To Cluster or Not — With the BillWarehouse · medium · ~12 minThe events table (2.1TB) is naturally ordered by load time. The support tools team runs ~45,000 point-lookup queries a day filtered by account_id — each scanning nearly the full table because account values are scattered across every micro-partition. Someone proposes clustering by account_id; someone else says 'clustering is a scam that just burns credits'. Settle it with arithmetic.Credits cost $3. Scans currently average 0.9TB per lookup ≈ 0.05 credits/query. Clustering would cut lookups to ~15GB ≈ 0.001 credits/query, with reclustering maintenance estimated at 60 credits/week (moderate churn).Evidence

Workload

Numbers

Checkpoints

Why do the lookups scan almost everything today?

Do the month: with clustering, daily total ≈ 45 (queries) + 8.6 (maintenance) vs 2,250 today. How many DOLLARS saved per 30-day month? (2,250 − 53.6 ≈ 2,196 credits/day × 30 × $3)

When is the anti-clustering camp RIGHT?

Present the decision to the team — including how you'd VERIFY the estimate before and after. Interview-style.