizing, multi‑cluster caps, and query‑level attribution. Auto‑suspend defaults to 600 seconds; reducing it to 60 seconds can recover $1,050 per month per warehouse by eliminating idle time. Right‑sizing warehouses based on ACCOUNT_USAGE.QUERY_HISTORY shows that 80% of queries finish in under 30 seconds, so a SMALL warehouse can handle most work, while a LARGE warehouse is reserved for the 20% that take longer. Multi‑cluster warehouses bill each running cluster; setting a realistic max‑cluster cap (3–5) and using economy scaling policy can cut costs from $400 to $5 per 1TB write workload. Query attribution allows teams to see which warehouse consumes how many credits, making it easier to adjust sizes and limits.
Key changes
- Set auto‑suspend to 30–60 seconds to recover idle time costs
- Right‑size warehouses based on ACCOUNT_USAGE.QUERY_HISTORY p50/p95/p99 durations
- Cap multi‑cluster warehouses to 3–5 clusters and use economy scaling policy
- Enable query attribution to track per‑warehouse credit consumption
- Switch from synchronous multi‑region active‑active replication to async cross‑region snapshots
- Use async cross‑region snapshots for analytics workloads instead of real‑time replication
- Apply a 10‑cluster cap with economy scaling policy for analytical workloads
- Track peak concurrency from QUERY_HISTORY to set realistic max‑cluster caps